CITY AND REGIONAL PLANNING 870.

03

Notes on Spatial Econometric Models

Philip A. Viton

March 2, 2010

Contents
1 Introduction 2

2 Spatial Autocorrelation 3

3 Spatial Neighbors and Weights 4

3.1 Spatial neighbors . . . . . . . . . . . . . . . . . . . . . . . . . . 4
3.2 Spatial weights . . . . . . . . . . . . . . . . . . . . . . . . . . . 5

4 Is there a spatial problem? 7

4.1 Moran’s Test . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8

5 Spatial Cross-Sectional Models 9

5.1 Spatial Durbin Model . . . . . . . . . . . . . . . . . . . . . . . . 10
5.2 SAR — The Spatial Autoregressive Model . . . . . . . . . . . . . 10
5.3 SEM — The Spatial Error Model . . . . . . . . . . . . . . . . . . 11
5.4 The General Spatial Model . . . . . . . . . . . . . . . . . . . . . 11
5.5 Interpretation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12

6 Spatial Panel Data Models 13

7 Estimation 14
7.1 ML Estimation of the SAR model . . . . . . . . . . . . . . . . . 15
7.2 ML Estimation of the SEM model . . . . . . . . . . . . . . . . . 16
7.3 Computational Considerations . . . . . . . . . . . . . . . . . . . 17

1
8 Software 18
8.1 The R system . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
8.2 Matlab . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
8.3 GeoDa . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21

1 Introduction

Following Bivand, Pebesma and Gómez-Rubio (2008) we can distinguish three

types of spatial analysis:

1. Spatial point processes: we observe a (typically complete) set of data points

in space. We ask whether the data exhibits any spatial clustering patterns,
or whether it exhibits what is referred to as Complete Spatial Randomness
(CSR). For example, consider a pattern of the locations of cancers in an
urban area: are they random, or are they clustered (and if, so, is there any
factor which is associated with the clustering)?

2. Geostatistical data: we observe data at a typically limited set of spatial

points. Based on these observations, we are interested in interpolating the
data to unobserved points. For example, we observe air quality at a set of
monitoring stations. What does this tell us about air quality in the region as
a whole?

3. Areal data: data is observed distributed into pre-defined spatial regions. A

typical question is whether what happens in one region is influenced by what
happens in other regions. For example: we observe crime data by state. Does
the crime rate in one state depend on the rate in other (neighboring) states?

This note discusses only the third problem, in the context of several standard
spatial econometric models. Our setting is a dataset of spatial regions: these could
be countries, states, census tracts, zip codes etc. For the most part we shall assume
that our dataset contains a single observation on each region (ie, an observation at
a single point in time), comprising a spatial cross-section. We suppose that there
are R regions in the dataset.

2
2 Spatial Autocorrelation

The essence of spatial analysis is that “space matters”, ie what happens in one
region is related to what happens in neighboring regions. This has been made more
precise in what Tobler (1979) refers to as the First Law of Geography: “Everything
is related to everything else, but closer things more so”. One way to approach this
is via the notion of spatial autocorrelation. According to Anselin and Bera (1998),

Spatial autocorrelation can be loosely defined as the coincidence of

value similarity with locational similarity. In other words, high or low
values for a random variable tend to cluster in space (positive spa-
tial autocorrelation) or locations tend to be surrounded by neighbors
with very dissimilar values (negative spatial autocorrelation). Of the
two types of spatial autocorrelation, positive autocorrelation is by far
the more intuitive. Negative spatial autocorrelation implies a checker-
board pattern of values and does not always have a meaningful sub-
stantive interpretation.

Formally, if yi and y j are realizations of a random variable y indexed by spatial

locations, then we have spatial autocorrelation if

Corr.yi ; y j / D E.yi y j / E.yi /E.y j / 6D 0

The problem that this raises can be seen from two perspectives. First, if we draw
a sample of R locations from a spatially autocorrelated random process, then ef-
fectively we have a sample size of one. (Compare a single draw from a bivariate
normal density: we get two values, but there is still only one draw). Unless we
have a panel-data setting, there is effectively no way to increase that sample size.
Secondly, the covariance (correlation) matrix for our sample is R R; and clearly
it will be impossible to estimate that many terms. The only possibility is to impose
some a priori structure on the problem to make it manageable. One obvious way
to do this is to assume some systematic pattern of spatial covariances (autocorrela-
tion), parsimoniously parametrized.

Spatial autocorrelation is something like temporal autocorrelation, but more

complicated. The reason is that temporal autocorrelation can only go one way:
what happens at one time can be influenced only what has happened in the past.
But spatial autocorrelation can potentially go in any direction: it is like saying
that what happens at any one point in time can be influenced by both the past

3
and the future. For this reason, we cannot simply transfer models of temporal
autocorrelation to the spatial context.

