Anda di halaman 1dari 13

Transportation Research Part B 44 (2010) 633–645

Contents lists available at ScienceDirect

Transportation Research Part B


journal homepage: www.elsevier.com/locate/trb

A spatial dynamic panel model with random effects applied


to commuting times
Olivier Parent a,*, James P. LeSage b
a
Department of Economics, University of Cincinnati, Cincinnati, OH 45221, United States
b
Fields Endowed Chair in Urban and Regional Economics, McCoy College of Business Administration, Department of Finance and Economics,
Texas State University – San Marcos, San Marcos, TX 78666, United States

a r t i c l e i n f o a b s t r a c t

Article history: A space–time filter is set forth for spatial panel data situations that include random effects.
Received 12 January 2010 We propose a general spatial dynamic specification that encompasses several spatiotempo-
Accepted 12 January 2010 ral models previously used in the panel data literature. We apply the model to the case of
highway induced travel demand. The theory of induced travel demand asserts that
increased highway capacity will induce growth in traffic for a number of reasons. Our
Keywords: model allows us to quantify the spatial spillover impacts of increased highway capacity
Spatial correlation
at one location in the network on travel times in neighboring locations and in future time
Dynamic panels
MCMC estimations
periods.
Ó 2010 Elsevier Ltd. All rights reserved.

1. Introduction

We introduce a dynamic panel data model with simultaneous spatial dependence that extends the literature on dynamic
panel data models to a spatial setting where the cross-sectional units are regions or points located in space. Simultaneous
spatial dependence allows for dependence of one region on neighboring regions at the same time period, producing a model
that can be used to quantify spatial spillovers (see LeSage and Pace, 2009). For example, monthly observations regarding
flows on a transportation road network that passes through neighboring regions will produce sample data consistent with
simultaneous spatial dependence in the flows. LeSage and Pace (2009) point out that given a monthly observation time
frame, daily or hourly congestion effects will produce monthly data observations that are best modeled using a contempo-
raneous spatial lag of the dependent variable flows. The partial derivative effects associated with changing the explanatory
variables in models that include simultaneous spatial dependence can be shown to reflect direct and indirect or spatial spill-
over effects. Specifically, the own-partial derivative for the rth explanatory variable, @yit =@X rit , represents the direct effect on
region i’s traffic flows (at time t) arising from a change in the rth explanatory variable for region i (at time t). The cross-par-
tial, @yjt =@X rit , measures the effect on regions other than i, where most of the spatial spillover impact falls on nearby or neigh-
boring regions.
The model found frequently in transportation applications (see e.g. Garrido, 2000) is the STAR (space–time autoregressive
model) of Pfeifer and Deutsch (1980). The STAR model excludes simultaneous spatial dependence. Cressie (1993, p. 450) crit-
icizes the lack of an adequate treatment of simultaneous spatial dependence in the STAR (and STARMA) models. The stan-
dard STAR specification found in the literature implicitly imposes that after conditioning on past observations, the process is
uncorrelated across space. Even Pfeifer and Deutsch, 1980 criticize their own model noting that after conditioning on the

* Corresponding author.
E-mail addresses: olivier.parent@uc.edu (O. Parent), jlesage@spatial-econometrics.com (J.P. LeSage).

0191-2615/$ - see front matter Ó 2010 Elsevier Ltd. All rights reserved.
doi:10.1016/j.trb.2010.01.004
634 O. Parent, J.P. LeSage / Transportation Research Part B 44 (2010) 633–645

past history of the process many observed panel data series display considerable contemporaneous spatial dependence. This
has lead to the introduction of a contemporaneous spatial lag of the dependent variable and random effects that can be moti-
vated as arising due to (1) common shocks affecting all the spatial units during the current time period and (2) idiosyncratic
shocks that spread to neighboring locations within the current time sampling interval. Cressie (1993, p. 450) points out that
we will tend to observe high values of simultaneous spatial dependence when the sampling interval is relatively long with
respect to the speed of the underlying diffusion process. This is of course the case with our annual panel data application.
We illustrate the usefulness of the dynamic spatial lag panel model using state-level travel times over the 2000–2006 per-
iod as they relate to highway capacities (and other pertinent explanatory variables). Our objective is to better understand
how space and time dynamics can explain the evolution of mean travel times to work between states. Nationwide, the trans-
portation system is experiencing number of challenges such as growing user demand and adequate investment to maintain
and enhance transportation infrastructure. In order to decide how much to spend on roads, Federal, state and local govern-
ments must evaluate the demand for travel on road networks and other public services. The primary source of highway fund-
ing at the state and federal levels is highway-user fees, including tolls and motor fuel taxes. Observed spatial dependence in
travel times suggests interdependence in both the problem and solution with a role for both states and national authorities
in funding and maintaining the transportation system. Following Noland (2001), we estimate for each state a dynamic panel
log linear relationship that relates mean travel-to-work times and a set of explanatory variables that include: highway
expenditures as a percentage of gross state product, traffic volume in urban areas, taxes on gasoline and the number of lane
miles as a measure of highway capacity. The latter variable is introduced to assess whether capacity expansion results in an
overall increase in traffic growth. In fact, increasing highway capacity should reduce the cost and time of traveling which
may induce additional vehicle traffic. Our estimation results show that higher taxes on gasoline as well as increased expen-
ditures on highways lead to a current time period reduction in travel times in the own- and neighboring states, and these
benefits spillover during future time periods in both the own- and other states. A focus of our investigation is on interpreting
the direct (own-state) and indirect spatial spillovers to neighboring states over time that arise in our space–time dynamic
model.
In Section 2, we set forth a spatial dynamic panel data model that includes random effects and discuss the impact of
changes in the explanatory variables by examining partial derivatives for the model that allow us to quantify direct and indi-
rect spatial effects as well as direct and indirect effects in future time periods. In Section 3 we propose and discuss Bayesian
estimation of the model using Markov Chain Monte Carlo (MCMC) methods. This approach greatly simplifies estimation of
this type of model where the numerous parameters make maximum likelihood estimation difficult. An empirical illustration
based on mean travel times to work is presented in Section 4 which allows us to illustrate interpretation of the various mea-
sures of impact relating to spillovers that occur in both space and time.

2. Model framework

The dynamic spatial lag panel model we introduce here reflects a departure from models that dynamically model space
and time dependence in the error terms (e.g., Su and Yang, 2007 and Parent and LeSage, 2008), which we label error models.
The error models can be expressed as shown in (1) and the accompanying definitions.

