Spatial Econometric Modeling of

Origin-Destination flows

James P. LeSage1
University of Toledo
Department of Economics
Toledo, OH 43606
[email protected]
and
R. Kelley Pace
LREC Endowed Chair of Real Estate
Department of Finance
E.J. Ourso College of Business Administration
Louisiana State University
Baton Rouge, LA 70803-6308
OFF: (225)-388-6256, FAX: (225)-334-1227
[email protected], www.spatial-statistics.com

October 18, 2005

1
The authors would like to acknowledge Randall W. Jackson and Wolfgang
Polasek for helpful discussions on this topic.
 Abstract

The traditional gravity model used to provide econometric estimates of vari-

ables influencing origin-destination flows is extended to explicitly take into
account spatial dependence in the flows. This is accomplished by intro-
ducing a spatial connectivity matrix that allows for three types of spatial
dependence in the flows from origins to destinations. Introducing this type
of connectivity or spatial weight structure for the flows allows conventional
spatial econometric estimation procedures to be used in modeling variation
in flows that arise when the origin-destination flow matrix is vectorized. A
family of alternative spatial econometric model specifications is set forth
along with an applied illustration based on state-level migration flows for
the 48 contiguous US states and the District of Columbia.

KEYWORDS: migration flows, spatial autoregression, Bayesian, max-

imum likelihood, spatial connectivity of origin-destination flows.
1 Introduction
This paper sets forth spatial econometric methods for modeling origin-
destination (OD) matrices containing interregional flows. These are gen-
eral data structures used in a variety of economic, geography and regional
science research contexts such as international trade flows, migration re-
search, transportation, network, and freight flow analysis, communications
and information flow research, journey-to-work studies, and regional and
interregional economic modeling. In contrast to typical spatial economet-
ric models where the sample involves n regions, with each region being an
observation, these models involve n2 OD pairs.
The term spatial interaction models has been used in the literature to
label models that focus on flows between origins and destinations, Sen and
Smith (1995). An objective of this type of modeling is to explain variation in
the level of flows between the n2 OD pairs. These models typically rely on a
function of the distance between an origin and destination and explanatory
variables pertaining to characteristics of both origin and destination regions.
They typically assume that spatial dependence between the sample of N 2
OD pairs will be captured by the distance function. With a few exceptions,
use of spatial lags typically found in spatial econometric methods have not
been used in these models. There has been widespread recognition of the
need for such models in disciplines such as population migration. Cushing
and Poot (2003, p. 317) provide a survey of migration research in which
they state that:
As noted in the introduction, no one has as yet seriously exploited
the potential of spatial econometrics in the migration literature. This
would seem to be a natural extension for migration research and one
with potentially greater importance at greater levels of geographic dis-
aggregation. A more complete consideration of the spatial dimension
in migration research is one of the key contributions that regional sci-
ence can make to this literature.
Questions surrounding how to parsimoniously structure the connectivity
of the sample of n2 origin-destination pairs that arise in modeling inter-
regional flows has remained a stumbling block. Conventional spatial au-
toregressive models rely on spatial weight structures constructed to reflect
connectivity between n regions. One focus of the presentation here is a
proposal for spatial weight structures that model dependence between the
n2 OD pairs in a fashion consistent with conventional spatial autoregressive
models.

1
 The notion that use of distance functions in conventional spatial interac-
tion models effectively capture spatial dependence in the interregional flows
being analyzed has been challenged in recent work by Porojon (2001) for the
case of international trade flows, Lee and Pace (2004) for retail sales and
unpublished work that utilizes both German and Canadian transportation
network flows. The residuals from conventional models were found to ex-
hibit spatial dependence, which could be exploited to improve the precision
of inference as well as prediction accuracy.
A family of successive spatial filtering models is introduced here that
represent an extension of the spatial regression models introduced in Anselin
(1988). Spatial regression models have served as the workhorse in applied
spatial econometric analysis, and the models introduced here should play
an important role in modeling interregional flow matrices. Another focus
of this study is maximum likelihood and Bayesian estimation of the models
introduced here. We demonstrate how simple extensions of widely available
software algorithms for implementing conventional spatial regression models
can be employed to estimate the models set forth here.

2 Interregional flows in a spatial regression con-

text
Let Y represent an n by n square matrix of interregional flows from each
of the n origin regions to each of the n destination regions where each of
the n columns of the flow matrix represents a different origin and the n
rows reflect destinations. We can produce an n2 by 1 vector of these flows
in two ways, one reflecting an origin-centric ordering as in (1), and the
other reflecting a destination-centric ordering as in (2). We obtain the
origin-centric ordering via y = vec(Y ), and the destination-centric ordering
via y (d) = vec(Y 0 ). These two orderings are related by the vec-permutation
matrix so that P y = y (d) , and by the properties of permutation matrices
y = P 1 y (d) = P 0 y (d) . For most of the discussion, we will focus on the
origin-centric ordering where the first n elements in the stacked vector y
reflect flows from origin 1 to all n destinations. The last n elements of this
vector represent flows from origin n to destinations 1 to n.1
1
Typically, the diagonal elements of a flow matrix containing flows within a region,
e.g., from origin 1 to destination 1, origin 2 to destination 2, etc., will be large relative
to the off-diagonal elements representing interregional flows. In fact, many of the interre-
gional flows will take on zero values, indicating the absence of flows from some origins to
particular destinations. In the following discussion, we ignore these issues with discussion

2
 l(o) o(o) d(o)
1 1 1
.. ..
. 1 .
n 1 n
.. .. ..
. . . (1)
n2 n + 1 n 1
.. .. ..
. . .
n 2 n n

l(d) o(d) d(d)

1 1 1
.. .. ..
. . .
n n 1
.. .. ..
. . . (2)
n2 n + 1 1 n
.. .. ..
. . .
n 2 n n

A conventional gravity model least-squares regression approach to ex-

plaining the variation in the vector of origin-destination flows would rely
on an n by k explanatory variables matrix that we label xd , containing k
characteristics for each of the n destinations. Given the format of our vector
y, where observations 1 to n reflect flows from origin 1 to all n destinations,
this matrix would be repeated n times to produce an n2 by k matrix of
destination characteristics that we represent as Xd for use in the regression.
Each vector j of Xd equals Xj , where is a n by 1 vector of ones. A
second matrix containing origin characteristics which we label Xo would be
constructed for use in the gravity model. This matrix would repeat the
characteristics of the first origin n times to form the first n rows of Xo , the
characteristics of the second origin n times for the next n rows of Xo and
so on, resulting in an n2 by k matrix of origin characteristics. Similarly,
each vector j of Xo is Xj . Typically, the distance from each origin to
destination is also included as an explanatory variable vector in the gravity
model, and perhaps non-linear terms such as distance-squared. We let D
represent an n2 by 1 vector of these distances from each origin to each des-
tination formed by stacking the columns of the origin-destination distance
and suggestions for dealing with these taken up later.