3 Spatial Neighbors and Weights

3.1 Spatial neighbors

The basis for most models is an indicator of whether one region is a spatial neigh-
bor of another; or equivalently, which regions are neighbors of a given region. This
is a square symmetric R R matrix with .i; j/ element equal to 1 if regions i and
j are neighbors of one another (or more generally, are spatially related), and zero
otherwise. By convention, the diagonal elements of this “spatial neighbors” ma-
trix are set to zero. As LeSage (1998) points out, there is an embarrassingly large
number of ways to construct such a matrix. These include:1

Linear contiguity: i and j are neighbors if they share (part of) a common
eastern or western border.2

Rook contiguity (so called after the movement of the chess piece): two re-
gions are neighbors if they share (part of) a common border (on any side).
In implementation, this is usually based on a small “snap distance” and says
that two regions share a common border if that border is longer than the snap
distance.

Bishop continuity: two regions are spatial neighbors if they meet at a “point”.
This is the spatial analog of two elements of a graph meeting at a vertex. In
implementation, one says that two regions are neighbors in this sense if their
common border is shorter than the “snap distance” referred to above.

Queen contiguity: this is the union of Rook and Bishop contiguity. Two
regions are neighbors in this sense if they share any part of a common border,
no matter how short.
1 There is a good review of the possibilities in Kelejian and Robinson (1995, appendix A). See
also LeSage and Pace (2009): the LeSage reference in the text is an early draft of this book.
2 Can anyone explain why it is ever interesting to privilege the east/west borders over the
north/south ones in a land/spatial context? I don’t see it.

4
 One can go further and define “second order” measures of contiguity: these
would count as neighbors regions sharing a border with a first-order neighbor ac-
cording to each of the criteria listed (so, we might have second-order rook neigh-
bors, etc). Another approach is distance-based, and considers two regions as neigh-
bors if, for example, their population-weighted centroids are within some given
distance of one another. This can obviously be expanded in several ways: one
could use different distances, or different weights for computing the centroids.

It is important not to proceed too mechanically here, but rather to consider

neighbors in light of the practical problem being studied. For example (using a case
mentioned by LeSage), suppose that two zip code regions meet only at a very short
border (effectively, at a point); but that one is considered as a bedroom community
for the other. Then in most cases one would not want to use a rook measure,
because the two regions would not turn out to be neighbors. An exception might
be where there is no direct transportation route between the two: for example, they
are separated be a river, and the nearest bridge involves gong through a third region.
In this case rook contiguity might be exactly what is wanted. It all depends on the
context.

One can construct spatial neighbors matrices “by hand”, but of course this is
tedious and error prone. The spatial analysis software described in section 8 can
construct the matrices for you, given, for example GIS data on the boundaries of
the regions, as represented in an ESRI shapefile.

3.2 Spatial weights

While in principle one could use spatial neighbors matrices directly, in practice one
generally uses slightly transformed matrices, usually referred to as spatial weights
matrices. The most common transformation is called “row-standardization” in
which the rows of the neighbors matrix are made to sum to unity. Let WQ with el-
ements wQ i j be a spatial neighbors matrix. To row-standardize this, we divide each
element in a row by the sum of the elements in the row. Thus a spatial weights
matrix W; with element wi j is defined by
X
wi j D wQ i j = wQ i j
j

5
(or, slightly more precisely, in light of the fact
P that region i may have no neighbors,
ie may be an island; Wi j D wQ i j = max.1; wQ i j /:3 The matrix W is also referred
to as a row-stochastic matrix, since, if there are no islands, each element is be-
tween zero and 1, and the rows sum to one, like probabilities. Note that, unlike the
neighbors matrix, the spatial weights matrix is no longer symmetric.

To see why row-standardization is intuitively attractive, consider any observa-

tion x, an R 1 vector, and look at W x. Since W (and WQ / are R R and x is
R 1; W x is the R 1 vector WQ x; with entry (row) k asPthe sum of those ele-
ments which are neighbors of region k, divided divided by wQ k j ; the count (total
number) of its the neighbors. Thus the k-th entry of W x is an average of x, where
the average is taken only over the neighbors of region k. (Recall that the diagonal
elements of WQ are zero, which means that the data for region k itself does not enter
this computation).

Example: Consider an R D 4 region system, with the following neighbor

relations:
Region Neighbors
1 2; 3; 4
2 1; 3
3 1; 2; 4
4 1; 3
Then 0 1
0 1 1 1
B 1 0 1 0 C
WQ D B
@ 1
C
1 0 1 A
1 0 1 0
P
In this system, region 1 has three neighbors, and the rest have two, so k wQ jk D
.3; 2; 3; 2/ and the row-standardized weights matrix is:
0 1
0 31 13 31
B 1 0 1 0 C
W DB @ 1 1 0 1 A:
2 2 C
3 3 3
1
2
0 12 0
3 In fact, some computational routines can fail if the data contains islands, and it is best (where
possible) to use a dataset without them. That is why one often sees analyses conducted at the level
of the 48 contiguous US states plus the District of Columbia, excluding Alaska and Hawaii.