Y ¼ iNT a þ Xb þ g ð1Þ
g ¼ Zl þ e
Qe ¼ m
0 1
w 0 ... 0
B / 1 ... 0C
B C
C¼B
B .. .. .. .. C
C ð2Þ
@ . . . .A
0 . . . / 1
B ¼ ðIN  qWÞ ð3Þ
Q ¼CB ð4Þ
Q ¼ INT  qIT  W  /L  IN þ ðq  /ÞL  W ð5Þ
 0
where Y ¼ y01 ; . . . ; y0T ; yt ¼ ðy1t ; . . . ; yNt Þ0 is the N 1 vector of observations for the tth time period, a is the intercept, iNT is
0
an NT  1 column vector of ones, and X ¼ x01 ; . . . ; x0T so that xt denotes the N  k matrix of non-stochastic regressors at time
t. The matrix L represents the time lag operator Lyt ¼ yt1 . The N  1 column vector l represents random effects with
li N 0; r2l , and it is typically assumed that l is uncorrelated with et . The matrix Z ¼ iT  IN , and the disturbance vector
e ¼ e01 ; . . . ; e0T 0 with et ¼ ðe1t ; . . . ; eNt Þ0 is assumed to exhibit spatial dependence measured by the parameter q as well as first
order time dependence reflected in the scalar parameter /, with covariance between space and time captured by the term
L  W and associated parameters q  /. The disturbance vector m ¼ ðm1t ; . . . ; mNt Þ0 is i:i:d: across i and t with zero mean and
variance r2m . The space filter matrix B ¼ ðIN  qWÞ is nonsingular, where the scalar spatial dependence parameter is q and
the N  N matrix W is assumed to be a known row stochastic spatial weight matrix with zeros on the diagonal. This matrix
defines the dependence between cross-sectional spatial units. We will also assume that W was created by row-normalizing a
O. Parent, J.P. LeSage / Transportation Research Part B 44 (2010) 633–645 635

symmetric matrix, so that all eigenvalues are less than or equal to one. Note that the variance for the error term g is equiv-
alent to

X ¼ r2l ðJT  IN Þ þ r2e INT ð6Þ

with J T ¼ iT i0T .
Parent and LeSage (2008) focusing on different values for w, discuss treatment of initial period values in the time filter
matrix C as endogenous or assumed exogenous and known, so we ignore these issues here in an effort to focus on spatial
spillovers. We note that for the error models there are no spatial spillovers because: @yit =@X rit ¼ br and @yjt =@X rit ¼ 0 consistent
with the usual assumption that observations from the cross-section of regions are independent. Baltagi et al. (2007) discuss
the error model and Parent and LeSage (2008) point out that these models can be viewed as consisting of filters for both
spatial and time dependence in the disturbances. Specifically, the Kronecker product Q of the space filter B ¼ IN  qW and
time filter C leads to a cross-product term and associated parameter hL  W, with an implied restriction h ¼ ðq  /Þ.
Yu et al. (2008) consider a dynamic spatial lag panel model that allows for both time and spatial dependence as well as a
cross-product term reflecting spatial dependence at a one-period time lag, as shown in (7).
yt ¼ /yt1 þ qWyt þ hWyt1 þ iN a þ xt b þ gt
gt ¼ l þ et ð7Þ
where et is i:i:d: across i and t with zero mean and variance r2e In .
The parameter h has to be estimated and thus model can be written as in (10) to emphasize the space–time filter applied
to the dependent variable vector.
Yu et al. (2008) discuss maximum likelihood estimation for the case where the constraint h ¼ q  / is relaxed and the
parameter h is treated as unrestricted. We discuss the relevance and implications of this constraint for Bayesian Markov
Chain Monte Carlo (MCMC) estimation of the model. In Parent and LeSage (2009), we show that use of this constraint can
greatly simplify the MCMC estimation procedure. In this case the model is equivalent to:
QY ¼ iNT a þ Xb þ g ð8Þ
g ¼ Zl þ e
Q ¼ INT  qIT  W  /L  IN þ ðq  /ÞL  W ð9Þ
Yu et al. (2008) do not model the observations of the first period and assume the dynamic space–time process is condi-
tional on this initial cross-section. Su and Yang (2007) and Parent and LeSage (2008) find that for the case of the dynamic
panel error model where spatial dependence arises in the disturbance terms, estimates and inference are sensitive to treat-
ment of the initial period observations in cases where there are a small number of time periods. For the case of dynamic
spatial lag panel data models the treatment of the initial period observations is even more important when drawing infer-
ences regarding direct versus spatial spillover impacts that arise from changes in the explanatory variables. A careful analysis
of issues related to treatment of the first period observation can be found in Parent and LeSage (2009). Assuming that the
first cross-section is unconditional and the within-sample observations and pre-sample values were generated by the same
process, they propose a prediction for the first cross-section that exploits spatial dependence between locations. For simplic-
ity of exposition, we assume that the first period is only subject to spatial dependence and set w ¼ 1.1 In this case, the Kro-
necker product Q, has to be replaced by the NT  NT matrix P and the model can be rewritten as:
PY ¼ iNT a þ Xb þ g ð10Þ
g ¼ Zl þ e
0 1
B 0
B .. C
B ð/IN þ hWÞ . C
B C
P¼B .. .. C ð11Þ
B C
@ . . A
0 . . . ð/IN þ hWÞ B
The associated data generating process (DGP) shown in (12).

Y ¼ P1 ½iNT a þ Xb þ Z l þ e ð12Þ


X
k
Y¼ SðrÞ X ðrÞ þ P1 ½iNT a þ Z l þ e ð13Þ
r¼1

SðrÞ ¼ P1 INT br ð14Þ

1
This assumption appears to be reasonable in our empirical study where we find evidence of weak time dependence. In addition, a comparison of this
approach with a methodology that utilizes an optimal prediction for the first cross-section produced similar results.
636 O. Parent, J.P. LeSage / Transportation Research Part B 44 (2010) 633–645

In (13) we let X ðrÞ denote the rth column from the NT  k matrix X, allowing us to express the DGP in a form suitable for
considering the partial derivative impacts that arise from changes in the rth explanatory variable. Of course, the values taken
by the rth explanatory variable change with time periods so we need to further elaborate expression (13). For future refer-
ence we note that the matrix P 1 takes the form of a lower-triangular block matrix, containing blocks with N  N matrices.
Specifically, it takes the simple structure shown in (15), when the restriction h ¼ q  / is imposed.
0 1
B1 0 ... 0
B .. .. C
B C
B D1 . . C
B C
B .. C
P1 ¼ B D2 D1 . C
B C
B. .. .. C
B. C
@. . . 0 A
DT1 DT2 . . . D1 B1
Di ¼ /i  B1 ð15Þ
We work with this restricted variant of the model which allows us to elaborate the DGP expression and calculate partial
derivatives with respect to changes in the explanatory variables in any time period on the dependent variable. This leads to
expressions for the own-period impact of changes in the explanatory variables taking the form in (16), where we note that
the N  N matrix of effects do not vary for values of t ¼ 1; . . . ; T. This is because our model assumes that the cross-section of
regions are identical across time as are the parameters br and q, which leads to the block diagonals of the matrix P 1 having
the same values. Of course, when the spatial dependence parameter q ¼ 0, we have the conventional partial derivative ðbr Þ
which is the same for all regions, reflecting a lack of dependence across the sample of regions.

