SPATIAL ECONOMETRICS

Luc Anselin

Bruton Center
School of Social Sciences
University of Texas at Dallas
Richardson, TX 75083-0688
[email protected]

April 26, 1999

ABSTRACT: Spatial econometric methods deal with the incorporation of spatial interaction and spatial
structure into regression analysis. The field has seen a recent and rapid growth spurred both by theoretical
concerns as well as by the need to be able to apply econometric models to emerging large geocoded data
bases. The review presented in this chapter outlines the basic terminology and discusses in some detail the
specification of spatial effects, estimation of spatial regression models, and specification tests for spatial
effects.
 SPATIAL ECONOMETRICS

Luc Anselin

1. Introduction
Spatial econometrics is a subfield of econometrics that deals with the treatment of spatial interac-
tion (spatial autocorrelation) and spatial structure (spatial heterogeneity) in regression models for
cross-sectional and panel data [Paelinck and Klaassen (1979), Anselin (1988a)]. As such, the field
is similar to geostatistics and spatial statistics, which form an important part of the methodologi-
cal toolbox of the physical scientist [e.g., Ripley (1981), Cressie (1993)]. However, spatial econo-
metrics is distinct from spatial statistics in the same sense as econometrics is distinct from
statistics in general. This albeit subtle distinction boils down to the central role attributed to the
theoretical model rather than the data when it comes to dictating the types of specifications that
are of interest in spatial econometrics.
A focus on location and spatial interaction has recently gained a more central place not
only in applied but also in theoretical econometrics.1 In the past, models that explicitly incorpo-
rated “space” (or geography) and therefore applications of spatial econometrics were primarily
found in specialized fields such as regional science, urban and real estate economics and eco-
nomic geography [e.g., recent reviews in Anselin (1992a), Anselin and Florax (1995a), Anselin
and Rey (1997), Pace et al. (1998)]. However, more recently, spatial econometric methods have
increasingly been applied in a wide range of empirical investigations in more traditional fields of
economics as well, including, among others, studies in demand analysis [Case (1991)], interna-
tional economics [Aten (1996)], labor economics [Topa (1996)], public economics and local public
finance [Case, Rosen and Hines (1993), Murdoch, Rahmatian and Thayer (1993), Holtz-Eakin
(1994), Brueckner (1998)], and agricultural and environmental economics [Benirschka and Bin-
kley (1994), Murdoch, Sandler and Sargent (1997), Nelson and Hellerstein (1997), Bell and Bocks-
tael (1999)]. In addition, there is a growing methodological literature in econometrics that deals
with alternative model specifications, estimators and test statistics geared to models that incor-
porate the “geography” of the data [recent examples include, among others, Conley (1996),
Driscoll and Kraay (1998), Pinkse and Slade (1998), Kelejian and Prucha (1999a)].

1 The origins of this reaction to the so-called Anglo-Saxon focus in economics on the time domain can be traced
to Isard (1956).
 This new attention to specifying, estimating and testing for the presence of spatial inter-
action in the mainstream of applied and theoretical econometrics can be attributed to two major
factors. One is a growing interest within theoretical economics in models that move away from
the atomistic agent as a decision maker acting in isolation to an explicit accounting for the inter-
action of that agent (its preferences, utility, etc.) with other heterogeneous agents in the system.
These new theoretical frameworks specify and study “direct” interaction between agents, in the
form of social norms, neighborhood effects, copy-catting and other peer group effects, and raise
interesting questions about how the individual interactions can lead to emergent collective
behavior and aggregate patterns. Examples of such models are found in the new macroeconom-
ics of Aoki (1994, 1996), in theoretical models of social interaction [Brock and Durlauf (1995),
Akerlof (1997)], interdependent preferences [Alessie and Kapteyn (1991)], models of evolving
trading structures [Ioannides (1990, 1997)], neighborhood spillover effects [Durlauf (1994), Borjas
(1995), Glaeser, Sacerdote and Scheinkman (1996)], and yardstick competition [Besley and Case
(1995), Bivand and Szymanski (1997)]. These frameworks also form some of the underpinnings
for empirical models that encompass strategic interaction between agents [e.g., Case, Rosen and
Hines (1993), Murdoch, Sandler and Sargent (1997), Brueckner (1998)]. Much of this literature is
inspired by principles developed in statistical mechanics, such as the study of interacting particle
systems and random field models [for a review, see Durlauf (1997)], which, interestingly, also
forms a basis for some of the spatial Markov field models developed in spatial statistics [e.g.,
Cressie (1993, Ch. 6)]. Related to this strand in the literature is the revived emphasis on the spa-
tial aspects of Marshallian externalities, agglomeration economies and other spillovers that are
central to the new economic geography reflected in the work of Arthur (1989), Krugman (1991a,
1991b, 1998), Glaeser et al. (1992) and others.
Paralleling and complementing this theoretical motivation, i.e., the need for a methodol-
ogy to handle spatial models, there has also been a second driver behind the increase in demand
for techniques to deal with spatial data from a practical, applied perspective. Irrespective of the
nature of the model of interest (whether spatial or non-spatial), the explosive diffusion of geo-
graphic information systems (GIS) technology and the associated availability of geo-coded socio-
economic data sets (i.e., data sets that contain the location of the observational units) has created
a need for specialized methods to deal with the distinguishing characteristics (primarily spatial
autocorrelation) of such geographic data [Anselin (1989)]. The use of GIS in conjunction with
spatial data analysis and modeling techniques is by now commonplace in applied economics
and policy analysis, especially in real estate and housing economics [Anselin (1998a), Can

—2—
(1998)], environmental and resource economics [Bockstael (1996), Geoghegan, Wainger and
Bockstael (1997)], and development economics [Nelson and Gray (1997)]. The recognition that
standard econometric techniques often fail in the presence of spatial autocorrelation, which is
commonplace in geographic (cross-sectional) data sets, is therefore a major second motivation
for the growing interest in this field.
Historically, spatial econometrics originated as an identifiable field in Europe in the early
1970s because of the need to deal with sub-country data in regional econometric models [e.g.,
Hordijk and Paelinck (1976)]. In the first book-length treatment of the subject, Paelinck and
Klaassen (1979, 5–11) defined the field in terms of five distinguishing characteristics: (a) the role
of spatial interdependence in spatial models; (b) the asymmetry in spatial relations; (c) the
importance of explanatory factors located in other spaces; (d) differentiation between ex-post
and ex-ante interaction; and (e) explicit modeling of space. In Anselin (1988a, 7), spatial econo-
metrics is defined as “the collection of techniques that deal with the peculiarities caused by space
in the statistical analysis of regional science models.” In other words, spatial econometrics deals
with methodological concerns that follow from the explicit consideration of spatial effects, such as
spatial autocorrelation and spatial heterogeneity. This yields four broad areas of interest: (a) the
formal specification of spatial effects in econometric models; (b) the estimation of models that
incorporate spatial effects; (c) specification tests and diagnostics for the presence of spatial effects;
and (d) spatial prediction (interpolation). In this brief review chapter, I will focus on the first three
concerns, since they fall within the central preoccupation of econometric methodology.
The remainder of the chapter is organized as follows. In section 2, I outline some founda-
tions and definitions. In section 3, the specification of spatial regression models is treated, includ-
ing the incorporation of spatial dependence in panel data models and models with qualitative
variables. Section 4 focuses on estimation and section 5 on specification testing. In section 7,
some practical implementation and software issues are addressed. Concluding remarks are for-
mulated in Section 8.

2. Foundations

2.1. Spatial Effects

In a regression context, spatial effects pertain to spatial dependence, or its weaker expression,

spatial autocorrelation, and to spatial heterogeneity.2 The latter is simply structural instability in the

2 In this chapter, I will use the terms spatial dependence and spatial autocorrelation interchangeably.
Obviously, the two are not identical, but typically, the weaker form is used, in the sense of a moment of a joint
distribution. Only seldom is the focus on the complete joint density [a recent exception can be found in Brett
and Pinkse (1997)].

—3—
form of non-constant error variances (heteroskedasticity) or model coefficients (variable coeffi-
cients, spatial regimes) and can be tackled by means of the standard econometric toolbox. How-
ever, there are three reasons why it is important to consider spatial heterogeneity explicitly. First,
the “structure” behind the instability is spatial (or geographic) in the sense that the location of the
observations is crucial in determining the form of the instability. For example, groupwise het-
eroskedasticity could be modeled as different error variances for different compact geographic
subsets of the data.
More formally, consider a set S of N geographical units (e.g., states, counties, census
tracts), partitioned into R non-overlapping compact subsets S r (with r = 1, 2, …, R ), such that
for any r, s ( r ≠ s ), S r ∩ S s = ∅ , and ∪ S r = S . Spatial groupwise heteroskedasticity
r = 1, …, R
would then follow in the form of spatially clustered error variances for observation i,
2
Var [ ε i ] = σ r when i ∈ S r . Similarly, variability in the regression coefficients could be specified
to correspond to so-called spatial regimes, or geographic subsets S r of the data where the model
slope is different, β i = β r , for an observation i ∈ S r [for further details, see Anselin (1988a,
1990a)]. Secondly, because the structure is spatial, heterogeneity often occurs jointly with spatial
autocorrelation, and standard econometric techniques are no longer appropriate [e.g., tests for
heteroskedasticity may be misleading, as illustrated in Anselin and Griffith (1988)]. Thirdly, in a
single cross-section, spatial autocorrelation and spatial heterogeneity may be observationally
equivalent.3 For example, a spatial cluster (i.e., observed in locations that are in close proximity)
of extreme residuals may be interpreted as due to spatial heterogeneity (e.g., groupwise het-
eroskedasticity) or as due to spatial autocorrelation (e.g., a spatial stochastic process yielding
clustered values). This requires that both aspects of the problem be structured very carefully to
obtain identifiability of the model parameters, and that one aspect can never be considered in
isolation from the other. In the remainder of the chapter, however, the main focus of attention
will be on spatial dependence.
Spatial autocorrelation, or the coincidence of value similarity with locational similarity,
can be formally expressed by the moment condition

Cov [ y i, y j ] = E [ y i y j ] – E [ y i ] ⋅ E [ y j ] ≠ 0 , for i ≠ j (2.1)

where i, j refer to individual observations (locations) and y i ( j ) is the value of a random variable
of interest at that location. This covariance becomes meaningful from a spatial perspective when
the particular configuration of nonzero i, j pairs has an interpretation in terms of spatial struc-

3 For similar concerns in a time series context, see, e.g., Heckman (1991).

—4—
ture, spatial interaction or the spatial arrangement of the observations.