3
matrix into a variable vector.2 This results in a regression model of the type
shown in (3)

y = + Xd d + Xo o + D + (3)
In (3), the explanatory variable matrices Xd , Xo represent n2 by k ma-
trices containing destination and origin characteristics respectively and the
associated k by 1 parameter vectors are d and o . The vector D denotes
the vectorized origin-destination distance matrix, and , are scalar pa-
rameters. For now we assume N (0, 2 In2 ), but generalizations will be
taken up later.

2.1 Spatial dependence in origin-destination flows

In contrast to the traditional regression-based gravity model, a spatial econo-
metric model of the variation in origin-destination flows would be charac-
terized by: 1) reliance on spatial lags of the dependent variable vector,
which we refer to as a spatial autoregressive model (SAR); 2) spatial lags
of the disturbance terms, which we label a spatial error model (SEM); or 3)
perhaps spatial lags of both kinds, which we denote as the general spatial
model (SAC). Spatial weight matrices represent a convenient and parsimo-
nious way to define the spatial dependence or connectivity relations among
observations.
In a typical cross-sectional model with n regions where each region rep-
resents an observation, the spatial weight matrix labelled W represents an
n by n sparse matrix. This matrix captures dependency relations between
the observations which represent regions. In this conventional case, the rows
i = 1, . . . , n of the matrix W are specified using the set of neighboring ob-
servations to each observation i. If we designate neighboring observations
using i , spatial dependence arises when an observation at one location, say
yi is dependent on neighboring observations yj , yj i . If Wij represents
the individual elements of the matrix W , then Wij > 0 when observation
yj i , that is yi depends upon yj . By convention, Wii = 0 to prevent an
observation from being defined as a neighbor to itself, and the matrix W is
typically row-standardized to have row sums of unity.
A key issue is how to construct a meaningful spatial weight matrix in
the case where the n2 by 1 vector of observations reflect flows from all ori-
gins to all destinations, rather than the typical case where each observation
2
The diagonal elements of the distance matrix containing distances from origin 1 to
destination 1, origin 2 to destination 2, etc., will be zero. We will have more to say about
this later.

4
represents a region. We can create a typical n by n first-order contiguity or
m nearest neighbors weight matrix W that reflects relations between the n
destinations/origin regions. This can be repeated using In W to create an
n2 by n2 row-standardized spatial weight matrix that we label Wo , shown
in (4), where 0 represents an n by n matrix of zeros.

W 0 ... 0
..
0 W 0 .

Wo = .. .. (4)

. 0 . 0
0 ... 0 W
Using this matrix to form a spatial lag of the dependent variable, Wo y,
(where Wo = In W with W row-standardized), we capture origin-based
spatial dependence relations using an average of flows from neighbors to each
origin region to each of the destinations. Intuitively, it seems plausible that
forces leading to flows from any origin to a particular destination region may
create similar flows from neighbors to this origin to the same destination.
This is what the spatial lag Wo y captures.
As an example, consider a single row i of the spatial lag vector Wo y
that represents flows from the origin state/region of Florida to the desti-
nation state/region of Washington. First-order contiguous neighbors to the
origin Florida are Alabama and Georgia, and neighbors to the destination
Washington are Oregon and Idaho. The spatial lag Wo y would represent an
average of the flows from Alabama and Georgia (neighbors to the origin) to
the destination state Washington.
A similar interpretation applies to other rows of the spatial lag Wo y.
For example when examining flows from the origin state of Alabama to
the destination state of Washington, the spatial lag Wo y would represent
an average of the flows from Florida, Georgia, Mississippi and Tennessee
(neighbors to the origin) to the destination state Washington.
A second type of spatial dependence that could arise in the gravity model
would be destination-based dependence. Intuitively, it seems plausible
that forces leading to flows from an origin state to a destination state may
create similar flows to nearby or neighboring destinations.
A spatial weight matrix that we label Wd can be constructed to capture
this type of dependence using W In , producing an n2 by n2 spatial weight
matrix that captures connectivity relations between the flows from an origin
state to neighbors of the destination state.
To provide an example of this we consider four regions located in a row
as presented in Table 1.

5
 Table 1: Location of 4 Regions in Space

Region #1 Region #2 Region #3 Region #4

The row-standardized first-order contiguity matrix associated with this re-

gional configuration is shown in (5).

0 1 0 0
1/2 0 1/2 0

W = (5)
0 1/2 0 1/2
0 0 1 0
For this example, Wd = W In takes the form shown in (6), where 0
represents an n by n matrix of zeros, and n = 4 in this example.

0 In 0 0
(1/2)In 0 (1/2)In 0

Wd = (6)
0 (1/2)In 0 (1/2)In
0 0 In 0
Using our example of flows from the origin state of Florida to the desti-
nation state of Washington, the spatial lag vector Wd y represent an average
of flows from Florida to Idaho and Oregon, states that neighbor the destina-
tion state of Washington. In our other example, where the origin state was
Alabama, the spatial lag vector Wd y would represent an average of flows
from Alabama to Idaho and Oregon, states that neighbor the destination
state of Washington.
To provide a more formal development of destination based dependence,
we employ the vec-permutation matrix P introduced previously. If we
adopted the destination-centric ordering, specification of the destination
weight matrix would be I W by the same logic as introduced in the de-
velopment of the origin weight matrix. Consequently, a destination weight
matrix under the origin-centric ordering would be P 0 (I W )P . Some re-
sults on Kronecker products lead to a further simplification of P 0 (I W )P .
Given that P is the vec-permutation matrix, by Corollary 4.3.10 in Horn
and Johnson (1991, p. 290), W I = P 0 (I W )P , and thus Wd = W I.
A third type of dependence to consider is reflected in the product Ww =
Wo Wd = (In W ) (W In ) = W W . This spatial weight matrix
reflects an average of flows from neighbors to the origin state to neighbors

6
of the destination state. One motivation for this matrix product might be
a spatial filtering perspective. We might envision a spatial autoregressive
model of the type shown in (7) based on successive filtering. We transform or
filter the dependent variable successively by (In2 1 Wo ), and (In2 2 Wd ).
The motivation is that we are removing destination dependence first and
subsequently origin dependence, or vice-versa.