6
Now if x D .x1 ; x2 ; x3 ; x4 /0 ; then
0 1 1
.x C
3 2
x3 C x4 /
B 1
.x C x3 / C
Wx D B
@
2 1 C
1
.x
3 1
C x2 C x4 / A
1
.x
2 1
C x3 /

which is indeed the vector of the average x over the neighbors of each region. Note
that in the case of a spatially weighted disturbance term, say u, W u will incorporate
the desired “backward and forward” autocorrelation that distinguishes spatial from
temporal autocorrelation.

Another possibility is what has been referred to as distance-based contiguity:

let di j be the distance between (centroids of) regions i and j. We define

wQ i j D 1 if di j < d

and zero otherwise, for a pre-specified d: In other words, regions i and j are spatial
neighbors when they are within d distance units of one another. Yet another possi-
bility starts by defining bi j as the proportion of the length of the border of region i
(ie, its perimeter) that is shared with region j; and then takes
bi j
wQ i j D
di j

The two parameters and can be either pre-specified (one possibility is to take
D 1 and D 2; leading to a gravity-model formulation) or estimated: in the
latter case we clearly have a non-linear problem. In a non-regional application
to a problem studied in sociology or political science, we could take wQ i j D 1 if
individuals i and j belong to the same social network, and study their behavior on
that assumption. In practice, most regional-type research begins with the simple
0=1 row-standardized contiguity matrix.

4 Is there a spatial problem?

The first question an analyst must ask is: does the data exhibit spatial autocorre-
lation or not? If not, then clearly we can be content with the standard non-spatial
models we have been studying so far. But if spatial autocorrelation is present, we
will need to formulate more complicated models.

7
4.1 Moran’s Test

The most common test for the existence of spatial autocorrelation is due to Patrick
Moran, and is usually referred to as Moran’s-I .4 The statistic is defined for a
particular data (or residual ) vector x by
P P
R i j wi j .x i x/.x
N j x/
N
I DP P P 2
i j wi j i .x i x/
N

where xN is the mean of x.5 Moran’s statistic has expectation

1
E.I / D
R 1
and there is a complicated expression for the variance:
R S4 S3 S1 .1 2R/
V.I / D P P
.R 1/.R 2/.R 3/. i j wi j /2

where
1 XX
S1 D .wi j C w ji /2
2 i j
!2
X X X
S2 D wi j C w ji
i j j
1
P
R i .x i N 4
x/
S3 D P 2
1 N 2
R i .x i x/
XX
S4 D .R 2 3R C 3/S1 R S2 C 3. wi j /2
i j

It can be shown that under the null hypothesis of no spatial autocorrelation, Moran’s
statistic is asymptotically normal. So the statistic
I E.I /
I D p
V.I /
4 Note that here I obviously does not refer to the identity matrix. There is usually no problem in
keeping the two separate, though if you work with matrix-based definitions of Moran’s test, there’s
room for ambiguity.
5 For the special case where xN D 0 (this includes the case of the OLS residuals) and where the
P
spatial weights matrix is row-stochastic (so j wi j D 1 for all i) the Moran statistic takes the simple
form I D x 0 W x=x 0 x:

8
is asymptotically standard normal (under the null), and we can use it just like a
p-value.

A few points to bear in mind about this:

The normality is an asymptotic result. One should be wary about this in a

setting where the number of regions is small. On the other hand, Anselin
and Florax (1995) has a simulation-based discussion of the properties of the
Moran statistic in small samples, and conclude that it performs quite well.

The computation of the statistic is relative to a given choice of the spatial

weights W . If in fact the pattern of spatial autocorrelation is generated by a
different set of weights, then the test can give spurious results.

The statistic assumes that the any other trends in the data have been elimi-
nated. If this is not so, the test can again give spurious results. Bivand et
al. (2008, p. 260) give an example where there is no spatial autocorrelation
in the data, but there is a simple linear trend, and where Moran’s test incor-
rectly indicates the presence of spatial autocorrelation. (De-trending the data
resolves the problem).

There are other tests for spatial autocorrelation in the literature, for example
Geary’s C statistic, defined as
P P
R 1 i j wi j .x i N 2
x/
CDP P P
i j wi j i .x i x/N 2

Note that this is not quite the inverse of Moran’s statistic. Moran’s statistic is usu-
ally said to be a measure of global spatial autocorrelation, while Geary’s statistic
is more sensitive to local autocorrelation.

5 Spatial Cross-Sectional Models

Our starting point is the linear-in-parameters cross-sectional model

y D X Cu

with the error term u classical.

9
5.1 Spatial Durbin Model

The spatial Durbin model is

y D X CWX Cu

which just adds average-neighbor values of the independent variables to the specifi-
cation. Example: the level of crime in region j depends on the intensity of policing
in j as well as on the intensity in neighboring jurisdictions. Apart from potential
problems of multicollinearity (recall that row-wise, X and W X are for different re-
gions because the diagonal elements of W are zero), this model poses no problems
for us.