@yt
ðrÞ
¼ B1 IN br ; t ¼ 1; . . . ; T ð16Þ
@xt
The result is an N  N matrix that shows how changes in each region’s explanatory variable value will impact the own-
and all other regions. This greatly complicates the task of interpreting parameter estimates from our model, so we follow
LeSage and Pace (2009) and propose scalar summary measures reflecting direct and indirect or spatial spillover effects. They
propose taking the sum of the rows (or columns) of the N  N matrix ðIN  qWÞ1 IN br and using an average to produce a sca-
ðrÞ
lar summary measure of the total impact that arises from changing xt on all N cross-sectional elements of yt (which is the
same for any time period in our model). The direct impact is summarized using the average of the diagonal elements of the
matrix ðIN  qWÞ1 IN br , and the difference between the total and direct effects reflect the indirect or spatial spillover effects.
They note that the indirect effects estimates represent a cumulation of the spillover effects falling on all neighbors, but de-
scribe how marginal effects can be calculated for the spillover effects falling on first-, second-, out to say m-order neighbors
using a partitioning scheme. This is possible using the fact that: ðIN  qWÞ1 ¼ IN þ qW þ q2 W 2 þ   , to produce a series of
marginal effects estimates (see LeSage and Pace, 2009).
In our dynamic panel model, changes to explanatory variable r at time t will have direct and indirect impacts on the own-
and other region dependent variable values at time t (captured by (16)), as well as impacts on both own- and other regions in
future time periods. This diffusion over space as time passes arises when the model includes non-zero time dependence cap-
tured by the parameter /. The partial derivatives for this situation are shown in (17) for the case where we change the
ðrÞ
explanatory variable x1 at time period 1, and measure the impacts at a one- through t-period horizon.

@yt
ðrÞ
¼ ðD1 þ D2 þ    þ Dt1 þ B1 ÞIN br ¼ ð/t1 þ /t2 þ    þ / þ 1ÞB1 IN br ð17Þ
@x1
We make a number of observations regarding the effects measures in (17). First, for longer time horizons the spatial dif-
fusion process will allow for a cumulation of: (1) the own-region effects from previous time periods (associated with /yt1 in
the model) and (2) the other-region effects from previous time periods (associated with the term hWyt1 in the model), and
the contemporaneous spatial effects (arising from the term qWyt in the model). The main diagonal matrices B1 in P 1 cap-
ture the contemporaneous spatial effects while the matrices Di ¼ /i  B1 capture the time and space–time diffusion effects.
For the case of the restricted model where h ¼ q  / so that there is space–time separability, diffusion over time and space
takes the form of time-discounting based on the time-dependence parameter /; /2 ; . . . of the contemporaneous spatial ef-
fects captured by the N  N matrix B1 .
This greatly simplifies interpretability of the dynamic responses for any number of time periods. Given estimates for the
parameters br ; q and / and the assumption that the matrix W is exogenous and fixed over time periods so that the spatial
configuration of the regions does not change, we can easily calculate dynamic responses for any number of time periods.
Third, we can calculate useful summary scalar measures that partition the total impacts into that portion attributable to
time dependence, spatial dependence, and space–time dependence (or diffusion). This can be done using a similar approach
to that of LeSage and Pace (2009) described for the case of a cross-sectional pure spatial dependence model.
ðrÞ
As a scalar summary measure of the direct spatial effects arising from a change in any period t of xt (spatial effects dont
change over time periods in this model), we use an average of the main diagonal from the matrix B1 IN br . This scalar
O. Parent, J.P. LeSage / Transportation Research Part B 44 (2010) 633–645 637

Table 1
An illustration of space, time and space–time diffusion effects estimates.

Effects Part A: q ¼ 0:2; / ¼ 0:8; h ¼ q  /


Direct Indirect Total
Spatial 0.0862 0.0135 0.0997
Time 1 0.0690 0.0108 0.0798
Time 2 0.0552 0.0087 0.0639
Time 3 0.0442 0.0069 0.0511
Time 4 0.0353 0.0055 0.0408
Time 5 0.0283 0.0044 0.0327
Time 6 0.0226 0.0035 0.0261
Time 7 0.0181 0.0028 0.0209
Time 8 0.0145 0.0023 0.0168
Time 9 0.0116 0.0018 0.0134
Cumulative total 0.3849 0.0603 0.4452

Effects Part B: q ¼ 0:8; / ¼ 0:2; h ¼ q  /


Spatial 0.1600 0.2739 0.4339
Time 1 0.0320 0.0548 0.0868
Time 2 0.0064 0.0110 0.0174
Time 3 0.0013 0.0022 0.0035
Time 4 0.0003 0.0004 0.0007
Time 5 0.0001 0.0001 0.0002
Time 6 0.0000 0.0000 0.0000
Time 7 0.0000 0.0000 0.0000
Time 8 0.0000 0.0000 0.0000
Time 9 0.0000 0.0000 0.0000
Cumulative total 0.2000 0.3424 0.5629

represents a summary measure of the direct spatial effects arising from a change in period t of the explanatory variable vec-
ðrÞ
tor xt on the own-region values of the dependent variable (where we average over all regional observations to produce the
scalar summary). Similarly, the indirect spatial effects can be summarized using an average of the off-diagonal of B1 IN br . This
scalar summary measure quantifies what are typically thought of as spatial spillover or other-region effects, where again we
ðrÞ
average the impact of changing explanatory variable vector xt on the other-region values of the dependent variable (where
we average over all regional observations to produce the scalar summary).
Scalar summary measures for the direct and indirect time effects can also be devised that measure decay over T periods of
ðrÞ
the own- and other-region effects on the dependent variable values arising from a change in any period t of variable xt at
time horizon T (period t þ T).
For the time horizon T direct time effects, which measures decay over T periods of the own-region effect arising from a
ðrÞ
change in period say t ¼ 1 of variable x1 , we use an average of the diagonal of:

ðD1 þ D2 þ    þ DT1 ÞIN br ¼ ð/ þ /2 þ    þ /T1 Þ  B1 IN br

as a scalar summary of these time horizon T effects. An average of the off-diagonal elements of:
ð/ þ /2 þ    þ /T1 Þ  B1 IN br serves as a scalar summary of the indirect time effects at horizon T, which measure diffusion
over time and space.2
As an illustration of these effects estimates, we generated a set of effects based on two different levels of spatial and time
dependence and a single parameter b ¼ 0:1, using a spatial weight matrix based on first-order contiguity of the lower 48
states plus the District of Columbia, which are shown in Table 1. Part A of the table shows scalar summary measures of
the effects estimates for spatial dependence ðq ¼ 0:2Þ that is relatively weaker than time dependence / ¼ 0:8, which leads
to larger time and space–time diffusion effects relative to the spatial effects. The values reported in the first row of the table
reflect cumulative spatial effects decomposed into direct, indirect and total effects. The direct effects correspond to own-par-
tial derivatives that measure the impact on region i from changes in the explanatory variable value of region i. However,
these include some feedback impacts discussed in LeSage and Pace (2009), since changes in region i influence the neighbors
and region i is in turn influenced by its neighbors. The indirect effects are cumulated over neighboring spatial regions, and
correspond to the cross-partial derivatives, and the final column shows the total effects which is the sum of the direct and
indirect effects. In our model, the spatial effects are separable from the time effects, and these do not change over time since
the spatial configuration of the regions remains the same and we restrict the spatial dependence parameter to be fixed over
all time periods.

2
For the case of more general models (where say the parameter q is allowed to change over time or the restriction that h ¼ q  / is not imposed), one could
still use the parameter estimates to numerically find the matrices Bi and Di ; i ¼ 1; . . . ; T that comprise the matrix P 1 in the scalar summary measures proposed
here.
638 O. Parent, J.P. LeSage / Transportation Research Part B 44 (2010) 633–645

The second through tenth rows of Part A of the table show the time period horizon effects again decomposed into scalar
summary measures of the direct, indirect and total effects. In this case, the scalar summary measures for direct effects reflect
the impact of changes made in region i in previous periods on region i in future periods, which arise from time dependence in
our dynamic model. The indirect effects scalar summaries capture the fact that changes in region i at one time period will
influence neighboring regions j – i in future periods due to the presence of terms like Wyt ; Wyt1 in our model. The scalar
summary measures for the direct effects were based on the average of the diagonal elements of the matrices Dt ; t ¼ 1; . . . ; 9
in (15), and the indirect effects were constructed from the off-diagonal elements of these matrices.
The last row in Part A of the table shows the cumulative sum of all rows, which would represent the cumulative (average)
impacts arising from a change to a single explanatory variable in (any) region i in period 0 over a 10 period time horizon. In
the case of Part A of the table where spatial dependence is relatively weak we see that less than 10% of the cumulative total
impact represents indirect or spatial spillover effects that arise over space and time diffusion. The (cumulative) time effects
represent the cumulative total effects minus the spatial effects, and in this scenario we see that over three-fourths of the
total 10 period horizon effects are attributable to time with less than one-fourth arising from spatial dependence in the
model.
Part B of Table 1 presents the scalar summary measures of effects for a scenario where spatial dependence is large
ðq ¼ 0:8Þ relative to weak time dependence ð/ ¼ 0:2Þ. In this case, we see relatively large indirect effects (spatial spillovers)
reported in the first row of Part B in the table, which are slightly less than twice the magnitude of the direct effects. While
this may seem counterintuitive, remember that the scalar summary measures of the spatial effects cumulate the indirect
effects over many neighboring regions, with the effect falling on individual neighbors being smaller than the direct effects
reported in the table (see LeSage and Pace, 2009). The other notable feature of the effects estimates shown in the weak time
dependence, strong spatial dependence scenario in Part B of the table is that the time effects are small and decay to zero very
quickly with increasing time horizons.

3. Estimation method

Dynamic panel data models with spatial effects are typically estimated using the Quasi-Maximum Likelihood (QML) esti-
mation (see Su and Yang, 2007; Yu et al., 2008). This estimation method is based on the maximization of a concentrated ver-
sion of the likelihood function that involves joint trivariate optimization over the parameters q; / and h. We note that the
complex form of the the disturbance covariance matrix leads to a concentrated likelihood expression which is not straight-
forward. Moreover the need to integrate over the random effects rules out simpler decomposition methods. However the
Bayesian MCMC estimation procedure based on conditional distributions allows us to estimate each parameter of interest
conditional on the others, which leads to some computational simplification. Convenient factorizations of the variance–
covariance matrix are available for each conditional distribution.
Bayesian estimation methods do not require integration over the random effects that appear in the likelihood. However
integration over these parameters can reduce serial dependence in the samples of parameters drawn (see Chib and Carlin,
1999).
A fully hierarchical Bayesian approach to estimation requires specification of the prior distributions for the parameters of
interest. Because this model will contain only proper prior distributions, the joint posterior distribution is proper. Bayesian
inference is based on the joint posterior distribution of the parameters given the data. Due to the hierarchical structure of
(10), the unnormalized form of the posterior distribution for the parameters takes the form:
         
p b; a; r2e r2l ; /; q; hjy / p yjc; l; r2e ; /; hq p ljr2l p r2l pðaÞpðbÞpðqÞp r2e pð/ÞpðhÞ ð18Þ

where the likelihood function is defined be pðyjÞ and prior distributions by pðÞ.
Direct evaluation of the joint posterior distribution involves multidimensional numerical integration and is not compu-
tationally feasible. In Appendix A, we set forth an MCMC sampling scheme which involves generating sequential samples
from the complete set of conditional posterior distributions for all parameters in the model.

4. An illustration of the model using travel-to-work times

Empirical analysis of commuting times typically relies on non-spatial models that assume observations for each region
are independent of those from neighboring regions (For an exception see Kirby and LeSage (2009)). Our illustrative applica-
tion assumes that commuting times exhibit both space and time dependence. Intuitively, road networks that transverse re-
gions would lead to increased traffic flows on the network having spillover impacts on observed travel or commuting times
from neighboring regions. There is an extensive literature on the determinants of commuting time (Kim, 2008; Van Omm-
eren and Fosgerau, 2009). One of the main challenges has been to explain variation in commute duration among metropol-
itan areas and its effect on urban sprawl.
While most studies focus on central-city or suburban averages (Song and Zenou, 2006), a few studies have analyzed the
impact of state and local spending on travel time. Even though commuting time is found to be largely heterogenous between
metropolitan areas, the state-level studies point to the same characteristics as important determinants. A novel aspect of our
empirical illustration is application of the dynamic space–time panel model in this setting.
O. Parent, J.P. LeSage / Transportation Research Part B 44 (2010) 633–645 639

Fig. 1. Mean travel times to work (averaged over the period 2000–2006).