(In2 1 Wo )(In2 2 Wd )y = + Xd d + Xo o + D + (7)

This leads to a model that includes the interaction term Ww = Wo Wd in
the sequence of spatial lags:

y = 1 Wo y + 2 Wd y 1 2 Ww y + + Xd d + Xo o + D + (8)

Using our example of flows from the origin state of Florida to the desti-
nation state of Washington, the spatial lag vector Ww y represent an average
of: flows from Alabama and Georgia (neighbors to the origin state) to Idaho
(a neighbor to the destination state), and flows from Alabama and Geor-
gia (neighbors to the origin state) to Oregon (a neighbor to the destination
state). In the case of our other example based on flows from the origin state
of Alabama to the destination state of Washington, the spatial lag vector
Ww y represent an average of: flows from Florida, Georgia, Mississippi and
Tennessee (neighbors to the origin state) to Idaho (a neighbor to the des-
tination state) and flows from Florida, Georgia, Mississippi and Tennessee
(neighbors to the origin state) to Oregon (a neighbor to the destination
state).
Note, the implementation of this does not require the actual formation of
the n2 by n2 matrices Wo , Wd , or Ww . Given arbitrary, conformable matrices
A, B, C, (C 0 A)vec(B) = vec(ABC) (Horn and Johnson, 1991, p. 255,
Lemma 4.3.1). Since Wo y = (I W )~(Y ), then Wo y = vec(W Y ). Similarly,
Wd = vec(Y W 0 ), and Ww = vec(W Y W 0 ). These expressions also aid in the
interpretation of origin-destination dependence. The algebra of Kronecker
products can be used to form moment matrices without dealing directly
with n2 by n2 matrices. For example, Xd0 i Xoj equals ( Xi )0 (Xj ) =
P P
Xi Xj . Also, the moment sub-matrices involving only origin variables
or destination variables are very simple (Xd0 i Xdj = Xo0 i Xoj = nXi0 Xj ).
In concluding we note that spatial lags involving the disturbance process
could also be constructed using weight matrices Wo , Wd and the product Ww .
This would allow for a model where spatial dependence arises in the error

7
terms of the model. In the successive filtering case this would take the form
in (9).

y = + Xd d + Xo o + D + u (9)
u = (In2 1 Wo )(In2 2 Wd )u +

2.2 Spatial model specifications for origin-destination flows

A family of nine different model specifications is proposed. We take as
a starting point a slight generalization of the successive spatial filtering
specification from (7) and (8), shown in (10) for the SAR model. The
generalization in (10) stems from relaxing the restriction from (8) that 3 =
1 2 .

y = 1 Wo y + 2 Wd y + 3 Ww y + + Xd d + Xo o + D + (10)

The generalized successive filtering model specification in (10) involves

the origin, destination weight matrices and their cross-product, along with
no restrictions on the parameters 1 , 2 and 3 . We set forth nine different
models that can be derived by placing various restrictions on the parameters
i , i = 1, . . . , 3. Since the statistical theory for testing parameter restrictions
is well-developed, this seems desirable from an applied specification search
viewpoint.

1. The restriction: 1 = 2 = 3 = 0, produces the least-squares model

where no spatial autoregressive dependence exists.
2. The restriction: 2 = 3 = 0, results in a model based on a single
weight matrix Wo , reflecting origin autoregressive spatial dependence.
3. The restriction: 1 = 3 = 0, produces a sibling model based on a
single weight matrix Wd for spatial dependence at the destinations.
4. The restriction: 1 = 2 = 0, creates another single weight matrix
model containing only Ww , reflecting dependence based on interaction
between origin and destination neighbors.
5. The restriction: 1 = 2 , 3 = 0, results in a model based on a single
weight matrix constructed using Wo + Wd . This reflects a lack of
separability between the impacts of origin and destination dependence
relations in favor of a cumulative impact.

8
 6. The restriction: 1 = 2 , 3 = 21 = 22 , produces another single
weight matrix model based on Wo + Wd + Ww . This reflects a lack
of separability between the impacts of origin, destination and origin-
destination interaction effects in favor of a cumulative impact.
7. The restriction: 3 = 0, leads to a model with separable origin and
destination autoregressive dependence embodied in the two weight ma-
trices Wo and Wd , while ruling out dependence between neighbors of
the origin and destination locations that would be captured by Ww .
8. The restriction: 3 = 1 2 results in a successive filtering or model
involving both origin Wo , and destination Wd dependence as well as
product separable interaction Ww , constrained to reflect the filter:
(In2 1 Wo )(In2 2 Wd ) = (In2 + 1 Wo + 2 Wd 1 2 Ww ).
9. No restrictions produces the ninth member of the family of models
based on an unrestricted variant of the filter: (In2 1 Wo )(In2
2 Wd ) = (In2 + 1 Wo + 2 Wd + 3 Ww )
Each of the single spatial weight matrix model specifications in 1) to 6)
would obey the usual properties of row-normalized weight matrices, allowing
use of existing algorithms for maximum likelihood (Pace and Barry, 1997),
Bayesian (LeSage, 1997) or generalized method of moments estimation esti-
mation (Kelejian and Prucha, 1999).
We note that specifications 1) to 6) based on single weight matrices are
also amenable to variants of spatial regression models of the type shown in
(11) to (13), which we label SAR, SEM and SAC models, respectively. In
these equations, we use Wj to denote the single spatial weight matrix.3

y = Wj y + + Xd d + Xo o + D + (11)
y = + Xd d + Xo o + D + u (12)
u = Wj u +
y = Wj y + + Xd d + Xo o + D + u (13)
u = Wj u +
We note that use of conventional algorithms for maximum likelihood,
Bayesian or generalized method of moments estimation of the spatial econo-
metric origin-destination interregional flow models becomes difficult as the
3
There are widely available algorithms for estimation of these alternative specifications,
e.g., the spatial econometrics toolbox, www.spatial-econometrics.com and spatial statistics
toolbox, www.spatial-statistics.com.

9
number of observations increases. For example, use of an origin-destination
flow matrix for the sample of approximately 3,100 US counties would re-
sult in sparse spatial weight matrices of dimension n2 by n2 where n2 =
9, 610, 000. Maximum likelihood and Bayesian estimation both require cal-
culation of the logged determinant for the n2 by n2 matrix (In2 Wj ).
While specialized approaches to calculating log-determinants of very large
matrices have been proposed by Pace and LeSage (2004) and Smirnov and
Anselin (2001), it turns out there are much more efficient approaches that
can exploit the special structure of matrices like Wd = In W , Wo = W In
and Ww = Wo Wd = W W . We turn attention to this topic in the next
section.

3 Estimation of spatial flow models

As already noted, SAR, SEM and SAC model specifications based on a
spatial weight matrix Wj = Wo , Wd , Ww , as well as sums of these such as,
Wj = Wo + Wd , Wj = Wo + Wd + Ww , can be implemented using standard
algorithms. For a model based on n = 50 states, this results in n2 =
2, 500, a situation where the log-determinant calculation as well as all other
calculations required for maximum likelihood or Bayesian estimation can be
completed rapidly (LeSage and Pace, 2004). Despite this, the structure of
the matrices Wo , Wd and Ww allow for computational improvements.