5.2 SAR — The Spatial Autoregressive Model

This model says that levels of the dependent variable y depend on the levels of y in
neighboring regions. It is thus a formulation of the idea of a spatial spillover: for
example, the number of crimes in a zip-code region may depend on the number of
crimes in adjacent (neighboring) zip-code regions. The formal model is

y D Wy C X C u

with u assumed to be classical. Note that W y makes sense since the diagonal
elements of W are zero, which implies that we do not have the circular specifica-
tion that y j on the left is influenced by the same y j on the right. Clearly we would
not want to run OLS on this model, since the presence of y on both the left and
right sides means that we have a correlation-between-errors-and-regressors prob-
lem, and the resulting estimates will be biased and inconsistent. But we can easily
obtain the reduced form as6

y D Wy C X C u
.I W /y D X C u
1
y D .I W/ X C .I W / 1u

(assuming that the inverse exists). We immediately see a couple of potential prob-
lems here. First, the new error term u D .I W / 1 u is no longer homoskedas-
6 Note that from now on we revert to the convention that I is the identity matrix, and not Moran’s
statistic.

10
tic. Second, and probably more fundamentally, the model is no longer linear-in-
parameters because of the new unknown parameter :7

5.3 SEM — The Spatial Error Model

In this model, the spatial influence comes only through the error terms: we have

y D X Cu
u D Wu C v
2
with v assumed to be normal with E.v/ D 0; E.vv 0 / D I (ie, completely classi-
cal). Solving the error specification for u we find

.I W /u D v
u D .I W / 1v

so that the model is

y D X C .I W / 1v
This is conceptually simpler than the SAR case because the only problems are
heteroskedasticity and non-linearity in .8

5.4 The General Spatial Model

It is possible to combine the SAR and SEM models:

y D W1 y C u
u D W2 u C v

where v is classical and both W1 and W2 are spatial weights matrices. (The de-
scription of this as The General Spatial Model is due to LeSage). One motivation
for this is as follows: suppose we have estimated a SAR model. We then test the
7 In fact, very nonlinear in ; as you can see even for the relatively small and straightforward W
used in the example on page 6: here .I W / 1 involves cubic powers of : On the other hand, the
determinant is surprisingly simple, though nonlinear in : it is 9 :
2 3 C7 2 9
8 Note that the literature contains a certain amount of variation in naming these models: what I
have been calling the SAR model may be referred to as the Spatial Lag Model; and what I call the
SEM model is sometimes referred to as the spatial autocorrelated (SAR) model.

11
residuals for spatial autocorrelation using (say) Moran’s test. If we cannot reject
the hypothesis that the residuals are (still) spatially autocorrelated, then this model,
which allows for both sources, may be appropriate.

A problem is the choice of forms (structure) for W1 and W2 : Theory provides no

guides; and if one naively takes W1 D W2 D W then one can run into identification
problems. Thus, this model does not appear to be used much in practice.

5.5 Interpretation

In the context of the SAR model, considering .I W /; one can show that

.1= max / .1= min /

where max and min are the largest and smallest eigenvalues of the spatial weights
matrix W: It can also be shown that if W is row-stochastic:

1 1

but note that there is no theoretical requirement that W actually be row-stochastic.

Now consider
.I W/ 1
If j j < 1; this can be expanded in a power series as
1 2
.I W/ DIC WC W2 C 3
W3 C : ::

(To see this formally, post-multiply both sides by .I W / ; then do the multipli-
cation on the right and note that all terms except I cancel). So suppose that W is
row-stochastic with no islands. Now, many analysts believe that in a spatial con-
text, it is natural to assume that 0 : for example, it would be odd to suppose
that an increase in crime in neighboring census tracts leads to a decrease in crime
in this tract. Finally, think of the series W; W 2 ; etc. W is the spatial weights of the
neighbors of a given region. W 2 is the weights of the neighbors of the neighbors,
W 3 is the weights of the neighbors of the neighbors of the neighbors, etc. So when
we look at
.I W/ 1X
we are examining
2
X C WX C W2X C 3
W3X C : : :

12
which is the sum of a series of decreasing influences of the entire spatial system.
In the case of the SEM we are looking at
.I W / 1u
which by the same reasoning can be interpreted as an autoregressive-in-space error
structure, with decreasing intensity as we move further away.