We propose a log–linear model so that coefficients and effects estimates can be interpreted as elasticities. As previously
explained the impact of a (percentage) change in each explanatory variable on travel times can be decomposed using the
model structure and estimates into: (1) spatial direct and indirect effects and (2) time direct and indirect effects. The
sum of these constitute the total space–time dynamic impact arising from a (percentage) change in an explanatory variable.
The model introduces random effects to capture heterogeneity at the regional level.
Our empirical illustration examines the relationship between state-level average travel-to-work times, using journey-to-
work data taken from the 2000–2006 Census. The data represent a cross-sectional annual time series (panel data) for the 49
US states over the years 2000–2006.
The dependent variable values for mean travel times to work (TT) were taken from the American Community Survey con-
ducted by the US Census. This measure is based on the elapsed time that it takes each day to get from home to work during a
reference week. The average is calculated by dividing the aggregate travel time to work by the number of workers. Mean
travel times remained relatively constant increasing on average by 0.7% over the period 2000 and 2006. As depicted by
Fig. 1, the longest travel times are located on the east coast with New York, Maryland, New Jersey and the District of Colum-
bia showing averages of 31.01, 30.27, 29.3 and 29.01 min, respectively. The lowest average times are located in north central
states where we see that North Dakota, South Dakota, Montana, Nebraska, Wyoming, Iowa, had average times of 15.77,
15.97, 16.82, 17.14, 17.65 and 18.14 min respectively.
The primary data source for the remaining variables are the Highway Statistics series published by the US Department of
Transportation (DOT), Federal Highway Administration (FHWA). The number of lane-miles measures the size of the highway
network.3 Total lane mile growth of the national highway system over this 7 year period was about 2.25%. The FHWA classifies
highways into functional systems depending on their role in the public highway network. Arterial highways, collector facilities
and local streets and roads are used here to calculate the total length of lane miles. The variable (ane Miles) represents the num-
ber of lane miles per square mile for each state and measures the density of the highway network for each state.
In the year 2004, there were about four million miles of public roads in the United States and the total amount spent on
these roads was more than $140 billion.4 For some states (such as Minnesota, New York) roadway spending occurs mostly at
the local government level, whereas in other states (such as Delaware, West Virginia, North Carolina) this spending is primarily

3
US DOT (FHWA, Highway Statistics, 2000–2006), Section V: Roadway Extent.
4
US DOT, FHWA, Highway Statistics 2004, Section IV: Highway Finance.
640 O. Parent, J.P. LeSage / Transportation Research Part B 44 (2010) 633–645

Table 2
Estimation results – 20,000 iterations.

Variables/parameters Posterior mean s.d. Lower 0.05 Upper 0.95


Constant 0.3218 0.0744 0.2073 0.4501
Expenditures 0.0318 0.0121 0.0518 0.0117
Traffic Volume 0.0448 0.0070 0.0333 0.0561
Lane Miles 0.0058 0.0089 0.0202 0.0086
Gas Tax 0.0358 0.0188 0.0660 0.0048
r2e 0.0008 7.6e05 0.0007 0.0009
r2l 0.0098 0.0021 0.0069 0.0137
q 0.7919 0.0149 0.7730 0.8163
/ 0.1810 0.0106 0.1643 0.1981
h 0.1782 0.0169 0.2107 0.1553

by the state government. The national government accounts for a small amount (about 1%) of direct expenditures on roadways.
Total spending was calculated using an estimate of the total disbursements for highways for all units of government provided by
the FHWA.5 In order to eliminate size effects, we converted our expenditure measure (Expenditures) to a percentage of Gross
State Product (GSP), where GSP represents the value of all goods and services produced by labor and capital located in a state.
The GSP measure is taken from the Bureau of Economic Analysis (BEA).
At the state level, the primary source of highway revenues are motor vehicle fees and taxes, and highway-user fees,
including gasoline and diesel fuel taxes. In order to avoid collinearity between explanatory variables, we use tax rates on
gasoline (Gas Tax) as a proxy for these revenues.6 This tax rate varies greatly from state to state. For example, over the period
2000–2006, the average state excise tax on gasoline varied from a high of 29.4 cents per gallon in Rhode Island to a low of 7.5
cents in Georgia.
Less traffic will also reduce travel time to work. Congestion costs on rural highways are negligible when compared to
costs associated with urban highways. We use a measure of average daily traffic volume in urban areas (Traffic Volume)
as a proxy for congestion.
Based on Noland (2001), we model the mean travel time to work for state i at period t ðTT i;t Þ using the following regres-
sion relationship:

X
N X
N
logðTT i;t Þ ¼ a þ q wi;j logðTT j;t Þ þ / logðTT i;t1 Þ þ hwi;j logðTT j;t1 Þ
j¼1 j¼1

þ b1 logðExpendituresi;t Þ þ b2 logðTraffic Volumei;t Þ


þ b2 logðLane Milesi;t Þ þ b2 logðGas Taxi;t Þ þ git
git ¼ li þ eit ð19Þ

4.1. Estimation results

Table 2 reports the regression results based on a row-normalized first-order contiguity spatial weight matrix. Markov
Chain Monte Carlo estimation results were based on a simulated chain where the first 80,000 samples are discarded as a
‘burn-in’ period, followed by 20,000 iterations that were collected to produce posterior summaries for the model parameters.
As already noted, these coefficient estimates do not represent partial derivative effects describing how the dependent var-
iable in this model responds to changes in explanatory variable values (as in conventional regression models that assume
observations are independent). For this, we must calculate the scalar summary impact/effects measures along with measures
of dispersion which can be used to draw inferences regarding statistical significance. We note that the coefficient estimates
for all variables except the Lane Miles were statistically significantly different from zero as indicated by the lower and upper
95% posterior density intervals reported in Table 2. This result does not support the theory of induced demand that increas-
ing carrying capacity of a road network will result in increased demand for travel. From the estimates for the time and spatial
dependence parameters we see that we have a case of large spatial dependence relative to weaker time dependence, similar
to the scenario presented in Part B of Table 1.
The effects estimates are reported in Table 3 for each variable in the model. Since the spatial effects predominate we re-
port these decomposed into direct, indirect and total as well as the cumulative total effects. The small differences between
the cumulative total effects and the spatial effects reflect the time effects. For example, in the case of highway expenditures
(as a percentage of GSP) we see a 0.059 direct cumulative effect value and a direct spatial effect of 0.0488, so the differ-
ence of 0.0102 represents cumulative direct time effects (which we calculated over a 10 year horizon).7 In comparison with
the coefficient estimate of 0.0318 from Table 2 for this variable, the direct effects estimate reported in Table 3 includes a feed-