3.1 The case of a single weight matrix Wk , k = o, d

We note that the concentrated log-likelihood function for model specifica-
tions 1) to 6) based on a single spatial weight matrix, which we denote Wj ,
concentrated with respect to the parameters and will take the form:

LogL() = C + log|In2 Wj | (n2 /2)log(e0 e()) (14)

Where e0 e() represents the sum of squared errors expressed as a function
of the scalar parameter alone after concentrating out the parameters , ,
and C denotes a constant not depending on (see LeSage and Pace, 2004).
The log-determinant of a matrix plays an important role in both maxi-
mum likelihood and Bayesian estimation of transformed random variables.
Specifically, the log-determinant ensures that the transformed continuous
random variable has a proper density. Otherwise, multiplication of a de-
pendent variable by a transformation such as I, where is a small positive

10
number, would reduce the magnitude of the estimation residuals to a neg-
ligible level. The log-determinant term serves as a penalty to prevent such
pathological transformations from obtaining an advantage in estimation.
Consequently, the likelihood is invariant to such scalings.
The log-determinant of the transformation is the trace of the matrix
logarithm of the transformation, and the Taylor series expansion of this has
a simple form for the positive definite matrix transformation In2 Wj ,
shown in (15).
t
X tr(Wjt )
ln |In2 Wj | = tr (ln(In2 Wj )) = (15)
t=1
t
For the case of destination or origin weight matrices, Wd = In W or
Wo = W In , which we designate Wk , k = o, d,

tr(Wkt ) = tr(Int W t ) = tr(Int ) tr(W t ) = n tr(W t ), (16)

and thus the trace of a square matrix of order n2 is simplified to a scalar
(n) times a trace of the n by n square matrix W .
t
X tr(Wkt )
ln |In2 Wk | = n = n ln |In W | (17)
t=1
t
Summarizing, for the case of a single spatial weight matrix Wk , k =
o, d, it should be possible to rely on algorithms for computing the logged
determinant of an n by n matrix ln |In W |, when working with a vector of
n2 origin-destination flows. For the earlier example of n = 3, 100 US counties
and n2 = 9, 610, 000, we should be able to solve these estimation problems
in a matter of seconds on desktop computers when using computationally
efficient sparse algorithms for the n by n log-determinant portion of the
problem (see LeSage and Pace, 2004).

3.2 Calculating logged determinants for the filtering models

7) to 9)
For the more general successive spatial filtering model specifications 7) to 9),
we observe that the order in which one transforms the dependent variable
makes no difference. The explanation for this can be seen by considering
the cross products in (18). The mixed-product rule for Kronecker products
indicates that the cross-product of W In and In W is W W , which is
the same as the cross-product of In W and W In .

11
 (In2 1 Wo )(In2 2 Wd ) = (In2 1 (In W ))(In2 2 (W In ))
= In2 1 (In W ) 2 (W In ) (18)
+1 2 (W W )

Because the log-determinant of a product is the sum of the log-determinants,

the overall log-determinant arising from the successive filtering approach is
quite simple as shown in (19).

ln |(In2 1 Wo ) (In2 2 Wd )| = n(ln |In 1 W | + ln |In 2 W |) (19)

As in the case of single spatial weights and conventional model specifi-

cations from the previous section, the logged determinant required for max-
imum likelihood estimation of successive filtering models can be calculated
using only n by n matrices rather than the large n2 by n2 matrices.
Drawing on the earlier discussion surrounding (17), the estimation chal-
lenge for the case of the most general spatial filtering model 9), with all
three parameters 1 , 2 and 3 unrestricted is to easily compute tr(Wft ) for
t = 1 m, where m is the largest moment computed, and Wf is defined in
(20).

Wf = 1 (In W ) + 2 (W In ) + 3 (W W ) (20)
The case of tr(Wf ) where t = 1 is immediate, and equals zero since
tr(W ) = 0. The case of tr(Wf2 ) is slightly more challenging as shown in
(21).

Wf2 = 21 (In W 2 ) + 22 (W 2 In ) + 23 (W 2 W 2 ) (21)

2 2
+ 21 2 (W W ) + 21 3 (W W ) + 22 3 (W W )

For the quadratic, there are 9 possible terms and 6 of these are unique.
Note, tr(Wf2 ) is the highest order term associated with W . Extrapolating,
computations of tr(Wft ) only require computing tr(W t ) based on the n by
n weight matrix W , a much less demanding task. Individual terms have the
form in (22).

i1 j2 k3 tr Woi Wdj Wwk = i1 j2 k3 tr(W (i+k) )tr(W (j+k) ) (22)

12
 Given a table of tr(W t ) for t = 1 m, each term involves the multiplica-
tion of five scalars. However, there are 3m terms, and this becomes difficult
for large m. Other than computing tr(W t ), none of these computations are
dependent upon n, and so it takes just as long for a problems with many
origins and destinations as for smaller problems. For small n, calculating
exact tr(W t ) requires little time. For large n, calculating tr(W t ) can be
approximate as in Barry and Pace (1999) who show how to do this with an
O(n) algorithm.
Given the m moments and the conditions on W , it becomes easy to
compute an relatively short interval containing the log-determinant as shown
in (23).

m tr(W t )
X Xm tr(W t )
X tr(W m)
f f f
ln |In2 Wf | + (23)
t=1
t t=1
t t=m+1
t

Pace and LeSage (2002) show how the moments tr(Wft ) must monotonically
decline for t > 1, and this sets up the bounds. The interval is narrow
provided (1 + 2 + 3 )m+1 /(m + 1) is reasonably small. A requirement for
stability is that 1 + 2 + 3 < 1, making this a reasonable presumption.
Summarizing, we derived a family of nine model specifications that em-
phasize different spatial connectivity relations between origin and destina-
tion regions. Since members of the family of specifications reflect models
based on parameter restrictions, these can be easily tested to draw infer-
ences regarding the nature of spatial dependence in any applied problem.
Potential computational problems that might plague estimation for models
involving n2 observations on origin-destination flows were eliminated by re-
ducing the troublesome logged determinant calculation to one involving only
traces of n by n matrices. As already noted, successive filtering of the type
described here could also be applied to the disturbance process, producing
a family of nine models of the type we have labelled SEM, or to both the
dependent variable and disturbance vectors resulting in nine more models
of the type we labelled SAC.