2.2. Spatial Weights and Spatial Lags

In a cross-sectional setting with N observations, there is insufficient information to esti-
mate the N by N covariance matrix (2.1) directly from the data. Even asymptotics are not helpful
2
since the number of covariances increases with N , whereas the sample size only grows with N.
In contrast, when repeated observations on the cross-section are available (as in a panel data set-
ting), it may be possible to exploit the other dimension (i.e., asymptotics in T) and obtain consis-
tent nonparametric cross-sectional covariance matrix estimates.4 In general however, it will be
necessary to impose a structure on the covariance.
There are three main approaches followed in the literature to address this issue: one
based on the specification of a spatial stochastic process; a second on the direct parametric repre-
sentation of the covariance structure; and a third where the covariance is left generally unspeci-
fied and treated in a non-parametric framework. I return to this issue in the next section. First, it
is important to consider the important concepts of spatial weights and spatial lag operator.
In parallel to time series analysis, spatial stochastic processes are categorized as spatial
autoregressive (SAR) and spatial moving average (SMA) processes, although there are several
important differences between the cross-sectional and time series contexts.5 Most importantly, in
contrast to the unambiguous notion of a “shift” along the time axis, there is no corresponding
concept in the spatial domain, especially when observations are located irregularly in space.6
Instead of the notion of shift, a spatial lag operator is used, which is a weighted average of random
variables at “neighboring” locations.7 Essential in this concept is the definition of a neighbor-
hood set for each location. This is obtained by specifying for each location i (as the row) the
neighbors as the columns corresponding to non-zero elements w ij in a fixed (non-stochastic) and
positive N by N spatial weights matrix W.8 Formally, a spatial lag for y at i is then expressed as

[ Wy ] i = ∑ w ij ⋅ y j , (2.2)
j = 1, …, N

4 Recent reviews of the pertinent issues are given in Frees (1995) and Driscoll and Kraay (1998).
5 See Anselin and Bera (1998) for an extensive and technical discussion.
6 On a regular lattice (e.g., square grid), one can conceive of north, south, east, west shifts [e.g., Whittle (1954)].
Such a model is often used in statistical mechanics, but it is too restrictive for economic applications.
7 In Anselin (1988a), the term spatial lag is introduced to refer to this new variable, to emphasize the similarity
to a distributed lag term rather than a spatial shift.
8 By convention, w ii = 0 , i.e., a location is never a neighbor of itself. This is arbitrary, but can be assumed
without loss of generality. For a more extensive discussion of spatial weights, see Anselin (1988a, Ch. 3), Cliff
and Ord (1981), Upton and Fingleton (1985). For a formal discussion of the related notion of a spatial markov
process, see Ripley (1988), Cressie (1993, pp. 402–410).

—5—
or, in matrix form, as

Wy (2.3)

where y is a N by 1 vector of observations on the random variable. Since for each i the matrix ele-
ments w ij are only non-zero for those j ∈ S i (where S i is the neighborhood set), only the match-

ing y j are included in the lag. For ease of interpretation, the elements of the spatial weights

matrix are typically row-standardized, such that for each i, ∑ w ij = 1 . Consequently, the spatial
j

lag may be interpreted as a weighted average (with the w ij being the weights) of the neighbors,

or as a spatial smoother.
It is important to note that the elements of the weights matrix are non-stochastic and
exogenous to the model. Typically, they are based on the geographic arrangement of the observa-
tions, or contiguity. Weights are non-zero when two locations share a common boundary, or are
within a given distance of each other. However, this notion is perfectly general and alternative
specifications of the spatial weights can be based on distance decay (inverse distance or inverse
distance squared) [Anselin (1980, Ch. 8)], on the structure of a social network [Doreian (1980)], on
economic distance [Case, Rosen and Hines (1993)], k nearest neighbors [Pinkse and Slade (1998)],
or on empirical flow matrices [e.g., pollutant emission flows in Murdoch, Sandler and Sargent
(1997), or trade-based interaction measures in Aten (1996, 1997)]. While this wide range of
options may suggest that the specification of spatial weights is largely arbitrary, this is not the
case. Due to the asymptotics required to obtain consistent and asymptotically normal estimators,
the range of dependence allowed by the structure of W must be constrained. Also, to avoid iden-
tification problems, the weights should truly be exogenous to the model [Manski (1993)]. In spite
of their lesser theoretical appeal, this explains the popularity of geographically derived weights,
since exogeneity is unambiguous.

2.3. Formal Expression of Spatial Autocorrelation

Spatial Stochastic Process Models
The most often used approach to express spatial autocorrelation is through the specification of a
spatial stochastic process, i.e., a functional relationship between a random variable at a given
location and this same random variable at other locations. The covariance structure then follows
from the nature of the process. For example, given the N by N weights matrix W, a N by 1 vector
of random variables y, and a N by 1 vector of i.i.d. random errors ε , a simultaneous spatial
autoregressive process is defined as

—6—
 –1
( y – µi ) = ρW ( y – µi ) + ε , or ( y – µi ) = ( I – ρW ) ε (2.4)

and a spatial moving average process as

y = λWε + ε , or y = ( I + λW )ε , (2.5)

where I is a N by N identity matrix, i a N by 1 vector of ones, µ is the common mean of the ran-
2
dom variables y i , the i.i.d. zero mean error terms ε i have a common variance σ , and ρ and λ
are respectively the autoregressive and moving average parameters.9 As in time series analysis,
it is also possible to write a spatial autoregressive process as an infinite spatial moving average,
since the elements of W are typically less than one and, in addition, in most cases ρ < 1 , such
that

2 2
( y – µi ) = Iε + ρWε + ρ W ε + … . (2.6)

The constraints imposed by the weights structure, together with the specific form of the
spatial process determine the variance-covariance matrix for y as a function of two parameters,
2
the variance σ and the spatial coefficient, ρ or λ . For the SAR structure in (2.4), this yields
(since E [ y – µi ] = 0 )

2 –1
Cov [ ( y – µi ), ( y – µi ) ] = E [ ( y – µi ) ( y – µi )′ ] = σ [ ( I – ρW )′ ( I – ρW ) ] . (2.7)

This is a full matrix, which implies that shocks at any location affect all other locations, through a
spatial multiplier effect (or, global interaction). In contrast, for an SMA process, the covariance is
(since E [ y ] = 0 )

2 2 2
Cov [ y, y ] = E [ yy′ ] = σ ( I + ρW ) ( I + ρW )′ = σ [ I + ρ ( W + W′ ) + ρ WW′ ] , (2.8)

which results in “local” interaction between a location and its first (through W) and second order
neighbors (through WW′ ), but does not yield a spatial multiplier as in the SAR model.
A major distinction between AR and MA processes in space compared to the time
domain is that even with i.i.d. error terms ε i , the diagonal elements in (2.7) and (2.8) are not con-
stant. Furthermore, the heteroskedasticity depends on the neighborhood structure embedded in
the spatial weights matrix W. Consequently, the process in y is not covariance-stationary. Station-
arity is only obtained in very rare cases, for example on regular lattice structures when each
observation has an identical weights structure, but this is of limited practical use. This lack of sta-

9 In spatial econometrics the main focus is on processes for which the joint distribution is modeled, or, a
simultaneous processes, in contrast to conditional processes. Simultaneity is more appropriate when the
interest is in inference, whereas a conditional approach is natural when the interest is in prediction or spatial
interpolation [see Anselin and Bera (1998) and Cressie (1993, p. 410) for technical details].

—7—
tionarity has important implications for the types of central limit theorems and laws of large
numbers that need to be invoked to obtain asymptotic properties for estimators and specification
test, a point that has not always been recognized in the literature.
A related specification, which also results in a heteroskedastic variance, is the spatial
error components model outlined in Kelejian and Robinson (1993, 1995, 1997). This spatial pro-
cess is very similar to a spatial MA process, but rather than pertaining to a single error term, it
contains two uncorrelated zero mean error components ψ and ε :

y = Wψ + ε (2.9)

where the term Wψ implies a smoothing of neighboring values or “regional” effect, and the
error term ε is location-specific. The resulting variance-covariance matrix is (since E [ y ] = 0 )

2 2
Cov [ y, y ] = E [ yy′ ] = σ ε I + σ ψ WW′ (2.10)
2 2
where σ ε and σ ψ are the variance components pertaining to the locational “innovation” and
the regional effect respectively. The range of interaction implied by (2.10) is a subset of that in the
spatial MA model, as seen from (2.8).

Direct Representation
A second commonly used approach to the formal specification of spatial autocorrelation
is to express the elements of the variance-covariance matrix in a parsimonious fashion as a
“direct” function of a small number of parameters and one or more exogenous variables. Typi-
cally, this involves an inverse function of some distance metric, for example,

2
Cov [ ε i ε j ] = σ f ( d ij, ϕ ) , (2.11)
2
where ε i and ε j are regression disturbance terms, σ is the error variance, d ij is the distance
∂f
separating observations (locations) i and j, and f is a distance decay function such that ------- < 0
∂
p
d
and f ( d ij, ϕ ) ≤ 1 , with ϕ ∈ Φ as a p by 1 vector of parameters on an open subset Φ of R . This
form is closely related to the variogram model used in geostatistics, although with stricter
assumptions regarding stationarity and isotropy.10 Using (2.11) for individual elements, the full
error covariance matrix follows as

10 The specification of spatial covariance functions is not arbitrary, and a number of conditions must be satisfied
in order for the model to be “valid” [details are given in Cressie (1993, pp. 61–63, 67–68 and 84–86)]. For a
stationary process, the covariance function must be positive definite and it is required that f → 0 as d ij → ∞ ,
which is ensured by the conditions spelled out here. Furthermore, most models only consider positive spatial
autocorrelation in this context. An exception is the so-called wave variogram, which allows both positive and
negative correlation due to the periodicity of the process. This model has not seen application outside the
physical sciences and is not considered here.