6 Spatial Panel Data Models

In this context, observations are indexed by i (location) and t (time). We assume

balanced panels: Most formulations of the model adopt an unobserved heterogene-
ity perspective: we have an unobserved time-invariant covariate i . If i is cor-
related with the observed covariates xit then we cannot absorb i into the distur-
bance term (because then it would induce correlation between errors and regres-
sors). Thus, in the case of the SAR model, we write
X
yit D wit j yi j C i C xit C u it
j

where wit j are the elements of the .i; t/ row of W and xit is the .i; t/ row of X;
excluding a constant. This is the (spatial) fixed-effects model. As in non-spatial
fixed effects, the problem comes from the need to estimate the ’s, which function
as region-specific intercepts. (Typically in these models T is much smaller than
R: this is another distinction from non-spatial panels): If the number of regions is
relatively small (a few dozen, perhaps as many at 50) then the model can be set
up with a bunch of dummy variables and estimated without problems. (In practice
it is a bit more difficult to set up, because, unlike the case of non-spatial panels,
we are assuming that the data is grouped by time period: ie all R observations on
the entire cross-section in the first period, followed by R observations on the entire
cross-section in the second period, etc: this is in order to make use of the spatial
weights matrix W meaningful).9 But what if we have upwards of 3000 regions (eg,
the set of all US counties)? We would be attempting to estimate more than 3000
intercepts, which is extremely problematic.

Fortunately, the problem can be solved in the spatial context in much the same
way as in the non-spatial context: by de-meaning the data, region-wise. However,
9 That is, suppose we organized the data as for non-spatial panels. Then the first row of data would
be region 1 in period 1, and the second would be region 1 in period 2, etc. In this case, W y would
not give the desired average (cross-sectional) spatial effect.

13
as Anselin, Gallo and Jayet (2008) observe, the computation of the means is com-
plicated by the spatial dependencies (the W matrix), and must be done carefully.10
But given a correct de-meaning, then just as in the non-spatial context, we obtain
a regression equation without the fixed effects (the ’s). We can then estimate this
equation, and reverse the transformation to recover estimates of the intercepts if
they are of interest to us. (Remember that one standard interpretation of the fixed
effects model is that the unobserved covariate is of no interest to us, ie is a nuisance
parameter). We would also need to compute estimates of the standard errors of the
intercepts; but this can be done, again relatively straightforwardly. We can also
formulate and estimate spatial two-way fixed effects models where there are both
temporally invariant and spatially invariant unobserved covariates.

Returning to the case of a temporally invariant covariate ( i /; if it can be as-

sumed not to be correlated with the observed x’s then we can in principle absorb it
into the disturbance term, resulting in the spatial random effects model. The prob-
lem is then to formulate a set of reasonable but parsimonious assumptions about
the joint distribution of the two components of the error term. This is more difficult
than in the non-spatial context, partly because we cannot just import the non-spatial
assumptions wholesale. For example, Elhorst (2009) describes a model in which
we have
yit D xit C u it
u it D i C "it
"it D W "it C vit
vit D vi;t 1 C z it
where i ~ N .0; 2 / is a set of region-specific effects, z it ~ N .0; 2 I /; vi0 ~ N .0; 2z =
2
1 / (vi0 is the initial value of the temporally autocorrelated disturbance vector)
and and z are assumed independent. Note that this is very general: it has both
spatial error autocorrelation via the parameter and temporal AR(1) autocorrela-
tion via the parameter : See also Kapoor, Kelejian and Prucha (2007).

7 Estimation

In this section we focus on estimation of the cross-sectional models only. For

spatial panels see, eg, Elhorst (2009).
10 As it happens, publication of Anselin et al. (2008) resulted in a substantial revision of the MAT-
LAB code for spatial panels described in the final section of these notes.

14
7.1 ML Estimation of the SAR model

The SAR model

y D Wy C X C u
can be written as

.I W /y D X C u
Ay D X C u

with u~N .0; 2 I /; A D .I W /; and W a row-standardized spatial weights ma-

trix. The model is usually estimated by maximum-likelihood. The log-likelihood
function is
2 2
ln L. ; ; / D .R=2/ ln .R=2/ ln Cln kAk .1=2 /.Ay X /0 .Ay X /

where kAk is the determinant of A: Anselin (1988, ch. 12) suggests a way to do the
estimation. Focussing first on it is straightforward to show that the ML estimator
is given by
1
b D .X 0 X / X 0 Ay
1 1
D .X 0 X / X0y .X 0 X / X 0W y
D b0 bL

where b0 D .X 0 X / 1 X 0 y and b L D .X 0 X / 1 X 0 W y: Inspection shows that b0 is the

coefficient vector from the OLS regression of y on X; while b L is from the OLS
regression of W y on X: So if is known, we could compute the ML estimate of :
Next, write the residuals of these two OLS regressions as

e0 D y X b0
eL D Wy X bL
2
It can be shown that the ML estimate of is

s 2 D .1=R/.e0 e L /0 .e0 eL /
2
so once again we could estimate if were known.