4 An applied illustration using state-level popula-

tion migration flows
To illustrate the family of spatial econometric models described here we use
state-level population migration flows as the dependent variable. Specifi-
cally, the growth rates in migration flows for the population 5 years and

13
over from the period covering 1985 to 1990 and flows for the period 1995 to
2000 were used.4 The sample was restricted to the 48 contiguous states plus
the District of Columbia resulting in n = 49 and n2 = 2, 401 observations.
The growth rates for flows of population within each state were set to zero
to emphasize flows between states which should exhibit the type of spatial
dependence of interest here.
Another benefit of the growth rates transformation is alleviation of the
problem noted earlier that arises with flows that are very large within re-
gions relative to many zero values for interregional flows. Figure 1 shows
a histogram of the annualized growth rates in the flows alongside a normal
probably density plot. From the figure we see some evidence of fat tails
reflecting more extremely large or small growth rates than one would expect
in a normal distribution. We also see the impact of setting 49 within-region
flows to zero values. An approach to dealing with the fat-tailed nature of
the distribution of flows during estimation will be illustrated in Section 5.
Explanatory variables for the matrices Xo , Xd for each state were taken
from the 1990 Census, with the exception of the unemployment rate variable,
which was constructed as the ratio of state-level unemployment rates in 1995
to 1990. These variables are documented in Table 2.
The motivation for including the age variables near retirement, and re-
tired is that these should exert an impact on migration decisions. Apriori,
we would expect that retired would increase flows from the origin, whereas
near retirement should decrease flows from both destinations and origins.
We note that an increase in flows is indicated by a positive coefficient esti-
mate and a decrease by a negative estimate.
Population that lived in another state in 1985 might increase flows at
both the origin and destination as this is an indicator of population mobility.
The effect of foreign born population seems unknown apriori, depending on
the mobility of this population relative to the average.
Population holding graduate and professional degrees should be the most
mobile, leading to increased flows at both the origin and destination, whereas
persons with less than ninth grade education should be less mobile. The
impact of associate degrees, college degrees, and sales jobs is less clear.
Rents and unemployment rates should increase flows at the origin and
decrease flows at the destination, whereas per capita income should decrease
flows at the origin and increase flows at the destination. The area variable
4
Available on the internet State-to-State Migration Flows: 1995 to 2000 Cen-
sus 2000 Special Reports. The data are based on a sample. For information
on confidentiality protection, sampling error, nonsampling error, and definitions, see
http://www.census.gov/prod/cen2000/doc/sf3.pdf.

14
was included to control for the impact of variation in the size of the states
on migration growth rates.
In addition to the variables included in the matrices Xo , Xd , the log of
distance from each origin to each destination was included in the model,
along with a constant term.
The family of nine model specifications described in Section 3 were es-
timated using maximum likelihood methods with a numerical hessian ap-
proach used to compute estimates of dispersion and tstatistics. The log-
likelihood function values for the family of nine models are shown in Table 3,
ordered from high to low, along with a likelihood ratio (LR) test of the re-
strictions imposed by each model versus the unrestricted model. It is clear
from the table that the Models 7, 8 and 9 based on the filtering specification
that contains separate spatial weight matrices for the origin and destination
provide a significantly higher likelihood than Models 2 through 6 that use a
single spatial weight matrix. There is also a noticeable drop in the likelihood
values when going from models 5 and 6 to models models 2, 3 and 4. We
note that models 5 and 6 are based on single weight matrices constructed
by summing information from both origin and destination weight matrices,
whereas to models 2, 3 and 4, are based on only and origin or only destina-
tion, or only the interaction weight matrices. This would seem to support
the notion that both origin and destination dependence/connectivity infor-
mation are important.
The LR tests indicate that the Model 7 restriction 3 = 0 does not
significantly reduce the likelihood function value. This restriction eliminates
the weight matrix Ww reflecting connectivity between neighbors to the origin
and neighbors to the destination. We might interpret this result as indicating
these relations are relatively unimportant in explaining growth rates in the
migration flows over our time period. The parameter estimates for the
unrestricted model are presented in Table 4, where we see that the parameter
3 is not significantly different from zero, consistent with the LR test results.
The LR test result for Model 8 based on the restriction that 3 = 1 2 ,
versus the unrestricted model rejects this restriction as consistent with the
sample data, at the 95 percent level.
Finally, it is clear that least-squares which ignores spatial dependence in
the growth rates of the migration flows and assumes these are independent
produces a much lower likelihood function value.
Turning to the parameter estimates from least-squares and the unre-
stricted spatial Model 9, shown in Table 4, we see estimates for 1 = 0.313
and 2 = 0.280, indicating spatial dependence of equal importance between:
neighbors to the origin and the destination, and neighbors to the destina-

15
tion and the origin. As indicated above, the estimate for 3 = 0.0072 is
not significantly different from zero, allowing us to infer that dependence
between neighbors to the origin and neighbors to the destination specified
by the weight matrix Ww is not important.
Turning attention to the parameter estimates, we see that distance is
positive and significant in the least-squares model, but insignificant in the
spatial model. (Distance was insignificant in all 8 spatial models.) Typically,
regression-based gravity models produce a negative influence of distance
on the flows, which seems intuitively appealing, whereas this is not the
case here. We note that if the true data-generating process was in fact a
model containing a spatial lag, then least-squares estimates are biased and
inconsistent (see LeSage and Pace, 2004).
From Table 4, we see many cases where least-squares estimates (in ab-
solute value terms) are larger than those from the spatial model, which is
typical of least-squares, since it attributes variation assigned to the spatial
lags of the dependent variable by the spatial model to explanatory variables.
For example, spatial model estimates for: D nearretirement, D diffstate,
D rents, D associate O college, O gradprof O rents, and O unemp, take on
values around one-half those from least-squares, while D unemployment is
an exception.
Three of the variables exhibit a change from significantly different from
zero to insignificant between the spatial and least-squares models: D sales,
O sales, and distance are all insignificant in the spatial model. Three other
variables change in the level of significance: D college, D unemp, have a
higher level of significance in the spatial model and O unemp has a lower
level of significance.
The estimates from the spatial filtering model indicate that: higher rents,
higher unemployment and more associate degrees lead to higher growth rates
in origin flows, but lower destination flows, as might be expected. Per capita
incomes also have the expected positive impact on destination flows, with
an insignificant impact on origin flows.
Persons near retirement (aged 60 to 64) reduces flows at both the ori-
gin and destination, as do college graduates. Persons with graduate and
professional degrees increase flows at both the origin and destination as do
persons that lived in a different state in 1985 than in 1990. This suggest
these groups reflect higher than average mobility, which seems plausible. In
contrast, foreign-born population reduces flows at both the origin and desti-
nation, suggesting lower mobility. Retired persons (aged 65 to 74) increase
flows only at the origin, having an insignificant impact on the destination.
Finally, area exerts a positive impact on flows at both the origin and

16
destination suggesting that states with larger physical areas exhibit higher
growth rates in migration flows when controlling for other factors.