—8—
 2
E [ εε′ ] = σ Ω ( d ij, ϕ ) (2.12)
2
where, because of the scaling factor σ , the matrix Ω ( d ij, ϕ ) must be a positive definite spatial
correlation matrix, with ω ii = 1 and ω ij ≤ 1 , ∀ i, j .11 Note that, in contrast to the variance for
the spatial process models (2.7), (2.8) and (2.10), the direct representation model does not induce
heteroskedasticity.
In spatial econometrics, models of this type have been used primarily in the analysis of
urban housing markets, e.g., in Dubin (1988, 1992), Olmo (1995) and Basu and Thibodeau (1998).
While this specification has a certain intuition, in the sense that it incorporates an explicit notion
of spatial clustering as a function of the distance separating two observations (i.e., positive spa-
tial correlation), it is also fraught with a number of estimation and identification problems [Anse-
lin (1999)].

Nonparametric Approaches
A recently suggested alternative specification of the spatial covariance matrix is to estimate its
elements non-parametrically, in the sense of not requiring an explicit spatial process or func-
tional form for the distance decay. This is common in the case of panel data, when the time
dimension is (considerably) greater than the cross-sectional dimension (T >> N) and the “spatial”
covariance is estimated from the sample covariance for the residuals of each set of location pairs
[e.g., in applications of Zellner’s SUR estimator, see Fiebig (1999)].
The new approaches are variants of the well known Newey-West (1987) heteroskedastic-
ity and autocorrelation consistent covariance matrix and have been used in the context of gener-
alized methods of moments (GMM) estimators of spatial regression models (see Section 4.3.).
Conley (1996) suggested a covariance estimator based on a sequence of weighted averages of
sample autocovariances computed for subsets of observation pairs that fall within a given dis-
tance band (or spatial window). Although not presented as such, this has a striking similarity to
the non-parametric estimation of a semi-variogram in geostatistics [see, e.g., Cressie (1993),
pp.69–70], but the assumptions of stationarity and isotropy required in the GMM approach are
stricter than those needed in variogram estimation. In a panel data setting, Driscoll and Kraay
(1998) use a similar idea, but avoid having to estimate the spatial covariances by distance bands.
This is accomplished by using only the cross-sectional averages (for each time period) of the
moment conditions, and by relying on asymptotics in the time dimension to yield an estimator
for the spatial covariance structure.

11 ω ii = 1 is ensured by selecting a functional form for f such that f ( d ij, ϕ ) = 1 for d ij = 0 .

—9—
2.4. Aymptotics in Spatial Stochastic Processes
As in time series analysis, the properties of estimators and tests for spatial series are derived
from the asymptotics for stochastic processes. However, these properties are not simply exten-
sions to two dimensions of the time series results. A number of complicating factors are present
and to date some formal results for the spatial dependence case are still lacking. While an exten-
sive treatment of this topic is beyond the scope of the current chapter, three general comments
are in order. First, the intuition behind the asymptotics is fairly straightforward in that regularity
conditions are needed to limit the extent of spatial dependence (memory) and heterogeneity of
the spatial series in order to obtain appropriate (uniform) laws of large numbers and central limit
theorems to establish consistency and asymptotic normality. In this context, it is important to
keep in mind that both SAR and SMA processes yield heteroskedastic variances, so that the
application of results for dependent stationary series are not appropriate.12 In addition to the
usual moment conditions that are similar in spirit to those for heterogeneous dependent pro-
cesses in time [e.g., Pötscher and Prucha (1997)], specific spatial conditions will translate into
constraints on the spatial weights and on the parameter space for the spatial coefficients [for
some specific examples, see, e.g., Anselin and Kelejian (1997), Kelejian and Prucha (1998, 1999a,
1999b), Pinkse and Slade (1998), and Pinkse (1998, 1999)]. In practice, these conditions are likely
satisfied by most spatial weights that are based on simple contiguity, but this is not necessarily
the case for general weights based on economic distance.
A second distinguishing characteristic of asymptotics in space is that the limit may be
approached in two different ways, referred to as increasing domain asymptotics and infill asymp-
totics.13 The former consists of a sampling structure where new “observations” are added at the
edges (boundary points), similar to the underlying asymptotics in time series analysis. Infill
asymptotics are appropriate when the spatial domain is bounded, and new observations are
added in between existing ones, generating a increasingly denser surface. Many results for
increasing domain asymptotics are not directly applicable to infill asymptotics [Lahiri (1996)]. In
most applications of spatial econometrics, the implied structure is that of an increasing domain.
Finally, for spatial processes that contain spatial weights, the asymptotics require the use
of CLT and LLN for triangular arrays [Davidson (1994, Ch. 19, 24)]. This is caused by the fact that
for the boundary elements the “sample” weights matrix changes as new data points are added

12 Specifically, this may limit the applicability of GMM estimators that are based on a central limit theorem for
stationary mixing random fields such as the one by Bolthausen (1982), used by Conley (1996).
13 Cressie (1993, pp. 100–101).

— 10 —
[i.e., the new data points change the connectedness structure for existing data points].14 Again,
this is an additional degree of complexity, which is not found in time series models.

3. Spatial Regression Models

3.1. Spatial Lag and Spatial Error Models

In the standard linear regression model, spatial dependence can be incorporated in two distinct
ways: as an additional regressor in the form of a spatially lagged dependent variable (Wy), or in
the error structure ( E [ ε i ε j ] ≠ 0 ). The former is referred to as a spatial lag model and is appropriate
when the focus of interest is the assessment of the existence and strength of spatial interaction.
This is interpreted as substantive spatial dependence in the sense of being directly related to a
spatial model [e.g., a model that incorporates spatial interaction, yardstick competition, etc.]. Spa-
tial dependence in the regression disturbance term, or a spatial error model can take any of the
forms outlined in Section 2.3. and is referred to as nuisance dependence. This is appropriate
when the concern is with correcting for the potentially biasing influence of the spatial autocorre-
lation, due to the use of spatial data (irrespective of whether the model of interest is spatial or not).
Formally, a spatial lag model, or a mixed regressive, spatial autoregressive model is
expressed as

y = ρWy + Xβ + ε (3.1)

where ρ is a spatial autoregressive coefficient, ε is a vector of error terms, and the other notation
is as before.15 Unlike what holds for the time series counterpart of this model, the spatial lag term
Wy is correlated with the disturbances, even when the latter are i.i.d. This can be seen from the
reduced form of (3.1),

–1 –1
y = ( I – ρW ) Xβ + ( I – ρW ) ε (3.2)

in which each inverse can be expanded into an infinite series as in (2.6), including both the
explanatory variables and the error terms at all locations (the spatial multiplier). Consequently,
the spatial lag term must be treated as an endogenous variable and proper estimation methods
must account for this endogeneity (OLS will be biased and inconsistent due to the simultaneity
bias).
A spatial error model is a special case of a regression with a non-spherical error term, in
which the off-diagonal elements of the covariance matrix express the structure of spatial depen-

14 See Kelejian and Prucha (1999a, 1999b).

15 For ease of exposition, the error term is assumed to be i.i.d., although various forms of heteroskedasticity can
be incorporated in a straightforward way [Anselin (1988a), Chapter 6].

— 11 —
dence. Consequently, OLS remains unbiased, but it is no longer efficient and the classical estima-
tors for standard errors will be biased. As outlined in Section 2.3., the spatial structure can be
specified in a number of different ways, and (except for the non-parametric approaches) results
in a error variance covariance matrix of the form

E [ εε′ ] = Ω ( θ ) , (3.3)

where θ is a vector of parameters, such as the coefficients in a SAR or SMA error process. When
the error process is SAR, the resulting model can also be expressed as a spatial lag specification,
in the form of spatial Durbin or spatial common factor model [Anselin (1980)]. The SAR error
model is

y = Xβ + ε and ε = λWε + u . (3.4)

–1 –1
Since ε = ( I – λW ) u , and thus y = Xβ + ( I – λW ) u , (3.4) is equivalent to

y = λWy + Xβ – λWXβ + ε , (3.5)

which is a spatial lag model with an additional set of spatially lagged exogenous variables (WX)
and a set of k nonlinear (common factor) constraints on the coefficients (the product of the spatial
autoregressive coefficient with the regression coefficients β should equal the negative of the coef-
ficients of WX). The similarity between the error model (3.5) and the “pure” spatial lag model
(3.1) will complicate specification testing in practice, since tests designed for a spatial lag alterna-
tive will also have power against a spatial error alternative, and vice versa.
Most spatial regression models used in practice are based on a single spatial weights
matrix. However, in principle, higher order models are possible as well, such as higher order
SAR models [e.g., Brandsma and Ketellapper (1979), Blommestein (1983, 1985)], spatial autore-
gressive, moving average specifications or SARMA [Huang (1984)], and models that include
both a spatially lagged dependent variable as well as a SAR error process [Case (1992)]. In the
specification of such higher order models, care must be taken to ensure that the weights are
unique, orthogonal (non-overlapping) and that all coefficients are identified.16

3.2. Spatial Dependence in Panel Data Models

When observations are available across space as well as over time, the additional dimension
allows the estimation of the full covariance of one type of association, using the other dimension
to provide the asymptotics (e.g., in SUR models with N << T). However, as in the pure cross-sec-
tional case, there is insufficient information in the NT observations to estimate the complete

16 For a more extensive discussion, see Blommestein (1985), Anselin and Smirnov (1996), and Anselin and Bera
(1998, pp. 251–252).