We can now use all this to write down a version of the log-likelihood function
in terms of only: the result is the concentrated log-likelihood, ln L which can
be seen to be

ln L D C .R=2/ ln[.1=R/.eo e L /0 .eo e L /] C ln kAk

15
where C does not involve any unknown parameters. We can now maximize ln L
with respect to and obtain the ML estimate of this parameter, and work back-
wards. In detail, the estimation steps are:

1. Regress y on X : this gives b0 : Compute the residual e0 D y X b0

2. Regress W y on X : this gives b L : Compute the residual e L D W y X bL
3. Find the that maximizes the concentrated log-likelihood function. Call it
O:

4. Given O ; compute b D bo O b L and s 2 D .1=R/.e0 O e L /0 .e0 O eL /

Note that steps (1) and (2) are simply OLS linear estimation problems; and that
step (3) is a one-parameter nonlinear optimization problem.

One problem with this is that, due to the stepwise nature of the estimation
process, you do not get estimates of the (joint) covariance matrix of all the es-
timated parameters. However, because they are maximum-likelihood estimates,
we know that they are asymptotically efficient, meaning that for large samples the
covariance matrix attains the Cramer-Rao lower bound, given by
E.@ 2 ln L=@ @ 0 / 1

2
where D . ; ; /: This in turn can be estimated by the numerical Hessian of
the log likelihood.

7.2 ML Estimation of the SEM model

Estimation of the cross-sectional SEM model

y D X Cu
u D Wu C v
v classical
2
(so that E(vv 0 / D I / which we can write as
y D X Cu
Bu D v
B D .I W/
v classical

16
is somewhat more complicated. The log-likelihood is
2 2
ln L D .R=2/ ln .R=2/ ln C ln kBk .1=2 /.y X /0 B 0 B.y X /

and, as previously described, Bv is heteroskedastic. We could estimate using

GLS, and an estimate of 2 is similar to the SAR case. The concentrated log-
likelihood is

ln L D C .R=2/ ln[.1=R/e0 B 0 Be] C ln kI Wk

where e D y X bGLS : The problem is that bGLS itself depends on (unlike the
SAR case). Anselin suggests an iterative procedure, essentially as follows:

1. Regress y on X: Call the coefficient estimate bOLS and compute the residual
vector e D y X bOLS

2. Use this e in the concentrated log-likelihood, and optimize to find O :

3. Use O to compute the GLS estimator bGLS and then a new residual vector
e D y X bGLS

4. First time or if the residuals have not converged: go back to step 2 and re-
estimate : Otherwise: go to step 5.

5. At this point we have a converged estimate of (say, O / and the associated

residual vector e, and a GLS estimator of : We can now estimate 2 by
.1=R/e0 B 0 Be:

7.3 Computational Considerations

It is important to understand that there are real computational difficulties in the case
where your data represents many regions. For example, suppose you have data by
US counties, of which there are about 3000 .D 3 103 /. Then W is a matrix of
about 9 106 (9 million) elements. Clearly it is going to be very difficult to work
with arrays this size. However, it turns out that for this particular case (the US
counties), and with a rook definition of contiguity, only about 12; 500 elements are
non-zero (LeSage and Pace (2009)). (A county in Ohio will not be a neighbor of a
county in California, or indeed of counties in most other states). So if we could just
keep track of which elements are not zero, we could save considerable space. That
is what sparse-matrix representations do. Instead of keeping track of the entire

17
matrix, they record, for each non-zero element only, its row, its column, and its
value. The result, for US counties, is that we need to keep track of only about
37; 500 ( = 12,500 3) numbers, rather than 9 million. Of course, for all this to
be practically useful, we need routines to work with sparse matrix representations,
eg to be able to multiply or invert them, without having to expand them into their
full (“dense”) forms.

A related problem is that even if W can be represented as a sparse matrix, .I

W / is not sparse and is also R R. However, note from the likelihood functions,
that all we really need is the log determinant of .I W / and not .I W/
itself. For small R one can compute the log determinant directly. For larger R one
approach is to note that
!
YR
ln kI W k D ln .1 i/
iD1

where the i are the eigenvalues of W; which are specific to W and do not change as
estimates of change. The advantage of this is that the eigenvalues need to be com-
puted only once. The disadvantage is that the computation can be time-consuming,
even once: for example, Kelejian and Prucha (1999, footnote 12) report for a case
of R D 1500 and with the average number of spatial neighbors (non-zero entries in
a row) equal to 10, computing the eigenvalues took 22 minutes; however, this was
on what is now considered very old and slow hardware. This slowness motivated
a search for a method to compute the log determinant that did not need to compute
the eigenvalues. Nowadays, with improved hardware and more RAM, this is less
of a problem. Still, there is room for improvement: a popular approach is a Monte
Carlo method due to Barry and Pace (1999).

8 Software

Most standard statistics packages do not contain estimation routines for spatial
econometric models.11 However, most packages do contain a built-in programming
language, so that in principle you could implement the estimation routines yourself,
following the sketches in the previous section. But obviously this is a second-best
solution: if you anticipate doing lots of work with spatial data, it is worth while
learning how to use a package that has spatial estimation routines built in.
11 I know that Limdep does not. Stata may contain an add-on for spatial econometrics, but I don’t
know for sure.