5 Extensions of the general spatial filtering model

to accommodate special issues
Some issues that were raised earlier regarding modeling of vectorized origin-
destination flow matrices were: 1) the fat-tailed nature of the distribution of
the vectorized flow matrix, even after transformation to growth rates over
time (as illustrated by Figure 1 for our state-level migration flows exam-
ple); 2) the presence of numerous zeros for large flow matrices as in the
case of migration flows between US counties, reflecting a lack of interaction
between numerous regions in the sample; and 3) the presence of large flows
on the diagonal of the flow matrix reflecting a large degree of intra-regional
connectivity;
Bayesian estimation procedures for conventional spatial models of the
type labelled SAR, SEM and SAC here have been set forth in LeSage (1997)
and for probit and tobit variants of these models in LeSage (2000). One im-
portant aspect of these estimation procedures which rely on Markov Chain
Monte Carlo methods (MCMC) (Gelfand and Smith, 1990) is that they can
accommodate sample y vectors that exhibit fat-tails, or follow a Student
tdistribution, rather than the conventional normal distribution. This sug-
gests that these methods might be able to overcome problems associated
with issue 1) above. Another feature of Bayesian MCMC estimation of
these models is that tobit extensions to accommodate sample censoring in
the y vector are relatively straightforward to implement, allowing a possible
solution to issues raised by 2) above. Regarding issue 3) above, it may be
the case that a two-regime, hierarchical or mixture model could be used
to describe intraregional variation in flows (within regions) separately from
interregional flows (between regions). There is a large literature on use of
Bayesian mixture and hierarchical spatial models that might be applicable
to issue 3) above (see Besag, York, and Mollie, 1991, Besag and Kooperberg,
1995, and Cressie 1995).
We illustrate one of these possible extensions by presenting estimates
based on a robust variant of the SAR model presented in LeSage (1997) for
our state-level migration flow growth rates. This involves Bayesian MCMC
estimation of a model that takes the same form as the general spatial fil-
tering model, but relaxes the constant variance assumption regarding the
disturbances in the data generating process. In place of N (0, 2 In2 ), we

17
assume that:

N [0, 2 diag(V )] (24)

V = v1 , v2 , . . . , vn2

To produce estimates for the n2 variance scalars in (24), we follow an

approach introduce by Geweke (1993), that places a 2 (r) prior on the
variance scalars vi with a mean of unity and a mode and variance that
depend on the hyperparameter r of the prior. Small values of r around 5
result in a prior that allows for the individual vi estimates to be centered
on their prior mean of unity, but deviate greatly from the prior value of
unity in cases where the model residuals are large. Large residuals are
indicative of outliers or origin-destination combinations that are atypical or
aberrant relative to the majority of the sample of origin-destination flows.
Geweke (1993) points to the equivalence of this modeling approach and the
assumption of disturbances that follow a Student tdistribution.
The MCMC estimation method samples sequentially from the complete
set of conditional distributions for all parameters in the model. Sampling
from the conditional distributions for the parameters and , when unin-
formative priors are assigned to these, is relatively straightforward as they
take known distributional forms. The conditional distribution for the o , d
parameters take the form of a kvariate multivariate normals, and that for
the parameter is a 2 (n2 ) distribution.5 In this model we must sample the
three parameters 1 , 2 , 3 as well as the variance scalars vi , i = 1, . . . , n2 .
The conditional distribution for the parameter 1 , conditional on the re-
maining parameters = (, , V, 2 , 3 ) take the form the form shown in
(25), with the conditionals for 2 and 3 taking similar forms.

p(1 |) = log|In2 1 Wo 2 Wd 3 Ww | (25)

2
n k
log{[e(1 , 2 , 3 )0 V 1 e(1 , 2 , 3 )]/(n2 k)}
2
e(1 , 2 , 3 ) = y 1 Wo 2 Wd 3 Ww Xo o Xd d D

Where we note the presence of the logged determinant term as in the case
of maximum likelihood estimation. We can rely on the same algorithms
for rapidly evaluating this expression in the context of Bayesian MCMC
estimation as in maximum likelihood. Sampling for the parameters i , i =
5
See LeSage (2004), pp. 232-233 for the exact expressions needed here.

18
1, 2, 3 is accomplished using expressions similar to (25) in a Metropolis-
Hastings algorithm based on a tuned normal random-walk proposal.
Table 5 presents the posterior means and highest posterior density (HPD)
intervals based on 0.05 and 0.95 percentiles for the parameters of the most
general spatial filtering model. Maximum likelihood estimates are also in-
cluded in the table to facilitate comparison. We see 19 parameters whose
posterior mean is different from zero based on the 0.95 HPD intervals, which
contrasts with 18 such parameters from maximum likelihood estimation us-
ing the 95 percent level of significance. Differences arise for four destination
characteristics: D sales, D foreignborn, D grade9 and D associate; and two
origin characteristics: O nearretirement, and O sales.
There are also differences in the magnitudes of the parameter estimates,
even when both Bayesian and maximum likelihood estimates are signifi-
cantly different from zero. For example, the MCMC estimate of D sales is
twice as large as maximum likelihood, whereas maximum likelihood esti-
mates for O college and O gradprof are about twice those from the robust
Bayesian model. and O rents is twice as large.
Finally, we see evidence of stronger spatial dependence in larger posterior
mean estimates for the parameters 1 and 2 . We note that the 0.05 and
0.95 HPD intervals for these two parameters do not include the maximum
likelihood estimates, suggesting a substantial increase in spatial dependence
when we account for non-constant variance.
Turning attention to the variance scalar estimates, these provide a diag-
nostic for observations (or OD pairs) that do not conform well to the model
relationship. Large estimates for these variances suggest an outlier or aber-
rant observation. Of the 2,401 observations, 2,233 of the posterior mean
values for the scalars vi had values of 3 or less, and 2,354 values of 6 or less.
There were 17 vi estimates whose posterior means exceeded a value of 10,
indicative of exceptionally high or low migration growth rates that could not
be explained by the models origin and destination variables/characteristics
or the spatial autoregressive dependence structure.
Table 6 shows the origin-destination state pairs for the 17 cases where the
posterior mean vi values exceeded 10 along with the mean vi estimate. These
observations would be downweighted by the inverse of the vi values during
MCMC estimation of the robust model. This is in contrast to the maximum
likelihood estimation procedure where all observations are assigned equal
weight. This accounts for the differences between the Bayesian and max-
imum likelihood estimates. From the table, we see that origin-destination
state pairs identified as outliers conform to intuition, reflecting the growth
rates in migration flows from mostly small states to other small states. This

19
is where we would expect to see large variances in the growth rates of mi-
gration flows over the 1985-90 and 1995-2000 periods, likely due to volatility
that arises in growth rates calculated on the basis of small levels of flows.