— 12 —
 2
( NT ) covariance matrix Cov [ y it y js ] ≠ 0 , (with i ≠ j and t ≠ s ) without imposing some structure.
For small N and large T, the asymptotics in the time domain can be exploited to obtain a non-
parametric estimate of cross-sectional dependence, while time dependence must be parameter-
ized. Similarly, for large N and small T, the asymptotics in the spatial domain can be exploited to
yield a non-parametric estimate of serial (time) dependence, while spatial dependence must be
parameterized. As in the pure cross-sectional case, the latter requires the use of a spatial weights
matrix. In each of these situations, asymptotics are only needed in one of the dimensions while
the other can be treated as fixed.
When both spatial as well as serial dependence are parameterized, a range of specifica-
tions can be considered, allowing different combinations of the two. For ease of exposition,
assume that the observations are stacked by time period, i.e., they can be considered as T time
slices of N cross-sectional units. Restricting attention to “lag” dependence, and with f(z) as a
generic designation for the regressors (which may be lagged in time and/or space), four types of
models can be distinguished.
(a) pure space-recursive, in which the dependence pertains to neighboring locations in a differ-
ent period, or,

y it = γ [ Wy t – 1 ] i + f ( z ) + ε it , (3.6)

where, using the same notational convention as in (2.2), [ Wy t – 1 ] i is the i-th element of

the spatial lag vector applied to the observations on the dependent variable in the previ-
ous time period (using a N by N spatial weights matrix for the cross-sectional units).
(b) time-space recursive, in which the dependence relates to the same location as well as the
neighboring locations in another period, or,

y it = λy it – 1 + γ [ Wy t – 1 ] + f ( z ) + ε it (3.7)
i

(c) time-space simultaneous, with both a time-wise and a spatially lagged dependent variable,
or,

y it = λy it – 1 + ρ [ Wy t ] i + f ( z ) + ε it (3.8)

where [ Wy t ] is the i-th element of the spatial lag vector in the same time period.
i
(d) time-space dynamic, with all forms of dependence, or,

y it = λy it – 1 + ρ [ Wy t ] + γ [ Wy t – 1 ] + f ( z ) + ε it . (3.9)
i i

In order to estimate the parameters of the time-space simultaneous model, asymptotics are

— 13 —
needed in the cross-sectional dimension, while for the time-space dynamic model, asymptotics
are needed in both dimensions. For the other models, the type of asymptotics required are deter-
mined by the dependence structure in the error terms. For example, the pure space-recursive
model with i.i.d. errors satisfies the assumptions of the classical linear model and can be esti-
mated by means of OLS.
Spatial lag and spatial error dependence can be introduced into the cross-sectional
dimension of traditional panel data models in a straightforward way. For example, in a spatial
SUR model, both autoregressive as well as regression parameters are allowed to vary by time
period, in combination with a non-parametric serial covariance. The spatial lag formulation of
such a model would be (in the same notation as before):

y it = ρ t [ Wy t ] i + x′ it β t + ε it (3.10)

2
with Var [ ε it ] = σ t and E [ ε it ε is ] = σ ts .17
An important issue to consider when incorporating spatial dependence in panel data
models is the extent to which fixed effects may be allowed. Since the estimation of the spatial
process models requires asymptotics in the cross-sectional domain ( N → ∞ ), fixed effects (i.e., a
dummy variable for each location) would suffer from the incidental parameter problem and no
consistent estimator exists. Hence, fixed cross-sectional effects are incompatible with spatial pro-
cesses and instead a random effects specification must be considered.

3.3. Spatial Dependence in Models for Qualitative Data

Empirical analysis of interacting agents requires models that incorporate spatial dependence for
discrete dependent variables, such as counts or binary outcomes [Brock and Durlauf (1995)]. This
turns out to be quite complex and continues to be an active area of research.18 While an extensive
discussion of the technical aspects associated with spatial discrete choice models is beyond the
scope of the current chapter, the salient issues may be illustrated with a spatial version of the pro-
bit model, which has recently received considerable attention.19
The point of departure is the familiar expression for a linear model in a latent (unob-
served) dependent variable y i∗

y i∗ = x i ′β + ε i (3.11)

17 For further details, see Anselin (1988a, 1988b). A recent application is Baltagi and Li (1999).
18 For an extensive treatment of these so-called auto-models, see Besag (1974), and also Cressie and Read (1989),
Cressie (1993, pp. 427–428), and Kaiser and Cressie (1997).
19 Methodological issues associated with spatial probit models are considered in Case (1992), McMillen (1992,
1995), Bolduc, Fortin and Gordon (1997), Heagerty and Lele (1998), Pinkse and Slade (1998) and Beron and
Vijverberg (1999).

— 14 —
where ε i is a random variate for which a given distribution is assumed (e.g., the normal for the
probit model). The realization of y i∗ is observed in the form of discrete events, y i = 1 for
y i∗ ≥ 0 , and y i = 0 for y i∗ < 0 . The discrete events are related to the underlying probability
model through the error term, for example, y i∗ ≥ 0 implies – x i ′β < ε i , and, therefore,

E [ y i ] = P [ y i = 1 ] = Φ [ x i ′β ] (3.12)

where Φ is the cumulative distribution function for the standard normal.

Spatial autocorrelation can be introduced into this model in the form of a spatial autore-
gressive process for the error term ε i in (3.11), or

ε i = λ ∑ w ij ε j + u i (3.13)
j

where λ is an autoregressive parameter, the w ij are the elements in the i-th row of a spatial
weights matrix, and u i may be assumed to be i.i.d. standard normal. As a consequence of the
spatial multiplier in the autoregressive specification, the random error at each location now
becomes a function of the random errors at all other locations as well. Its distribution is multi-
variate normal with N by N variance-covariance matrix

–1
E [ εε′ ] = [ ( I – λW )′ ( I – λW ) ] , (3.14)

As pointed out in Section 2.3., besides being non-diagonal, (3.14) is also heteroskedastic. Conse-
quently, the usual inequality conditions that are at the basis of (3.12) no longer hold, and
P [ – x i ′β < ε i ] cannot be derived from the univariate standard normal distribution, but rather
must be expressed explicitly as the marginal distribution of a N-dimensional multivariate normal
vector, whose variance-covariance matrix contains the autoregressive parameter λ . This is non-
standard and typically not analytically tractable, which greatly complicates estimation and spec-
ification testing. Similar issues are faced in the spatial lag model for a latent variable.20
When the dependent variable is a count or proportion, the latent variable model is not
used, but instead the dependent variables is modeled directly as a random variable with a Poisson
or binomial distribution. Introducing spatial dependence into these models is not trivial, since the
multivariate versions of discrete distributions are not as tractable as the multivariate normal. This
is typically handled by means of hierarchical Bayesian modeling and estimation is based on com-
putation intensive methods such as Gibbs sampling and Markov Chain Monte Carlo.21

20 For an extensive discussion, see Beron and Vijverberg (1999).

21 For a recent review of the issues, see, e.g., Clayton and Kaldor (1987), Besag, York and Mollie (1991), Besag et
al. (1995), Waller, Carlin and Xia (1997), Waller et al. (1997) and Diggle, Tawn and Moyeed (1998).

— 15 —
4. Estimation

4.1. Maximum Likelihood Estimation

Maximum likelihood (ML) estimation of spatial lag and spatial error regression models was first
outlined by Ord (1975).22 The point of departure is an assumption of normality for the error
terms ε i in (3.1) or u i in (3.4). The joint likelihood then follows from the multivariate normal dis-
tribution for y. Unlike what holds for the classic regression model, the joint log likelihood for a
spatial regression does not equal the sum of the log likelihoods associated with the individual
observations. This is due to the two-directional nature of the spatial dependence, which results in
a Jacobian term that is the determinant of a full N by N matrix, e.g., I – ρW (spatial lag).
For the SAR error model, the log likelihood is based on the multivariate normal case, e.g.,
as used in the general treatments of Magnus (1978) and Breusch (1980). Since ε ∼ MVN ( 0, Σ ) , it
2 –1
follows that, with ε = y – Xβ and Σ = σ [ ( I – λW )′ ( I – λW ) ] ,

2
lnL = – ( N ⁄ 2 )ln ( 2π ) – ( N ⁄ 2 )lnσ + ln I – λW
(4.1)
2
– ( 1 ⁄ 2σ ) ( y – Xβ )′ ( I – λW )′ ( I – λW ) ( y – Xβ )

Closer inspection of the last term in (4.1) reveals that, conditional upon λ , a maximization of the
log likelihood is equivalent to the minimization of the sum of squared residuals in a regression of
a spatially filtered dependent variable y∗ = y – λWy on a set of spatially filtered explanatory
variables X∗ = X – λWX . The first order conditions for β̂ ML indeed yield the familiar general-
ized least squares estimator:

–1
β̂ ML = [ ( X – λWX )′ ( X – λWX ) ] ( X – λWX )′ ( y – λWy ) (4.2)
2
and, similarly, the ML estimator for σ follows as:

2
σ̂ ML = ( e – λWe )′ ( e – λWe ) ⁄ N (4.3)

with e = y – Xβ̂ ML . However, unlike the time series case, a consistent estimator for λ cannot be
obtained from the OLS residuals and therefore the standard two-step FGLS approach does not
apply.23 Instead, the estimator for λ must be obtained from an explicit maximization of a concen-
trated likelihood function [for details, see Anselin (1988a, Chapter 6) and Anselin and Bera
(1998)]. Also, while the SAR error model estimator (4.2) yields an expression in spatially filtered

22 Other classic treatments of ML estimation in spatial models can be found in Whittle (1954), Besag (1974),
Mardia and Marshall (1984), Warnes and Ripley (1987), and Mardia and Watkins (1989).
23 For a formal demonstration, see Anselin (1988a). In Kelejian and Prucha (1997) it is also shown that an
approach based on the spatial Durbin specification yields an inconsistent estimator.