18
8.1 The R system

R has probably the most extensive set of spatial statistics support, not limited to
the econometric models discussed in this note. It is a free and open-source system,
available at http://www.r-project.org/ for most computer platforms (includ-
ing Windows, the Mac, and most flavors of Unix). The downside it that it has a
fairly steep learning curve (at least I find that it has) and that it is command-driven
system: no point-and-click menus etc.

The core R system is supplemented by a very extensive package system, pro-

viding routines to do all sorts of statistical and data-analysis tasks. R has become
the preferred platform for professional statisticians, so it is extraordinarily com-
plete: about the only thing I have found it lacking is estimation of the mixed-logit
(random-parameters logit) model. It has some of the best graphics available for
any system, even commercial.

There are three packages of interest in our spatial context: sp, spdep and
splm.12 The sp package contains routines for constructing spatial neighbors and
weights matrices, and can manipulate ESRI shapefiles. spdep has cross-sectional
econometric estimation routines, using the weights matrices from sp;while splm
has routines for some panel-data models, also using weights matrices from sp.13
Bivand et al. (2008) describes the main facilities provided by sp and spdep: this
book is available for free download (for OSU faculty and students) from the Springer
website.14 Chapter 9 describes the construction of spatial weights matrices, and
Chapter 10 illustrates some of the estimation routines. Briefly, the cross-sectional
estimation routines are:
12 There is a general view of the facilities available for spatial analysis in R at http://
cran.r-project.org/web/views/Spatial.html. See also http://geodacenter.asu.edu/
r-spatial-projects.
13 The panel-data routines are not available as a standard R package. You
can install them from the R command line via install.packages("splm",
repos="http://R-Forge.R-project.org"), but it may be easier to install it by hand af-
ter getting the materials from http://r-forge.r-project.org/projects/splm/. Just put the
zip file somewhere on your hard disk, then, from within R do Packages -> Insall from local zip
files.
14 In fact, most of the books in the “Use R!” series are available for download as PDF files. This
includes a very helpful book (Spector (2007)) on managing data in R, including facilities for im-
port and export. I have produced a version of Bivand et al. (2008) re-formatted to put 2 book
pages on one physical (landscape) page. The re-formatted version is available on the (restricted,
OSU people only) site: http://facweb.knowlton.ohio-state.edu/pviton/courses/misc/
bivand-crop.zip.

19
 Function Estimates
lm Basic non-spatial linear regression model
lagsarlm Cross-sectional spatial autoregressive (SAR) model
lagsarlm with type="mixed" Cross-sectional spatial Durbin model
errorsarlm Cross-sectional spatial error (SEM) model

The panel-data models are estimated with package splm. The package allows
estimation of both fixed-effects and random effects models: for random effects, the
specification is as described in section 6. The routines are:

Function Estimates
spfelm with effects="spfe" Panel-data spatial fixed-effects model
spfelm with effects="tpfe" Panel-data spatial time-effects model
spfelm with effects="sptpfe" Panel-data spatial 2-way fixed-effects model
spreml with errors="semssre" Panel-data spatial random effects model

and there are variants of the random effects model for special cases of no spatial
autocorrelation ( D 0/ and/or no time autocorrelation ( D 0/:

It is not clear how effective these routines are with large cross-sections, where
the computational problems can be significant: in this case you may want to con-
sider Matlab, see below. For the panel data models in package splm, the docu-
mentation may not be available if you use an R version earlier than 2.10 (because
the R system changed the way it processes help files). The package still works,
though, and I have a PDF version of a pre-print of a paper by Giovanni Millo and
Gianfranco Piras which is to be submitted to the Journal of Statistical Software;
this gives the essential documentation. I have placed a copy on the CRP 780.03
restricted website.

8.2 Matlab

Matlab is a general-purpose matrix-oriented programming language, available on

an essentially free year-to-year license to OSU students and faculty. Licenses ex-
pire at the beginning of each academic year, after which the system will not start
until you obtain a new license. For students, this means that once you leave OSU
you will need to obtain a commercial license in order to continue (or even revise)
your work (unless you go to an institution which also has a Matlab license).

20
 Unlike R, Matlab is not primarily a data-analysis system. In particular, it has
no facilities for referring to variables by name: what you do in Matlab is read your
data into a matrix, and then refer to columns of that matrix by column number.
So a typical regression command might be “regress column 12 of matrix DATA on
columns 1,3 6 7 and 9 of DATA”. This is obviously less convenient than a system
which allows you to say “regress crime against pop, income, and age”.