5
US DOT, (FHWA, Highway Statistics, 2000–2006), Section IV: Highway Finance.
6
The correlation coefficient between expenditure (as a percentage of GSP) and the tax rate on gasoline equals 0.137.
7
Due to the small time dependence coefficient, this horizon captures almost all of the time effects which decay towards zero rapidly.
O. Parent, J.P. LeSage / Transportation Research Part B 44 (2010) 633–645 641

Table 3
Scalar summary estimates of the space–time effects.

Lower 0.05 Median Upper 0.95


Expenditures variable
Spatial effects
Direct 0.083407 0.048807 0.010809
Indirect 0.156660 0.092173 0.021737
Total effects 0.233411 0.140924 0.032546
Cumulative effects
Direct 0.102342 0.059171 0.012973
Indirect 0.204370 0.111191 0.026572
Total effects 0.298933 0.172271 0.039546
Traffic Volume variable
Spatial effects
Direct 0.053139 0.074144 0.096082
Indirect 0.092764 0.135430 0.187322
Total effects 0.145902 0.207579 0.283065
Cumulative effects
Direct 0.062497 0.089022 0.117899
Indirect 0.099455 0.165021 0.250217
Total effects 0.160431 0.254916 0.368116
Lane Miles variable
Spatial effects
Direct 0.037872 0.008424 0.017995
Indirect 0.066108 0.016733 0.032451
Total effects 0.103980 0.025159 0.050335
Cumulative effects
Direct 0.047036 0.010378 0.021320
Indirect 0.086128 0.019928 0.037217
Total effects 0.133164 0.031151 0.058536
Tax variable
Spatial effects
Direct 0.129011 0.058787 0.002307
Indirect 0.248453 0.118148 0.004175
Total effects 0.369817 0.178068 0.006483
Cumulative effects
Direct 0.143939 0.067821 0.002619
Indirect 0.232245 0.111850 0.003924
Total effects 0.375942 0.177749 0.006544

back loop that arises in our space–time-dynamic panel model. An implication is that even interpreting the coefficients from
Table 2 would lead to incorrect inferences regarding the true (own-partial derivative) impacts that arise from changes in the
explanatory variables for this model.
The total effect associated with increasing highway expenditures (measured here as a percentage of GSP) is 0.172
suggesting that (on average across all states and a 10 year time horizon) a 100% increase (doubling) in state-level high-
way expenditures would reduce travel times to work by 17.2% cumulatively over space and time. Of this travel time
reduction, 5.9% (the cumulative direct effect) would be captured by the own-state increasing highway expenditures
and 11.1% (the cumulative indirect effect) would spillover to other states and in future time periods. Of course, for
an individual neighboring state the spillover benefit in travel-to-work time reduction would likely be less than the direct
own-state benefit of 5.9%, consistent with intuition that spatial spillovers represent a second-order effect. However, recall
we are cumulating the spillover benefits over many neighboring states and accounting for space–time diffusion over a
10 year time horizon. We note that state spending on highways averaged around 1.2% of GSP on average across the time
periods and states in our sample. So, the 100% increase envisioned here would require around 1% more of GSP to be
spent on highways.
For states like New York, Maryland, New Jersey and the District of Columbia where the average travel times are around
30 min, the direct effect estimate of 5.9% would mean a 1 min 40 s reduction in travel-to-work times. The 11.1% spillovers
cumulated over neighboring states and the 10 year time horizon amounts to a 3 min 20 s reduction in travel times. So, the
total effect associated with the 17.2% reduction would mean around 5 min reduction in travel-to-work times associated with
the relatively large doubling of highway expenditures, spread over commuters located in space and over future time periods.
For the western states where travel-to-work times average around 15 min this would represent around 2 and one-half min-
utes cumulative total reduction in travel-to-work times.
642 O. Parent, J.P. LeSage / Transportation Research Part B 44 (2010) 633–645

Table 4
Estimation results for the random effects – 20,000 iterations.

Region Posterior mean s.d. Lower 0.05 Upper 0.95


Alabama 0.051 0.018 0.082 0.022
Arizona 0.044 0.020 0.011 0.077
Arkansas 0.058 0.018 0.088 0.028
California 0.039 0.021 0.004 0.073
Colorado 0.089 0.019 0.059 0.120
Connecticut 0.041 0.020 0.073 0.009
Delaware 0.017 0.022 0.053 0.018
Dist. of Col. 0.134 0.035 0.077 0.193
Florida 0.029 0.021 0.063 0.004
Georgia 0.018 0.028 0.028 0.064
Idaho 0.010 0.021 0.025 0.044
Illinois 0.133 0.020 0.100 0.165
Indiana 0.064 0.019 0.095 0.034
Iowa 0.109 0.018 0.139 0.079
Kansas 0.097 0.018 0.127 0.067
Kentucky 0.041 0.018 0.071 0.011
Louisiana 0.033 0.018 0.004 0.063
Maine 0.009 0.020 0.025 0.041
Maryland 0.150 0.018 0.120 0.181
Massachusetts 0.048 0.020 0.015 0.080
Michigan 0.012 0.019 0.044 0.020
Minnesota 0.121 0.018 0.091 0.151
Mississippi 0.009 0.018 0.021 0.038
Missouri 0.043 0.019 0.013 0.074
Montana 0.030 0.023 0.008 0.067
Nebraska 0.083 0.019 0.114 0.052
Nevada 0.046 0.023 0.007 0.084
New Hampshire 0.078 0.029 0.031 0.126
New Jersey 0.065 0.024 0.026 0.105
New Mexico 0.030 0.023 0.067 0.008
New York 0.121 0.020 0.089 0.153
North Carolina 0.052 0.018 0.082 0.022
North Dakota 0.066 0.021 0.100 0.033
Ohio 0.089 0.020 0.121 0.057
Oklahoma 0.077 0.018 0.107 0.048
Oregon 0.059 0.019 0.092 0.028
Pennsylvania 0.075 0.021 0.109 0.041
Rhode Island 0.012 0.023 0.049 0.026
South Carolina 0.069 0.019 0.099 0.038
South Dakota 0.059 0.020 0.093 0.025
Tennessee 0.030 0.018 0.060 0.000
Texas 0.005 0.022 0.042 0.031
Utah 0.021 0.020 0.055 0.012
Vermont 0.079 0.021 0.114 0.044
Virginia 0.013 0.019 0.018 0.044
Washington 0.121 0.018 0.091 0.152
West Virginia 0.102 0.021 0.067 0.137
Wisconsin 0.088 0.019 0.120 0.057
Wyoming 0.039 0.023 0.077 0.001