— 16 —
variables, this is not the case for the SMA error model, since its variance (2.8) requires the inverse
of a N by N matrix in order to obtain a FGLS expression. This model has therefore not seen much
application in empirical practice [see, e.g., Sneek and Rietveld (1999)].24
The log-likelihood for the spatial lag model is obtained using the same general principles
[see Anselin (1988, Chapter 6) for details] and takes the form

2
lnL = – ( N ⁄ 2 )ln ( 2π ) – ( N ⁄ 2 )lnσ + ln I – ρW
(4.4)
2
– ( 1 ⁄ 2σ ) ( y – ρWy – Xβ )′ ( y – ρWy – Xβ )

The minimization of the last term in (4.4) corresponds to OLS, but since this ignores the log Jaco-
bian ln I – ρW , OLS is not a consistent estimator in this model. As in the spatial error model,
there is no satisfactory two-step procedure and estimators for the parameters must be obtained
from an explicit maximization of the likelihood. This is greatly simplified since both β̂ ML and
2
σ̂ ML can be obtained conditional upon ρ from the first order conditions:

–1
β̂ ML = ( X′X ) X′ ( y – λWy ) (4.5)
–1 –1
or, with β̂ 0 = ( X′X ) X′y , e 0 = y – Xβ̂ 0 , β̂ L = ( X′X ) X′Wy , e L = y – Xβ̂ L ,

β̂ ML = β̂ 0 – ρβ̂ L (4.6)

and

2
σ̂ ML = ( e 0 – ρe L )′ ( e 0 – ρe L ) ⁄ N (4.7)

This yields a concentrated log-likelihood in a single parameter, which is straightforward to opti-

mize by means of direct search techniques [see Anselin (1980, 1988a) for derivations and details].
Both spatial lag and spatial error models are special cases of a more general specification
that may include forms of heteroskedasticity as well. This also provides the basis for ML estima-
tion of spatial SUR models with spatial lag or spatial error terms [Anselin (1980, Ch. 10)]. Simi-
larly, ML estimation of error components models with spatial lag or spatial error terms can be
implemented as well. Spatial models with discrete dependent variables are typically not esti-
mated by means of ML, given the prohibitive nature of evaluating multiple integrals to deter-
mine the relevant marginal distributions.25

24 Note that the spatial error components specification (2.10) and the direct representation models (2.12) also
suffer from this problem.
25 For details, see, e.g., McMillen (1992), Pinkse and Slade (1998), Beron and Vijverberg (1999), and also, for
general principles, Poirier and Ruud (1988).

— 17 —
 Finally, it is important to note that models with spatial dependence do not fit the classical
framework [e.g., as outlined in Rao (1973)] under which the optimal properties (consistency,
asymptotic efficiency, asymptotic normality) of ML estimators are established. This implies that
these properties do not necessarily hold and that careful consideration must be given to the explicit
formulation of regularity conditions. In general terms, aside from the usual restrictions on the vari-
ance and higher moments of the model variables, these conditions boil down to constraints on the
range of dependence embodied in the spatial weights matrix.26 In addition, to avoid singularity or
explosive processes, the parameter space for the coefficient in a spatial process model is restricted
to an interval other than the familiar −1, +1. For example, for a SAR process, the parameter space is
1 ⁄ ω min < ρ < 1 ⁄ ω max , where ω min and ω max are the smallest (on the real line) and largest eigen-
values of the spatial weights matrix W. For row-standardized weights, ω max = 1 , but ω min > – 1 ,
such that the lower bound on the parameter space is less than −1 [Anselin (1980)]. This must be
taken into account in practical implementations of estimation routines.

4.2. Spatial Two Stage Least Squares

The endogeneity of the spatially lagged dependent variable can also be addressed by means of
an instrumental variables or two stage least squares approach [Anselin (1980, 1988a, 1990b),
Land and Deane (1992), Kelejian and Robinson (1993), Kelejian and Prucha (1998)]. As demon-
strated in Kelejian and Robinson (1993), the choice of an instrument for Wy follows from the con-
ditional expectation in the reduced form (3.2),

–1 2 2
E [ y X ] = ( I – ρW ) Xβ = Xβ + ρWXβ + ρ W Xβ + … . (4.8)

Apart from the exogenous variables X (which are always instruments), this includes their spatial
lags as well, suggesting WX as a set of instruments.27
Under a set of reasonable assumptions that are easily satisfied when the spatial weights
are based on contiguity, the spatial two stage least squares estimator achieves the consistency
and asymptotic normality properties of the standard 2SLS [see, e.g., the theorems spelled out in
Schmidt (1976)].28 A straightforward extension is the application of 3SLS to the spatial SUR
model with a spatial lag [Anselin (1988a, Chapter 10)]. However, 2SLS is not appropriate to
obtain a consistent estimator for the nuisance parameter in a spatial error model, for example
based on the spatial Durbin formulation [(3.5)], as demonstrated by Kelejian and Prucha (1997).

26 For a careful consideration of these issues, see Kelejian and Prucha (1999a).
27 Higher lags may be included as well to improve precision of the estimator, although this may also increase its
bias. When other (i.e., non-spatial) endogenous variables are included in the specification, they will require
instruments as well, in the standard fashion.
28 For technical details, see, e.g., Kelejian and Robinson (1993), Kelejian and Prucha (1998).

— 18 —
The spatial 2SLS model is extended to include spatial error components in Kejejian and Robinson
(1993) and SAR errors in Kelejian and Prucha (1998).

4.3. Method of Moments Estimators

Recently, a number of approaches have been outlined to estimate the coefficients in a spatial
error model as an application of general principles underlying the method of moments. Kelejian
and Prucha (1999a) develop a set of moment conditions that yield estimation equations for the
parameter of a SAR error model. Specifically, assuming an i.i.d. error vector for u in (3.4), the fol-
lowing three conditions readily follow

2
E [ u′u ⁄ N ] = σ
2
E [ u′W′Wu ⁄ N ] = σ ( 1 ⁄ N )tr ( W′W ) (4.9)
E [ u′Wu ⁄ N ] = 0

where tr is the matrix trace operator. Replacing u by e – λWe (with e as the vector of OLS resid-
2 2
uals) in (4.9) yields a system of three equations in the parameters λ , λ and σ . Kelejian and
Prucha (1999a) suggest the use of nonlinear least squares to obtain a consistent generalized
moment estimator for λ from this system, which can then be used to obtain consistent estimators
for the β in a FGLS approach. Since the λ is considered as a nuisance parameter, its significance
(as a test for spatial autocorrelation) cannot be assessed, but its role is to provide a consistent esti-
mator for the regression coefficients.29
A different approach is taken in the application of Hansen’s (1982) generalized method of
moments estimator (GMM) to spatial error autocorrelation in Conley (1996). This estimator is the
standard minimizer of a quadratic form in the sample moment conditions, where the covariance
matrix is obtained in non-parametric form as an application of the ideas of Newey and West
(1987). Specifically, the spatial covariances are estimated from weighted averages of sample
covariances for pairs of observations that are within a given distance band from each other. Note
that this approach requires covariance stationarity, which is only satisfied for a restricted set of
spatial processes (e.g., it does not apply to SAR error models).
Pinkse and Slade (1998) use a set of moment conditions to estimate a probit model with
SAR errors. However, they focus on the induced heteroskedasticity of the process and do not
explicitly deal with the spatial covariance structure.30

29 A recent application of this method is given in Bell and Bockstael (1999). An extension of this idea to the
residuals of a spatial 2SLS estimation is provided in Kelejian and Prucha (1998).
30 See also Case (1992) and McMillen (1992) for a similar focus on heteroskedasticity in the spatial probit model.

— 19 —
4.4. Other Estimation Methods
A number of other approaches have been suggested to deal with the estimation of spatial regres-
sion models. An early technique is the so-called coding method, originally examined in Besag
and Moran (1975) and Besag (1977).31 This approach consists of selecting a subsample from the
data such that the relevant neighbors are removed (a non-contiguous subsample). This in effect
eliminates the simultaneity bias in the spatial lag model, but at the cost of converting the model
to a conditional one and with a considerable reduction of the sample size (down to 20% of the
original sample for irregular lattice data). The advantage of this approach is that standard meth-
ods may be applied (e.g., for discrete choice models). However, it is not an efficient procedure
and considerable arbitrariness is involved in the selection of the coding scheme.
Another increasingly common approach consists of the application of computational esti-
mators to spatial models. A recent example is the recursive importance sampling (RIS) estimator
[Vijverberg (1997)] applied to the spatial probit model in Beron and Vijverberg (1999).
A considerable literature also exists on Bayesian estimation of spatial models, but a
detailed treatment of this is beyond the current scope.32

5. Specification Tests

5.1. Moran’s I
The most commonly used specification test for spatial autocorrelation is derived from a statistic
developed by Moran (1948, 1950a, 1950b) as the two-dimensional analog of a test for univariate
time series correlation [see also Cliff and Ord (1972, 1973)]. In matrix notation, Moran’s I statistic
is

I = ( N ⁄ S 0 ) ( e′We ⁄ e′e ) (5.1)

with e as a vector of OLS residuals and S 0 = ∑ ∑ w ij , a standardization factor that corresponds

i j
to the sum of the weights for the non-zero cross-products. The statistic shows a striking similar-
ity to the familiar Durbin-Watson test.33 Inference for Moran’s I is based on a normal approxima-
tion, using a standardized z-value obtained from expressions for the mean and variance of the
statistic [Cliff and Ord (1972, 1981)]. Alternatively, as shown in Tiefelsdorf and Boots (1995), an
exact test can be constructed based on numerical integration and the principles outlined in Imhof
(1961) and Koerts and Abrahamse (1968).

31 See also the discussion in Haining (1990, pp. 131–133).

32 For a review, see the references in footnote 21, and also Hepple (1995) and LeSage (1997).
33 For example, for row-standardized weights, S 0 = N , and I = e′We ⁄ e′e . See Anselin and Bera (1998) for an
extensive discussion.

— 20 —
 Moran’s I test has been shown to be locally best invariant [King (1981)] and consistently
outperforms other tests in terms of power in simulation experiments [see, e.g., Bartels and
Hordijk (1977), Anselin and Rey (1991), Anselin and Florax (1995b), Kelejian and Robinson
(1998)]. Its application has been extended to residuals in 2SLS regression in Anselin and Kelejian
(1997), and to generalized residuals in probit models in Pinkse (1998, 1999). General formal con-
ditions and proofs for the asymptotic normality of Moran’s I in a wide range of regression mod-
els are given in Pinkse (1998) and Kelejian and Prucha (1999b). The consideration of Moran’s I in
conjunction with spatial heteroskedasticity is covered in Kelejian and Robinson (1998, 1999).34

5.2. ML Based tests

When spatial regression models are estimated by maximum likelihood, inference on the spatial
autoregressive coefficients may be based on a Wald or asymptotic t-test (from the asymptotic
variance matrix) or on a likelihood ratio test [see Anselin (1988, Chapter 6), Anselin and Bera
(1998)]. Both approaches require that the alternative model (i.e., the spatial model) be estimated.
In contrast, a series of test statistics based on the Lagrange Multiplier (LM) or Rao Score (RS)
principle only require estimation of the model under the null. The LM/RS tests also allow for the
distinction between a spatial error and a spatial lag alternative.35
A LM/RS test against a spatial error alternative was originally suggested by Burridge
(1980) and takes the form

2 2
LM err = [ e′We ⁄ ( e′e ⁄ N ) ] ⁄ [ tr ( W + W′W ) ] (5.2)
2
This statistic has an asymptotic χ ( 1 ) distribution and, apart from a scaling factor, corresponds
to the square of Moran’s I.36 From several simulation experiments [Anselin and Rey (1991),
Anselin and Florax (1995b)] if follows that Moran’s I has slightly better power than the LM err
test in small samples, but the performance of both tests becomes indistinguishable in medium
and large size samples. The LM/RS test against a spatial lag alternative was outlined in Anselin
(1988c) and takes the form

2
LM lag = [ e′Wy ⁄ ( e′e ⁄ N ) ] ⁄ D (5.3)

–1 2 2
where D = [ ( WXβ )′ ( I – X ( X′X ) X′ ) ( WXβ ) ⁄ σ ] + tr ( W + W′W ) .This statistic also has an
2
asymptotic χ ( 1 ) distribution.