6 Conclusions
We set forth a method for incorporating spatial autoregressive structures in
conventional regression-based gravity models. This extension allows a family
of conventional spatial regression models that explicitly model the spatial
dependence structure between cross-sectional observations to be employed
in modeling origin-destination flows.
The approach introduced here allows for application of conventional spa-
tial regression algorithms for estimation and inference in the case of small
samples. In addition, we provide a solution for more realistic cases where
large samples involving flows between US counties numbering nearly 10 mil-
lion observations can be estimated in a matter of seconds. This requires only
slight modification to existing algorithms. Much of what we have learned
about maximum likelihood and Bayesian estimation of spatial regression
models can be immediately applied to origin-destination flow modeling.
As an extension to conventional spatial autoregressive and spatial error
models, we introduce a family of models that subsume these conventional
models as a special case. Simple tests of parameter restrictions that produce
varying specifications for spatial dependence can be carried out, resolving
contentious model specification issues that often arise.
Three special issues that arise in origin-destination flow modeling were
discussed: 1) the fat-tailed nature of the distribution of the vectorized flow
matrix; 2) the presence of numerous zeros for large flow matrices reflecting a
lack of interaction between numerous regions in the sample; and 3) the pres-
ence of large flows on the diagonal of the flow matrix reflecting a large degree
of intra-regional connectivity. Solutions to some of these problems may be
possible by drawing on past work from conventional spatial econometrics.
As an illustration, robust Bayesian Markov Chain Monte Carlo estimates
for the model introduced in this study were presented. These estimates ac-
commodate the fat-tailed nature of the distribution of the vectorized flow
matrix.

20
7 References
Besag, J. E., York, J.C. and Mollie, A. (1991). Bayesian image
restoration, with two applications in spatial statistics (with discus-
sion), Annals of the Institute of Statistical Mathematics, Vol. 43, pp.
1-59.

Besag, J. E. and Kooperberg, C.L. (1995). On conditional and in-

trinsic autoregressions, Biometrika, Vol. 82, pp. 733-746.

Cressie, N. (1995). Bayesian smoothing of rates in small geographic

areas, Journal of Regional Science, Vol. 35, pp. 659-673.

Cushing, Brian, and Jacques Poot (2003), Crossing Boundaries and

Borders:: Regional Science Advances in Migration Modelling, Papers
in Regional Science, Vol. 83, pp. 317-338.

Gelfand, Alan E., and A.F.M Smith. (1990), Sampling-Based Ap-

proaches to Calculating Marginal Densities, Journal of the American
Statistical Association, Vol. 85, pp. 398-409.

Geweke J. (1993), Bayesian Treatment of the Independent Student t

Linear Model, Journal of Applied Econometrics, Vol. 8, pp. 19-40.

Horn, R.A., and C.R. Johnson (1991). Topics in Matrix Analysis,

Cambridge: Cambridge University Press.

LeSage, J.P. (1997) Bayesian Estimation of Spatial Autoregressive

Models, International Regional Science Review, Volume 20, number
1&2, pp. 113-129.

LeSage, J.P. (2000) Bayesian Estimation of Limited Dependent Vari-

able Spatial Autoregressive Models, Geographical Analysis, Volume
32, number 1, pp. 19-35.

LeSage, J.P. (2004) Spatial Regression Models, in Numerical Issues

in Statistical Computing for the Social Scientist, John Wiley & Sons,
Inc., Micah Altman, Jeff Gill and Michael McDonald (eds.), pp. 199-
218.

LeSage, J.P. and R. Kelley Pace (2004) Introduction Advances in

Econometrics: Volume 18: Spatial and Spatiotemporal Econometrics,
(Oxford: Elsevier Ltd), pp. 1-32.

21
Lee,Ming-Long and R. Kelley Pace (2005) Spatial Distribution of
Retail Sales, Journal of Real Estate Finance and Economics, Volume
31, number 1, pp. 53-69.

Pace, R.K. and James P. LeSage (2002), Semiparametric Maximum

Likelihood Estimates of Spatial Dependence, Geographical Analysis,
January 2002, Volume 34, Number 1, pp. 75-90.

Pace, R.K, and J.P. LeSage (2004) Techniques for Improved Approx-
imation of the Determinant Term in the Spatial Likelihood Function,
Computational Statistics and Data Analysis, 2004, Volume 45, pp.
179-196.

Porojan, A. (2001) Trade Flows and Spatial Effects: The Gravity

Model Revisited, Open Economic Review, Volume 12, pp. 265-280.

Sen, Ashish and Tony E. Smith (1995), Gravity Models of Spatial In-
teraction Behavior, Heidelberg: Springer-Verlag.

Smirnov, O. and L. Anselin (2001). Fast Maximum Likelihood Esti-

mation of Very Large Spatial Autoregressive Models: a Characteristic
Polynomial Approach. Computational Statistics and Data Analysis
35, 301-319.

22
 Table 2: Explanatory variables used in the model

Variable name Description

young log (population aged 22-29/population in 1990)
near retirement log (population aged 60-64/ population in 1990)
retired log (population aged 65-74/ population in 1990)
sales log (proportion of work force in sales in 1990)
diffstate log (proportion of population living in a different state in 1985)
foreign born log (proportion of foreign born population in 1990)
grade9 log (< 9th grade as highest degree in 1990/population > age 25)
associate log (associate degree in 1990/population > age 25)
college log (college as highest degree in 1990/population > age 25)
grad prof log (graduate or professional degree in 1990/population > age 25)
rents log (median rent in 1990)
unemp ratio of 1995 unemployment rate to 1990 unemployment rate
pc income log (per capita income in 1990)
area log (1990 state area in square miles)

Table 3: Log Likelihoods for alternative models

Model Log Likelihood LR test versus Critical Value

Model 1 ( = 0.05)
Model 9: 1 , 2 , 3 unrestricted 605.1585
Model 7: 3 = 0 605.1245 0.0680 2 (1) = 3.84
Model 8: 3 = 1 2 602.9628 4.3914 2 (1) = 3.84
Model 5: 1 = 2 , 3 = 0 597.7079 14.9012 2 (2) = 5.99
Model 6: 1 = 2 , 3 = 21 582.5549 45.2072 2 (2) = 5.99
Model 2: 2 = 3 = 0 534.8818 140.5534 2 (2) = 5.99
Model 3: 1 = 0, 3 = 0 516.1341 178.0488 2 (2) = 5.99
Model 4: 1 = 0, 2 = 0 469.8518 270.6134 2 (2) = 5.99
Model 1: 1 = 0, 2 = 0, 3 = 0 388.5083 433.3004 2 (3) = 7.82