Consistent with findings by Shirley and Winston (2004), our results confirm that increasing highway infrastructure
spending can produce economic benefits by reducing travel times to work. A desirable aspect of our model is the ability
to quantify direct and indirect (spatial spillover) benefits associated with changes in such expenditures.
The effects estimates for the Traffic Volume variable are positive suggesting an increase in this variable would lead to long-
er travel-to-work times. Specifically, we see that a doubling (a 100% increase) of traffic volume would lead to a 25% cumu-
lative total impact on travel times, with 20% of this 25% total impact representing direct and indirect spatial effects and the
remaining 5% being diffusion over space and time. Of the total spatial effect, spillovers or in this case congestion effects aris-
ing from the interconnected highway network represent two-thirds of the total effects. Using the long and short travel-to-
work times of 30 and 15 min, this implies that a doubling in traffic volume would raise travel times by 7.5 min and 3.75 min
respectively.
Changes in the Lane Miles variable exhibit a negative but statistically insignificant (using the 0.05 and 0.95 intervals) im-
pact on travel-to-work times. This is true of the spatial as well as time and cumulative total effects, as well as the decom-
position of these into direct and indirect effects components.
The Gas Tax variable exerts a negative and significant impact on travel-to-work times, having a total cumulative effect of
0.177. This is similar in magnitude to the highway expenditures variable, suggesting that a doubling of taxes on gasoline
would lead to a reduction in travel times of 17.7%. Almost all of this total cumulative effect represents spatial effects which
O. Parent, J.P. LeSage / Transportation Research Part B 44 (2010) 633–645 643

Fig. 2. Estimation for the random effects (insignificant effects are in white).

have a value of 17.80 compared to 17.77 for the cumulative total effect. This implies that the space–time diffusion effects
have a small positive magnitude of 0.03 that accounts for the slightly smaller cumulative total effect relative to the spatial
effect. Of course, since motorists in border areas are free to purchase gas in a neighboring state and many commuters cross
state borders during their commute to work, we might expect to see such large spatial effects as well as large indirect spatial
effects that capture the spillover impact that would accompany a single state raising its gasoline tax. There is some literature
that considers uniform mileage taxes versus fuel taxes (Parry and Small, 2005), concluding that mileage taxes would reduce
mileage-related externalities more over the long-run because vehicle fuel economy responds to gasoline prices (which of
course includes taxes). Our model’s ability to decompose the direct and indirect impacts would allow these types of issues
to be considered in a context where interdependencies exist between mileage traveled in neighboring regions, which could
shed additional light on these relationships.
A final comment regards use of a space–time dynamic model that includes a simultaneous spatial lag of the dependent
variable ðWyt Þ as well as a conventional time lag ðyt1 Þ and space–time lagged dependent variable ðWyt1 Þ. LeSage and Pace
(2009) provide a host of alternative motivations for these types of models (see Chapters 2 and 7). One of these motivations is
that commuters will take time to adjust their travel-to-work habits in response to changes in things like highway congestion
(traffic volume), highway infrastructure (lane miles and highway expenditures), and gasoline taxes (see Pace and LeSage,
forthcoming, Chapter 7). As in the case of costs of adjustments in time–dynamic models, where time lags of the dependent
variable appear as a result of such adjustment costs, simultaneous and space–time lags of the dependent variable appear
when our model allows for spatial as well as time dependence.

4.2. Random effects estimates

Turning to the random effects estimates, these are presented in Table 4 and depicted in Fig. 2.8 Results show strong po-
sitive effects in the Western and Northeastern regions. Maryland (0:150) and the District of Columbia ð0:134Þ have the largest
positive values. As explained in Appendix A, the posterior distribution for the random effects l is based on the difference be-
tween the space–time filter of the dependent variable Y and the explanatory variables: u ¼ Y  qðIT  WÞY  /Y 1  hðIT
WÞY 1  Xb  iNT a. The combination of very high commuting times for a region and its neighbors along with high levels of

8
Note that we just report the 95% highest posterior density (HPD) interval for the random effect estimates. We do not perform any test to reveal whether or
not each of these estimates is different from zero (see Scott and Berger (2006), for the multiple testing problem).
644 O. Parent, J.P. LeSage / Transportation Research Part B 44 (2010) 633–645

gasoline taxes or low levels of highway expenditures (as percentage of GSP) will result in large positive values for these random
coefficients. In fact the District of Columbia has the lowest share of expenditure on highways (less than 0.5% of its GSP on aver-
age over the 7 year period) as compared to an average around 1.2% and high values for states like Montana whose share is more
than 2.5% of its GSP. Maryland has one of the largest tax rates on gasoline (23.5%), compared to a national average of 20.8% over
the period. Similar observations can be made for the states of Washington and West Virginia. On the other hand, despite the low
average travel times to work, central states like Iowa and Kansas have a very high share of their GSP allocated to expenditures
on highways (1.61% and 1.39%, respectively). This could explain the negative values for the random effects seen for these states.
An important point is that the random effects can be used for more than the typical purpose of controlling for regional
heterogeneity. Since we have an estimate for the posterior distribution of each region’s random effect, valid inference regard-
ing significance of these effects can be carried out. Given our discussion surrounding interpretation of the effects as reflecting
differences between the space–time filter for the dependent variable and predictions made by the mean part of the model
(the linear combination of explanatory variables and their associated coefficient estimates), one might also rely on such
inferences as a model diagnostic.