34 See also Kelejian and Robinson (1992) for an alternative test statistic.
35 Moran’s I is not based on an explicit alternative and has power against both [see Anselin and Rey (1991)].
36 As shown in Anselin and Kelejian (1997) these tests are asymptotically equivalent.

— 21 —
 Since both tests have power against the other alternative, it is important to take account
of possible lag dependence when testing for error dependence and vice versa. This can be imple-
mented by means of a joint test [Anselin (1988c)] or by constructing tests that are robust to the
presence of local misspecification of the other form [Bera and Yoon (1993), Anselin et al. (1996)].
The LM/RS principle can also be extended to more complex spatial alternatives, such as
higher order processes, spatial error components and direct representation models [Anselin
(1999)], to panel data settings [Anselin (1988b)], and to probit models [Pinkse (1998, 1999), Pinkse
and Slade (1998)]. A common characteristics of the LM/RS tests against spatial alternatives is
2
that they do not lend themselves readily to a formulation as a NR expression based on an auxil-
iary regression. However, as recently shown in Baltagi and Li (1998), it is possible to obtain tests
for spatial lag and spatial error dependence in a linear regression model by means of Davidson
and MacKinnon’s (1984, 1988) double length artificial regression approach.

6. Implementation Issues
To date, spatial econometric methods are not found in the main commercial econometric and sta-
tistical software packages, although macro and scripting facilities may be used to implement
some estimators [Anselin and Hudak (1992)]. The only comprehensive software to handle both
estimation and specification testing of spatial regression models is the special-purpose SpaceStat
package [Anselin (1992b, 1998b)]. Maximum likelihood estimation of spatial models is also
included in the Matlab routines of Pace and Barry (1998) and estimation of spatial error models is
part of the S+Spatialstats add-on to S-Plus [MathSoft (1996)].37
In contrast to maximum likelihood estimation, method of moments and 2SLS can easily
be implemented with standard software, provided that spatial lags can be computed. This
requires the construction of a spatial weights matrix which must often be derived from informa-
tion in a geographic information system [e.g., Can (1996)]. Similarly, once a spatial lag can be
computed, the LM/RS statistics are straightforward to implement.
The main practical problem is encountered in maximum likelihood estimation where the
Jacobian determinant must be evaluated for every iteration in a nonlinear optimization proce-
dure. The original solution to this problem was suggested by Ord (1975), who showed how the
log Jacobian can be decomposed in terms that contain the eigenvalues of the weights matrix ω i ,

37 Neither of these toolboxes include specification tests and S+Spatialstats has no routines to handle the spatial
lag model.

— 22 —
 n
ln I – ρW = ∑ ln ( 1 – ρω i ) . (6.1)
i=1

This is easy to implement in a standard optimization routine by treating the individual elements
in the sum as observations on an auxiliary term in the log-likelihood [see Anselin and Hudak
(1992)]. However, the computation of the eigenvalues quickly becomes numerically unstable for
matrices of more than 1,000 observations. In addition, for large data sets this approach is ineffi-
cient in that it does not exploit the high degree of sparsity of the spatial weights matrix. Recently
suggested solutions to this problem fall into two categories. Approximate solutions avoid the
computation of the Jacobian determinant, but instead approximate it by a polynomial function or
by means of simulation methods [e.g., Griffith (1992), Griffith and Sone (1995), Martin (1993),
Barry and Pace (1999)]. Exact solutions are based on Cholesky or LU decomposition methods
that exploit the sparsity of the weights [Pace (1997), Pace and Barry (1997a, 1997b)], or use a char-
acteristic polynomial approach [Smirnov and Anselin (1999)]. While much progress has been
made, considerable work remains to be done to develop efficient algorithms and data structures
to allow for the analysis of very large spatial data sets.

7. Concluding Remarks
This review chapter has been an attempt to present the salient issues pertaining to the
methodology of spatial econometrics. It is by no means complete, but it is hoped that sufficient
guidance is provided to pursue interesting research directions. Many challenging problems
remain, both methodological in nature as well as in terms of applying the new techniques to
meaningful empirical problems. Particularly in dealing with spatial effects in models other than
the standard linear regression, much needs to be done to complete the spatial econometric tool-
box. It is hoped that the review presented here will stimulate statisticians and econometricians to
tackle these interesting and challenging problems.

References
Akerlof, George A. (1997). Social distance and social decisions. Econometrica 65, 1005–1027.
Alessie, Rob and Arie Kapteyn (1991). Habit formation, interdependent preferences and demo-
graphic effects in the almost ideal demand system. Economic Journal 101, 404–419.
Anselin, L. (1980). Estimation methods for spatial autoregressive structures. Regional Science Disserta-
tion and Monograph Series 8. Field of Regional Science, Cornell University, Ithaca, N.Y.
Anselin, L. (1988a). Spatial Econometrics: Methods and Models. Kluwer Academic, Dordrecht.
Anselin, L. (1988b). A test for spatial autocorrelation in seemingly unrelated regressions. Econom-
ics Letters 28, 335–341.
Anselin, L. (1988c). Lagrange multiplier test diagnostics for spatial dependence and spatial heter-

— 23 —
 ogeneity. Geographical Analysis 20, 1–17.
Anselin, L. (1989). What is special about spatial data? Alternative perspectives on spatial data
analysis. Technical Report 89–4, National Center for Geographic Information and Analysis,
University of California, Santa Barbara, CA [contained on CD-Rom “Fundamental research
in geographic information and analysis”, Santa Barbara, Ca: NCGIA]
Anselin, L. (1990a). Spatial dependence and spatial structural instability in applied regression
analysis. Journal of Regional Science 30, 185–207.
Anselin, L. (1990b). Some robust approaches to testing and estimation in spatial econometrics.
Regional Science and Urban Economics 20, 141–163.
Anselin, L. (1992a). Space and applied econometrics. Special Issue, Regional Science and Urban
Economics 22.
Anselin, L. (1992b). SpaceStat, a Software Program for the Analysis of Spatial Data. National Center
for Geographic Information and Analysis, University of California, Santa Barbara, CA.
Anselin, L. (1998a). GIS research infrastructure for spatial analysis of real estate markets. Journal
of Housing Research 9, 113–33.
Anselin, L. (1998b). SpaceStat Version 1.90. http://www.spacestat.com.
Anselin, L. (1999). Rao’s score test in spatial econometrics. Journal of Statistical Planning and Infer-
ence (forthcoming).
Anselin, L. and A. Bera (1998). Spatial dependence in linear regression models with an introduc-
tion to spatial econometrics. In: A. Ullah and D. E. A. Giles, Eds., Handbook of Applied Eco-
nomic Statistics, pp. 237–289. New York: Marcel Dekker.
Anselin, L. and R. Florax (1995a). Introduction. In L. Anselin and R. Florax (Eds.), New Directions
in Spatial Econometrics, pp. 3–18. Berlin: Springer-Verlag.
Anselin, L. and R. Florax (1995b). Small sample properties of tests for spatial dependence in
regression models: some further results. In L. Anselin and R. Florax (Eds.), New Directions in
Spatial Econometrics, pp.21–74. Berlin: Springer-Verlag.
Anselin, L. and D. Griffith (1988). Do spatial effects really matter in regression analysis. Papers,
Regional Science Association 65, 11–34.
Anselin, L. and S. Hudak (1992). Spatial econometrics in practice, a review of software options.
Regional Science and Urban Economics 22, 509–536.
Anselin, L. and H. H. Kelejian (1997). Testing for spatial error autocorrelation in the presence of
endogenous regressors. International Regional Science Review 20, 153–182.
Anselin, L. and S. Rey (1991). Properties of tests for spatial dependence in linear regression mod-
els. Geographical Analysis 23, 112–131.
Anselin, L. and S. Rey (1997). Introduction to the special issue on spatial econometrics. Interna-
tional Regional Science Review 20, 1–7.
Anselin, L. and O. Smirnov (1996). Efficient algorithms for constructing proper higher order spa-
tial lag operators. Journal of Regional Science 36, 67–89.
Anselin, L., A. Bera, R. Florax and M. Yoon (1996). Simple diagnostic tests for spatial depen-
dence. Regional Science and Urban Economics 26, 77–104.
Aoki, Masanao (1994). New macroeconomic modelling approaches. Hierarchical dynamics and
mean-field approximation. Journal of Economic Dynamics and Control 18, 865–77.
Aoki, Masanao (1996). New Approaches to Macroeconomic Modeling. Cambridge: Cambridge Uni-
versity Press.
Arthur, W. Brian (1989). Increasing returns, competing technologies and lock-in by historical small
events: the dynamics of allocation under increasing returns to scale. Economic Journal 99, 116–31.
Aten, Bettina (1996). Evidence of spatial autocorrelation in international prices. Review of Income
and Wealth 42, 149–63.
Aten, Bettina (1997). Does space matter? International comparisons of the prices of tradables and
nontradables. International Regional Science Review 20, 35–52.