James LeSage has written a general-purpose econometrics toolbox for Matlab,

available free from www.spatial-econometrics.com. Besides containing es-
timation routines for standard single-equation and simultaneous equations linear
models, it also contains routines for estimating spatial models.15 Because of Mat-
lab’s efficient sparse matrix storage facilities, it is relatively fast to estimate even
very large models (for example, panel-data models involving the set of 3000 US
counties). The main functions are:

Function Estimates
ols(y,X) Basic non-spatial linear model
sar(y,X,W) Spatial cross-sectional autoregressive (SAR) model
sem(y,X,W), Spatial cross-sectional error (SEM) model
sac(y,X,W1,W2) General cross-sectional spatial model
sem_panel_FE(y,X,W1,W2,T) spatial fixed-effects panel data model
sem_panel_RE(y,X,W1,W2,T) spatial random-effects panel data model

with the W’s being row-standardized spatial weights matrices. The panel-data mod-
els (for balanced panels only) are due to J. Paul Elhorst, and are as specified in
section 6. It is very important to obtain the latest (2009) versions of the panel-data
code: earlier versions did the de-meaning incorrectly. See Anselin et al. (2008) and
Elhorst (2009). The spatial toolbox comes with a few examples, and I have found
it relatively easy to adapt the examples to my own estimation needs without having
to learn too much about Matlab.

8.3 GeoDa

This is a free Windows program written by Luc Anselin for the construction for
spatial weights matrices and estimation of the cross-sectional SAR and SEM mod-
15 Ken Train has also produced Matlab code for estimating the mixed-logit model, enhancing the
attractiveness of this platform for computationally intensive estimation tasks.

21
els. It is available at http://geodacenter.asu.edu/. If you are running Win-
dows Vista it is important to obtain the version referred to as OpenGeoDa, which
is the only version to run under this version of the Windows operating system.
Anselin and his colleagues at Arizona State University support a general spatial
analysis mailing list called The Openspace List, which entertains questions on all
forms of spatial analysis and econometrics (not limited to GeoDa; Roger Bivand is
a frequent contributor). There is apparently also under development an enhanced
set of spatial analysis tools based on the (free and open-source) Python program-
ming language. Note that the R package sp can read spatial-weights matrices
produced by GeoDa.

GeoDa is very fast and easy to use — it is based on a standard Windows point-
and-click-type interface — but personally I find it a little too much of a “black box”
for my taste. Users cannot extend the estimation routines, so that if you need to do
something not provided by the system, you’re out of luck. My feeling is that an
investment in learning a bit of either R or Matlab programming will in the long run
be more productive.

References

Anselin, Luc, Spatial Econometrics: Methods and Models, Dordrecht, The

Netherlands: Kluwer Academic Publishers, 1988.

and A. K. Bera, “Spatial Dependence in Linear Regression Models with an

Introduction to Spatial Econometrics,” in A. Ullah and D. Giles, eds., Hand-
book of Applied Economic Statistics, New York, N.Y.: Marcel Dekker, 1998,
pp. 237–289.

and Raymond J. G. M. Florax, “Small Sample Properties of Tests for Spatial

Dependence in Regression Models: Some Further Results,” in Luc Anselin
and Raymond J. G. M. Florax, eds., New Directions in Spatial Econometrics,
Berlin: Springer, 1995, pp. 21–74.

, Julie Le Gallo, and Hubert Jayet, “Spatial Panel Econometrics,” in Láslo

Mátyás and Patrick Sevestre, eds., The Econometrics of Panel Data, New
York, N.Y.: Springer, 2008, pp. 625–659.

Barry, R. P. and R. K. Pace, “Monte Carlo Estimates of the Log Determinant of

Large Sparse Matrices,” Linear Algebra and Its Applications, 1999, 289, 41–
54.

22
Bivand, Roger S., Edzer J. Pebesma, and Virgilio Gómez-Rubio, Applied Spatial
Data Analysis with R Use R!, Secaucus, N.J.: Springer, 2008.

Elhorst, J. Paul, “Spatial Panel Data Models,” in Mandfred M. Fischer and Arthur
Getis, eds., Handbook of Applied Spatial Analysis, Berlin: Springer, 2009.

Kapoor, M., Harry H. Kelejian, and Ingmar R. Prucha, “Panel Data Models
with Spatially Autocorrelated Error Components.,” Journal of Econometrics,
2007, 140 (1), 97–130.

Kelejian, Harry H. and Dennis P. Robinson, “Spatial Correlation: A Suggested

Alternative to the Autoregressive Model,” in Luc Anselin and Raymond J.
G. M. Florax, eds., New Directions in Spatial Econometrics, Berlin: Springer,
1995, pp. 75–95.

and Ingmar R. Prucha, “A Generalized Moments Estimator for the Au-

toregressive Parameter in a Spatial Model,” International Economic Review,
1999, 40 (2), 509–533.

LeSage, James P., Spatial Econometrics, www.spatial-econometrics.org, 1998.

LeSage, James P. and R. Kelly Pace, Introduction to Spatial Econometrics, Boca

Raton: CRC Press, 2009.

Spector, Phil, Data Manipulation With R Use R!, Secaucus, N.J.: Springer, 2007.

Tobler, W., “Cellular Geography,” in S. Gale and G. Olssohn, eds., Philosophy in

Geography, Dordrecht: Reidel, 1979, pp. 379–386.

23