5. Conclusion

In this paper we set forth a Bayesian MCMC approach to estimation of a spatial dynamic panel data model with random
effects, and illustrate the model in an analysis of state-level mean travel-to-work times. The model extends time–dynamic
panel models to allow for both simultaneous spatial dependence as well as space–time diffusion of effects. We interpret the
model specification as representing a space–time filter which in the absence of (significant or substantive) spatial depen-
dence reflects a classic space–time autoregressive (STAR) model with random effects parameters (for each region in the mod-
el). In the absence of (significant or substantive) time dependence, our model collapses to a simultaneous spatial panel
model with random effects parameters (for each region in the model).
An implication of our model specification is that changing the value of an explanatory variable in one region at a single time
period can impact other regions in the same time period as well as the same and other regions in future time periods. A focus
was therefore proper interpretation of these spatial, time and space–time-diffusion impacts. The ability to partition the impacts
into direct and indirect effects using an extension of the scalar summary measures proposed by LeSage and Pace (2009) should
have valuable policy implications. For example, an increase in highway expenditures in a single region will have direct impacts
on travel times in the region itself and spatial spillover impacts on neighboring regions due to the nature of transportation net-
works that are shared across regions. There will also be future impacts that arise as a result of diffusion over time to the own and
neighboring regions. Since a change in each observation in each time period can potentially impact all other regions and all fu-
ture time periods, we have a partial derivative matrix for the impacts of dimension NT  NT associated with each explanatory
variable in the model. For this reason, we follow LeSage and Pace (2009) and develop scalar summary measures of the impacts
based on averaging the diagonal and off-diagonal elements associated with sub-matrices of the lower triangular block matrix
that represents these impacts. These can be interpreted as own-region or direct impacts that include feedback effects arising
from the interdependent structure of the model as well as cross-region and time-diffusion impacts that arise from both spatial
and time dependence. Our scalar summary measures of direct, indirect and total spatial effects, direct, indirect and total time
effects allow one to interpret estimation results in a fashion more typical of that used in regression analysis.
To illustrate these interpretative issues we adopt a log–linear relationship leading to effects estimates that represent elas-
ticities. Using a dataset on mean travel times for the 48 continental US states and the District of Columbia over the period
2000–2006, we estimate a relation between mean travel times to work, highways expenditures and lane miles, gasoline
taxes and traffic volumes. Our results illustrate how one would go about interpreting the impacts arising from changes in
such things as gasoline taxes, expenditures on highway infrastructure or traffic volumes. We find evidence of substantial
spatial spillovers and relatively weaker time dependence leading to much smaller time impacts accruing over future periods
when initial period impacts could diffuse to neighboring regions.

Appendix A. We suppose that the prior distribution pðc; r2u ; q; /; hÞ is such that all parameters are a priori independent.
Concerning the parameters c ¼ ða; b0 Þ, we estimate separately the intercept term a and the parameters b assuming a non-
hierarchical prior of the independent Normal-Gamma variety. Thus,

 
a  N a0 ; M1
a ð20Þ
 
b  N b0 ; M 1
b

e  Gðv 0 =2; S0 =2Þ


r2 ð21Þ
l  Gðv 1 =2; S1 =2Þ
r2
We use diffuse priors so that the location parameters a0 and b0 centered on zero, with precision parameters M 1 1
a and M b
12
equal to 10 . Parameters for the Gamma priors are all set to 0.001, and we adopt uniform priors for /; q and h over the
interval (1, 1).
O. Parent, J.P. LeSage / Transportation Research Part B 44 (2010) 633–645 645

Having the posterior distribution of the explanatory variables b conditional on the random effects l is not desirable be-
cause these two sets of parameters tend to be highly correlated which can create problems with mixing for the Markov Chain
estimation procedure. We use the method proposed by Chib and Carlin (1999) who suggest first sampling b marginalized
over l and then sampling l conditioned on b.
Posterior distributions are standard and can be found in Koop et al. (2007, Chapter 12) or LeSage and Pace (2009).

References

Baltagi, B.H., Song, S.H., Jung, B.C., Koh, W., 2007. Testing for serial correlation, spatial autocorrelation and random effects using panel data. Journal of
Econometrics 140 (1), 5–51.
Chib, S., Carlin, B.P., 1999. On MCMC sampling in hierarchical longitudinal models. Statistics and Computing 9 (1), 17–26.
Cressie, N.A.C., 1993. Statistics for Spatial Data. John Wiley & Sons, New York.
FHWA, 2000–2006. FHWA, Highway Statistics 2000–2006 Federal Highway Administration. US Department of Transportation, Washington, DC.
Garrido, R., 2000. Spatial interaction between the truck flows through the Mexico–Texas border. Transportation Research Part A 34 (1), 2333.
Kim, C., 2008. Commuting time stability: a test of a co-location hypothesis. Transportation Research Part A 42 (3), 524–544.
Kirby, D.K., LeSage, J.P., 2009. Changes in commuting to work times over the 1990 to 2000 period. Regional Science & Urban Economics 39 (4), 460–471.
Koop, G., Poirier, D.J., Tobias, J.L., 2007. Bayesian Econometric Methods. Cambridge University Press.
LeSage, J., Pace, R.K., 2009. An Introduction to Spatial Econometrics. CRC Press, Taylor-Francis.
Noland, R., 2001. Relationships between highway capacity and induced vehicle travel. Transportation Research Part A 35 (1), 4772.
Pace, R.K., LeSage, J.P., forthcoming. Spatial econometrics. In: Gelfand, A.E., Diggle, P., Fuentes, M., Guttorp, P. (Eds.), Handbook of Spatial Statistics. Chapman
& Hall.
Parent, O., LeSage, J.P., 2008. A Space–time Filter for Panel Data Models Containing Random Effects. Working Paper.
Parent, O., LeSage, J.P., 2009. Spatial Dynamic Panel Data Models With Random Effects. Working Paper.
Parry, I.W.H., Small, K.A., 2005. Does Britain or the United States have the right gasoline tax? American Economic Review 95 (4), 1276–1289.
Pfeifer, P.E., Deutsch, S.J., 1980. A three-stage iterative procedure for space–time modeling. Technometrics 22 (1), 35–47.
Scott, J.G., Berger, J.O., 2006. An exploration of aspects of Bayesian multiple testing. Journal of Statistical Planning and Inference 136 (7), 2144–2162.
Shirley, C., Winston, C., 2004. Firm inventory behavior and the returns from highway infrastructure investments. Journal of Urban Economics 55 (2), 398–
415.
Song, Y., Zenou, Y., 2006. Property tax and urban sprawl: theory and implications for US cities. Journal of Urban Economics 60 (4), 519–534.
Su, L., Yang, Z., 2007. QML Estimation of Dynamic Panel Data Models with Spatial Errors. Singapore Management University Manuscript.
Van Ommeren, J., Fosgerau, M., 2009. Workers marginal costs of commuting. Journal of Urban Economics 65 (1), 38–47.
Yu, J., de Jong, R., Lee, L.F., 2008. Quasi-maximum likelihood estimators for spatial dynamic panel data with fixed effects when both n and T are large. Journal
of Econometrics 146 (1), 118–134.

Anda mungkin juga menyukai