— 24 —
Baltagi, Badi and Dong Li (1998). Double length artificial regressions for testing spatial depen-
dence. Working Paper, Department of Economics, Texas A&M University, College Station, TX.
Baltagi, Badi and Dong Li (1999). Prediction in the panel data model with spatial correlation. In
L. Anselin and R. Florax (Eds.), Advances in Spatial Econometrics. Heidelberg: Springer-Verlag
(forthcoming).
Barry, R.P. and R. K. Pace (1999). Monte Carlo estimates of the log determinant of large sparse
matrices. Linear Algebra and its Applications 289, 41–54.
Bartels, C.P.A. and L. Hordijk (1977). On the power of the generalized Moran contiguity coeffi-
cient in testing for spatial autocorrelation among regression disturbances. Regional Science and
Urban Economics 7, 83–101.
Basu, Sabyasachi and Thomas G. Thibodeau (1998). Analysis of spatial autocorrelation in hous-
ing prices. The Journal of Real Estate Finance and Economics 17, 61–85.
Bell, Kathleen P. and Nancy E. Bockstael (1999). Applying the generalized moments estimation
approach to spatial problems involving micro-level data. The Review of Economics and Statis-
tics 81 (in press)
Benirschka, M. and J.K. Binkley (1994). Land price volatility in a geographically dispersed mar-
ket. American Journal of Agricultural Economics 76, 185–95.
Bera, A. and M. J. Yoon (1993). Specification testing with misspecified alternatives. Econometric
Theory 9, 649–658.
Beron, Kurt J. and Wim P.M. Vijverberg (1999). Probit in a spatial context: a Monte Carlo
approach. In L. Anselin and R. Florax (Eds.), Advances in Spatial Econometrics. Heidelberg:
Springer-Verlag (forthcoming).
Besag, J. (1974). Spatial interaction and the statistical analysis of lattice systems. Journal of the
Royal Statistical Society B 36, 192–225.
Besag, J. (1977). Efficiency of pseudo-likelihood estimators for simple Gaussian fields. Biometrika
64, 616–618.
Besag, J. and P.A.P. Moran (1975). On the estimation and testing of spatial interaction in Gaussian
lattice processes. Biometrika 62, 555–562.
Besag, J., J. York and A. Mollie (1991). Bayesian image restoration with two applications in spatial
statistics. Annals of the Institute of Statistical Mathematics 43, 1–59.
Besag, J., P. Green, D. Higdon and K. Mengersen (1995). Bayesian computation and stochastic
systems. Statistical Science 10, 3–66.
Besley, T. and A. Case (1995). Incumbent behavior: vote-seeking, tax-setting and yardstick com-
petition. American Economic Review 85, 25–45.
Bivand, R. and S. Szymanski (1997). Spatial dependence through local yardstick competition:
theory and testing. Economics Letters 55, 257–265.
Blommestein, H. (1983). Specification and estimation of spatial econometric models: a discussion
of alternative strategies for spatial economic modelling. Regional Science and Urban Economics
13, 250–271.
Blommestein, H. (1985). Elimination of circular routes in spatial dynamic regression equations.
Regional Science and Urban Economics 15, 121–130.
Bockstael, Nancy E. (1996). Modeling economics and ecology: the importance of a spatial per-
spective. American Journal of Agricultural Economics 78, 1168–1180.
Bolduc, Denis, Bernard Fortin and Stephen Gordon (1997). Multinomial probit estimation of spa-
tially interdependent choices: an empirical comparison of two new techniques. International
Regional Science Review 20, 77–101.
Bolthausen, E. (1982). On the central limit theorem for stationary mixing random fields. The
Annals of Probability 10, 1047–1050.
Borjas, George J. (1995). Ethnicity, neighborhoods, and human-capital externalities. American Eco-
nomic Review 85, 365–90.

— 25 —
Brandsma, A.S. and R.H. Ketellapper (1979). A biparametric approach to spatial autocorrelation.
Environment and Planning A 11, 51–58.
Brett, Craig and Joris Pinkse (1997). Those taxes all over the map! A test for spatial independence
of municipal tax rates in British Columbia. International Regional Science Review 20, 131–51.
Breusch, T. (1980). Useful invariance results for generalized regression models. Journal of Econo-
metrics 13, 327–340.
Brock, William A. and Steven N. Durlauf (1995). Discrete choice with social interactions I: Theory.
NBER Working Paper No. W5291. Cambridge, MA: National Bureau of Economic Research.
Brueckner, Jan K. (1998). Testing for strategic interaction among local governments: the case of
growth controls. Journal of Urban Economics 44, 438–67.
Burridge, P. (1980). On the Cliff-Ord test for spatial autocorrelation. Journal of the Royal Statistical
Society B 42, 107–108.
Can, A. (1996). Weight matrices and spatial autocorrelation statistics using a topological vector
data model. International Journal of Geographical Information Systems 10, 1009–1017.
Can, A. (1998). GIS and spatial analysis of housing and mortgage markets. Journal of Housing
Research 9, 61–86.
Case, A. (1991). Spatial patterns in household demand. Econometrica 59, 953–965.
Case, A. (1992). Neighborhood influence and technological change. Regional Science and Urban
Economics 22, 491–508.
Case, A., H.S. Rosen and J.R. Hines (1993). Budget spillovers and fiscal policy interdependence:
evidence from the States. Journal of Public Economics 52, 285–307.
Clayton, D.G. and J. Kaldor (1987). Empirical Bayes estimates of age-standardized relative risks
for use in disease mapping. Biometrics 43, 671–681.
Cliff, A. and J. K. Ord (1972). Testing for spatial autocorrelation among regression residuals. Geo-
graphical Analysis 4, 267–284.
Cliff, A. and J. K. Ord (1973). Spatial Autocorrelation. London: Pion.
Cliff, A. and J. K. Ord (1981). Spatial Processes: Models and Applications. London: Pion.
Conley, T.G. (1996). Econometric modelling of cross-sectional dependence. Ph.D. Dissertation.
Department of Economics, University of Chicago, Chicago, IL.
Cressie, N. (1993). Statistics for Spatial Data. Wiley, New York.
Cressie, N. and T. Read (1989). Spatial data analysis of regional counts. Biometrical Journal 31,
699–719.
Davidson, J. (1994). Stochastic Limit Theory. Oxford: Oxford University Press.
Davidson, R. and J.G. MacKinnon (1984). Model specification tests based on artificial regressions.
International Economic Review 25, 485–502.
Davidson, R. and J.G. MacKinnon (1988). Double-length artificial regressions. Oxford Bulletin of
Economics and Statistics 50, 203–217.
Diggle, P., J.A. Tawn and R.A. Moyeed (1998). Model-based geostatistics. Applied Statistics 47,
299–350.
Doreian, P. (1980). Linear models with spatially distributed data, spatial disturbances or spatial
effects. Sociological Methods and Research 9, 29–60.
Driscoll, John C. and Aart C. Kraay (1998). Consistent covariance matrix estimation with spa-
tially dependent panel data. The Review of Economics and Statistics 80, 549–60.
Dubin, R. (1988). Estimation of regression coefficients in the presence of spatially autocorrelated
error terms. Review of Economics and Statistics 70, 466–474.
Dubin, R. (1992). Spatial autocorrelation and neighborhood quality. Regional Science and Urban
Economics 22, 433–452.
Durlauf, Steven N. (1994). Spillovers, stratification and inequality. European Economic Review 38,
836–845.
Durlauf, Steven N. (1997). Statistical mechanics approaches to socioeconomic behavior. In W.

— 26 —
 Brian Arthur, Steven N. Durlauf and David A. Lane (Eds.), The economy as an evolving complex
system II, pp. 81–104. Reading, MA: Addison-Wesley.
Fiebig, Denzil G. (1999). Seemingly unrelated regression. In Baltagi (Ed.), Companion in Theoretical
Econometrics. Oxford: Basil Blackwell.
Frees, Edward W. (1995). Assessing cross-sectional correlation in panel data. Journal of Economet-
rics 69, 393–414.
Geoghegan, J., L. Wainger and N. Bockstael (1997). Spatial landscape indices in a hedonic frame-
work: an ecological economics analysis using GIS. Ecological Economics 23, 251–64.
Glaeser, E. L., H. D. Kallal, J. A. Scheinkman and A. Schleifer (1992). Growth in cities. Journal of
Political Economy 100, 1126–52.
Glaeser, E. L., B. Sacerdote and J. A. Scheinkman (1996). Crime and social interactions. Quarterly
Journal of Economics 111, 507–48.
Griffith, D. (1992). Simplifying the normalizing factor in spatial autoregressions for irregular lat-
tices. Papers in Regional Science 71, 71–86.
Griffith, D. and A. Sone. (1995). Trade-offs associated with normalizing constant computational
simplifications for estimating spatial statistical models. Journal of Statistical Computation and
Simulation 51, 165–183.
Haining, R. (1990). Spatial Data Analysis in the Social and Environmental Sciences. Cambridge Uni-
versity Press, Cambridge.
Hansen, Lars P. (1982). Large sample properties of generalized method of moments estimators.
Econometrica 50, 1029–1054.
Heagerty, Patrick J. and Subhash R. Lele (1998). A composite likelihood approach to binary spa-
tial data. Journal of the American Statistical Association 93, 1099–1111.
Heckman, J. (1991). Identifying the and of past: distinguishing state dependence from heteroge-
neity. American Economic Review 81, 75–79.
Hepple, L. W. (1995). Bayesian techniques in spatial and network econometrics: 2. Computa-
tional methods and algorithms. Environment and Planning A 27, 447–469.
Holtz-Eakin, D. (1994). Public-sector capital and the productivity puzzle. Review of Economics and
Statistics 76, 12–21.
Hordijk, L. and J. Paelinck (1976). Some principles and results in spatial econometrics. Recherches
Economiques de Louvain 42, 175–97.
Huang, J. S. (1984). The autoregressive moving average model for spatial analysis. Australian
Journal of Statistics 26, 169–178.
Imhof, J.P. (1961). Computing the distribution of quadratic forms in normal variables. Biometrika
48, 419–426.
Ioannides, Y. M. (1990). Trading uncertainty and market form. International Economic Review 31,
619–38.
Ioannides, Y. M. (1997). Evolution of trading structures. In W. Brian Arthur, Steven N. Durlauf
and David A. Lane (Eds.), The economy as an evolving complex system II, pp. 129–167. Reading,
MA: Addison-Wesley.
Isard, Walter (1956). Location and Space-Economy. Cambridge: MIT Press.
Kaiser, Mark S. and Noel Cressie (1997). Modeling Poisson variables with positive spatial depen-
dence. Statistics and Probability Letters 35, 423–432.
Kelejian, H. and I. Prucha (1997). Estimation of spatial regression models with autoregressive
errors by two stage least squares procedures: a serious problem. International Regional Science
Review 20, 103–111.
Kelejian, H. and I. Prucha (1998). A generalized spatial two stage least squares procedure for esti-
mating a spatial autoregressive model with autoregressive disturbances. Journal of Real Estate
Finance and Economics 17, 99–121.
Kelejian, H. and I. Prucha (1999a). A generalized moments estimator for the autoregressive