23
Table 4: Estimates from least-squares and the unrestricted spatial Model 9

Spatial model Least-squares

Variable Coefficient t-statistic(plevel) Coefficient t-statistic(plevel)
constant -7.2281 -11.230 (0.0000) -11.4183 -5.503 (0.0000)
D nearretirement -0.5407 -2.465 (0.0138) -0.8669 -3.236 (0.0012)
D retired 0.1420 1.001 (0.3166) 0.1848 1.096 (0.2729)
D sales 0.1245 1.152 (0.2491) 0.2438 2.022 (0.0432)
D diffstate 0.3454 2.498 (0.0125) 0.7137 4.724 (0.0000)
D foreignborn -0.3464 -1.580 (0.1140) -0.1772 -0.733 (0.4633)
D grade9 -0.0817 -3.443 (0.0006) -0.1013 -3.886 (0.0001)
D associate -0.1329 -3.620 (0.0003) -0.1995 -4.918 (0.0000)
D college -0.1371 -1.976 (0.0482) -0.1289 -1.659 (0.0971)
D gradprof 0.0289 0.508 (0.6110) -0.0265 -0.416 (0.6768)
D rents -0.5884 -4.360 (0.0000) -0.9086 -6.205 (0.0000)
D area 0.0126 1.549 (0.1213) 0.0258 2.867 (0.0041)
D unemp -0.7892 -3.444 (0.0006) -0.6102 -2.465 (0.0137)
D pcincome 0.4928 3.314 (0.0009) 0.6564 3.880 (0.0001)
O nearretirement -0.4955 -2.695 (0.0071) -0.8755 -3.268 (0.0010)
O retired 0.4520 3.377 (0.0007) 0.7546 4.476 (0.0000)
O sales -0.1505 -1.373 (0.1697) -0.2365 -1.961 (0.0498)
O diffstate 0.8783 6.160 (0.0000) 1.6001 10.593 (0.0000)
O foreignborn -1.3203 -5.908 (0.0000) -1.7875 -7.400 (0.0000)
O grade9 0.0163 0.697 (0.4856) 0.0375 1.437 (0.1506)
O associate 0.0718 1.913 (0.0558) 0.1969 4.856 (0.0000)
O college -0.2834 -4.232 (0.0000) -0.5823 -7.497 (0.0000)
O gradprof 0.1971 3.502 (0.0005) 0.3167 4.976 (0.0000)
O rents 0.6938 5.446 (0.0000) 1.2287 8.391 (0.0000)
O area 0.0456 5.486 (0.0000) 0.0571 6.342 (0.0000)
O unemp 0.4999 2.308 (0.0211) 1.2287 4.964 (0.0000)
O pcincome -0.1371 -1.256 (0.2091) -0.3004 -1.775 (0.0758)
log(distance) 0.0018 0.492 (0.6226) 0.0084 2.172 (0.0299)
1 0.3135 13.105 (0.0000)
2 0.2800 11.198 (0.0000)
3 -0.0072 -0.172 (0.8627)
2 0.0675 0.0819

24
Table 5: Estimates from maximum likelihood and robust Bayesian versions
of spatial Model 9

Bayesian MCMC Maximum Likelihood

Variable Posterior Lower Upper Coefficient t-statistic
mean 0.05 HPD 0.95 HPD (plevel)
constant -5.1139 -7.6374 -2.7329 -7.2281 -11.230 (0.0000)
D nearretirement -0.4835 -0.8174 -0.1512 -0.5407 -2.465 (0.0138)
D retired 0.1346 -0.0674 0.3472 0.1420 1.001 (0.3166)
D sales 0.2328 0.0954 0.3757 0.1245 1.152 (0.2491)
D diffstate 0.2122 0.0204 0.3861 0.3454 2.498 (0.0125)
D foreignborn -0.4218 -0.7004 -0.1506 -0.3464 -1.580 (0.1140)
D grade9 -0.0309 -0.0619 0.0001 -0.0817 -3.443 (0.0006)
D associate -0.0455 -0.0957 0.0015 -0.1329 -3.620 (0.0003)
D college -0.1222 -0.2137 -0.0329 -0.1371 -1.976 (0.0482)
D gradprof 0.0589 -0.0174 0.1334 0.0289 0.508 (0.6110)
D rents -0.5811 -0.7516 -0.4058 -0.5884 -4.360 (0.0000)
D area -0.0003 -0.0115 0.0109 0.0126 1.549 (0.1213)
D unemp -0.7896 -1.0800 -0.4954 -0.7892 -3.444 (0.0006)
D pcincome 0.4612 0.2689 0.6606 0.4928 3.314 (0.0009)
O nearretirement -0.2168 -0.5271 0.0966 -0.4955 -2.695 (0.0071)
O retired 0.2413 0.0446 0.4425 0.4520 3.377 (0.0007)
O sales -0.1673 -0.3067 -0.0153 -0.1505 -1.373 (0.1697)
O diffstate 0.7839 0.5884 0.9744 0.8783 6.160 (0.0000)
O foreignborn -1.0458 -1.3278 -0.7674 -1.3203 -5.908 (0.0000)
O grade9 0.0143 -0.0162 0.0456 0.0163 0.697 (0.4856)
O associate 0.0701 0.0191 0.1212 0.0718 1.913 (0.0558)
O college -0.1808 -0.2746 -0.0883 -0.2834 -4.232 (0.0000)
O gradprof 0.1013 0.0266 0.1735 0.1971 3.502 (0.0005)
O rents 0.5575 0.3791 0.7335 0.6938 5.446 (0.0000)
O area 0.0365 0.0245 0.0480 0.0456 5.486 (0.0000)
O unemp 0.5656 0.2792 0.8501 0.4999 2.308 (0.0211)
O pcincome -0.0936 -0.2870 0.1064 -0.1371 -1.256 (0.2091)
log(distance) 0.0045 -0.0010 0.0091 0.0018 0.492 (0.6226)
1 0.3821 0.3424 0.4262 0.3135 13.105 (0.0000)
2 0.3408 0.2952 0.3907 0.2800 11.198 (0.0000)
3 -0.0512 -0.1240 0.0197 -0.0072 -0.172 (0.8627)

25
Table 6: Origin-Destination pairs for variance scalar estimates greater than
10

Origin State Destination State Posterior

mean vi
RI ND 144.32
DE DC 58.38
VT SD 44.65
VT NE 38.81
ND RI 32.65
DE ND 28.47
VT WY 24.57
MT DE 18.54
RI WV 18.53
WY RI 18.23
DE SD 18.11
ND CT 14.63
SD CT 14.54
NH SD 13.06
VT MN 12.40
ID DE 11.33
IA VT 10.52

26
 450
Histogram
Normal plot
400

350

300

250

27
Frequency
200

150

100

50

0
0.4 -0.3 -0.2 -0.1 0 0.1 0.2 0.3
Figure 1: Distribution of Annualized Migration Growth Rates

Migration flow growth rates 1985-90 to 1995-2000