— 27 —
 parameter in a spatial model. International Economic Review (forthcoming).
Kelejian, H.H. and I. Prucha (1999b). On the asymptotic distribution of the Moran I test statistic
with applications. Working Paper, Department of Economics, University of Maryland, Col-
lege Park, MD.
Kelejian, H.H. and D.P. Robinson (1992). Spatial autocorrelation: a new computationally simple
test with an application to per capita country police expenditures. Regional Science and Urban
Economics 22, 317–333.
Kelejian, H.H. and D.P. Robinson (1993). A suggested method of estimation for spatial interde-
pendent models with autocorrelated errors, and an application to a county expenditure
model. Papers in Regional Science 72, 297–312.
Kelejian, H.H. and D.P. Robinson (1995). Spatial correlation: a suggested alternative to the
autoregressive model. In L. Anselin and R. Florax (Eds.), New Directions in Spatial Economet-
rics, pp. 75–95. Berlin: Springer-Verlag.
Kelejian, H.H. and D.P. Robinson (1997). Infrastructure productivity estimation and its underly-
ing econometric specifications: a sensitivity analysis. Papers in Regional Science 76, 115–131.
Kelejian, H.H. and D.P. Robinson (1998). A suggested test for spatial autocorrelation and/or het-
eroskedasticity and corresponding Monte Carlo results. Regional Science and Urban Economics
28, 389–417.
Kelejian, H.H. and D.P. Robinson (1999). The influence of spatially correlated heteroskedasticity
on tests for spatial correlation. In L. Anselin and R. Florax (Eds.), Advances in Spatial Econo-
metrics. Heidelberg: Springer-Verlag (forthcoming).
King, M. (1981). A small sample property of the Cliff-Ord test for spatial correlation. Journal of the
Royal Statistical Society B 43, 263–264.
Koerts, J. and A.P.I. Abrahamse (1968). On the power of the BLUS procedure. Journal of the Amer-
ican Statistical Association 63, 1227–1236.
Krugman, Paul (1991a). Increasing returns and economic geography. Journal of Political Economy
99, 438–99.
Krugman, Paul (1991b). Geography and trade. Cambridge, MA: MIT Press.
Krugman, Paul (1998). Space: the final frontier. Journal of Economic Perspectives 12, 161–174.
Lahiri, S.N. (1996). On the inconsistency of estimators under infill asymptotics for spatial data.
Sankhya A 58, 403–417.
Land, K. and G. Deane (1992). On the large-sample estimation of regression models with spatial-
or network-effects terms: a two stage least squares approach. In P. Marsden (Ed.), Sociological
Methodology, pp. 221–248. San Francisco: Jossey-Bass.
LeSage, James P. (1997). Bayesian estimation of spatial autoregressive models. International
Regional Science Review 20, 113–129.
Magnus, J. (1978). Maximum likelihood estimation of the GLS model with unknown parameters
inthe disturbance covariance matrix. Journal of Econometrics 7, 281–312. (Corrigenda, Journal of
Econometrics 10, 261).
Manski, C. F. (1993). Identification of endogenous social effects: the reflection problem. Review of
Economic Studies 60, 531–542.
Mardia, K.V. and R.J. Marshall (1984). Maximum likelihood estimation of models for residual
covariance in spatial regression. Biometrika 71, 135–146.
Mardia, K.V. and A.J. Watkins (1989). On multimodality of the likelihood for the spatial linear
model. Biometrika 76, 289–295.
Martin, R. J. (1993). Approximations to the determinant term in Gaussian maximum likelihood esti-
mation of some spatial models. Communications in Statistics: Theory and Methods 22, 189–205.
MathSoft (1996). S+SpatialStats User’s Manual for Windows and Unix. Seattle, WA: MathSoft, Inc.
McMillen, D. P. (1992). Probit with spatial autocorrelation. Journal of Regional Science 32, 335–48.
McMillen, D. P. (1995). Spatial effects in probit models: a Monte Carlo investigation. In L. Anselin and

— 28 —
 R. Florax (Eds.), New Directions in Spatial Econometrics, pp. 189–228. Heidelberg: Springer-Verlag.
Moran, P. A. P. (1948). The interpretation of statistical maps. Biometrika 35, 255–260.
Moran, P. A. P. (1950a). Notes on continuous stochastic phenomena. Biometrika 37, 17–23.
Moran, P. A. P. (1950b). A test for the serial dependence of residuals. Biometrika 37, 178–181.
Murdoch, J.C., M. Rahmatian and M. A. Thayer (1993). A spatially autoregressive median voter
model of recreation expenditures. Public Finance Quarterly 21, 334–350.
Murdoch, James C., Todd Sandler and Keith Sargent (1997). A tale of two collectives: sulphur
versus nitrogen oxides emission reduction in Europe. Economica 64, 281–301.
Nelson, Gerald C. and David A. Gray (1997). The use of spatial analysis at the World Bank. Environ-
mental Economics Series Paper No. 56. The World Bank, Washington DC.
Nelson, Gerald C. and Daniel Hellerstein (1997). Do roads cause deforestation? Using satellite
images in econometric analysis of land use. American Journal of Agricultural Economics 79, 80–88.
Newey, Whitney K. and Kenneth D. West (1987). A simple, positive semi-definite, heteroskedas-
ticity and autocorrelation consistent covariance matrix. Econometrica 55, 703–708.
Olmo, J. C. (1995). Spatial estimation of housing prices and locational rents. Urban Studies 32,
1331–1344.
Ord, J.K. (1975). Estimation methods for models of spatial interaction. Journal of the American Sta-
tistical Association 70, 120–126.
Pace, R. K. (1997). Performing large spatial regressions and autoregressions. Economics Letters 54,
283–291.
Pace, R. K. and R. Barry (1997a). Sparse spatial autoregressions. Statistics and Probability Letters 33,
291–297.
Pace, R. K. and R. Barry (1997b). Quick computation of spatial autoregressive estimators. Geo-
graphical Analysis 29, 232–246.
Pace, R. K. and R. Barry (1998). Spatial Statistics Toolbox 1.0. Real Estate Research Institute, Lousi-
ana State University, Baton Rouge, LA.
Pace, R. Kelley, Ronald Barry, C.F. Sirmans (1998). Spatial statistics and real estate. Journal of Real
Estate Finance and Economics 17, 5–13.
Paelinck, J. and L. Klaassen (1979). Spatial Econometrics. Saxon House, Farnborough.
Pinkse, Joris (1998). Asymptotic properties of the Moran and related tests and a test for spatial
correlation in probit models. Working Paper, Department of Economics, University of British
Columbia, Vancouver, BC.
Pinkse, Joris (1999). Moran-flavored tests with nuisance parameters. In L. Anselin and R. Florax
(Eds.), Advances in Spatial Econometrics. Heidelberg: Springer-Verlag (forthcoming).
Pinkse, Joris and Margaret E. Slade (1998). Contracting in space: an application of spatial statis-
tics to discrete-choice models. Journal of Econometrics 85, 125–54.
Poirier, Dale J. and Paul A. Ruud (1988). Probit with dependent observations. Review of Economic
Studies 55, 593–614.
Pötscher, B.M. and I. R. Prucha (1997). Dynamic Nonlinear Econometric Models. Springer, Berlin.
Rao, C. R. (1973). Linear Statistical Inference and its Applications (2nd Ed.). Wiley, New York.
Ripley, B. D. (1981). Spatial Statistics. Wiley, New York.
Ripley, B. D. (1988). Statistical Inference for Spatial Processes. Cambridge: Cambridge University Press.
Schmidt, P. (1976). Econometrics. New York: Marcel Dekker.
Smirnov, O. and L. Anselin (1999). Fast maximum likelihood estimation of very large spatial
autoregressive models: a characteristic polynomial approach. Working Paper, Bruton Center,
University of Texas at Dallas, Richardson, TX.
Sneek, J.M. and P. Rietveld (1999). On the estimation of the spatial moving average model. In L.
Anselin and R. Florax (Eds.), Advances in Spatial Econometrics. Heidelberg: Springer-Verlag
(forthcoming).
Tiefelsdorf, M. and B. Boots (1995). The exact distribution of Moran’s I. Environment and Planning

— 29 —
 A 27, 985–999.
Topa, Giorgio (1996). Social interactions, local spillovers and unemployment. Ph.D. Dissertation,
Department of Economics, University of Chicago.
Upton, G. J. and B. Fingleton (1985). Spatial Data Analysis by Example. Volume 1: Point Pattern and
Quantitative Data. Wiley, New York.
Vijverberg, Wim (1997). Monte Carlo evaluation of multivariate normal probabilities. Journal of
Econometrics 76, 281–307.
Waller, Lance, Bradley Carlin and Hong Xia (1997). Structuring correlation within hierarchical
spatio-temporal models for disease rates. In T. Gregoire, D. Brillinger, P. Diggle, E. Russek-
Cohen, W. Warren and R. Wolfinger (Eds.), Modelling Longitudinal and Spatially Correlated
Data, pp. 309–319. New York: Springer-Verlag.
Waller, Lance, Bradley Carlin, Hong Xia and Alan Gelfand (1997). Hierarchical spatio-temporal
mapping of disease rates. Journal of the American Statistical Association 92, 607–617.
Warnes, J.J. and B.D. Ripley (1987). Problems with likelihood estimation of covariance functions
of spatial Gaussian processes. Biometrika 74, 640–642.
Whittle, P. (1954). On stationary processes in the plane. Biometrika 41, 434–449.

— 30 —