See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/350327936

Spatial Models in Econometric Research

Preprint · March 2021

DOI: 10.13140/RG.2.2.26447.20641

CITATIONS READS
2 4,947

1 author:

Luc Anselin
University of Chicago
290 PUBLICATIONS 61,022 CITATIONS

SEE PROFILE

All content following this page was uploaded by Luc Anselin on 23 March 2021.

The user has requested enhancement of the downloaded file.

 Spatial Models in Econometric Research∗

Luc Anselin

Center for Spatial Data Science

The University of Chicago, Chicago, IL 60637

anselin@uchicago.edu

January 4, 2021

∗
 Summary

Spatial models are a recent but growing addition to econometric

research. They are characterized by an attention to the location of ob-

servations (i.e., ordered spatial locations) and the interaction among

them. Specifically, spatial models formally express spatial interaction

by including variables observed at other locations into the regression

specification. This can take different forms, mostly based on an aver-

aging of values at neighboring locations, through a so-called spatially

lagged variable, or spatial lag. The spatial lag can be applied to the de-

pendent variable, to explanatory variables and/or to the error terms.

This yields a range of specifications for cross-sectional dependence, as

well as for static and dynamic spatial panels.

A critical element in the spatially lagged variable is the definition

of neighbor relations in a so-called spatial weights matrix. Historically,

the spatial weights matrix has been taken to be given and exogenous,

but recent research has focused on estimating the weights from the

data and on accounting for potential endogeneity in the weights.

Due to the uneven spacing of observations, and the complex grow-

ing of data to obtain asymptotics, specialized laws of large numbers

and central limit theorems need to be developed, since results from

time series analysis are no longer applicable. This has yielded an

active body of research into the asymptotics of spatial models.

Key words: spatial econometrics, spatial dependence, spatial in-

teraction, spatial weights, asymptotics

1 Introduction

Spatial models are a fairly recent addition to the toolbox of mainstream

econometric research. Originally developed primarily in regional science and

quantitative geography, spatial econometrics has become a growing compo-

nent of theoretical and applied research in econometrics (for a historical

overview of the development of the field, see Anselin 2010).

Spatial econometrics can be defined as a “subset of econometric methods

that is concerned with the spatial aspects present in cross-sectional and space-

time observations. Variables related to location, distance and arrangement

(topology) are treated explicitly in model specification, estimation, diagnostic

checking and prediction” (Anselin 2006). The spatial aspects are referred to

as spatial effects, consisting of spatial dependence and spatial heterogeneity,

as defined in Anselin (1988).

Spatial dependence is a special form of cross-sectional dependence, in

that the structure of the correlation between pairs of observations is derived

from a spatial ordering, determined by the relative position (distance, spa-

tial arrangement) of the observations in geographic space, or, in general, in

any network space (with observations as nodes connected by edges). While

sometimes seen as a straightforward extension of time series dependence into

two dimensions, this is incorrect. A special set of methods is required to deal

with the distinct nature of spatial dependence, specifically, with the implied

feedback effects (Anselin 1988).

1
 Spatial heterogeneity is a special case of structural instability, a familiar

problem in standard econometrics. The heterogeneity is structured spatially,

i.e., with respect to locations in space, or regional subsets of observations.

This structure in turn informs the specification of heterogeneity, spatially

varying coefficients, random coefficients and spatial regimes (a form of spatial

fixed effects) (Anselin 1988).

Whereas early on, apart from Anselin (1980, 1988), Cliff and Ord (1981),

and later LeSage and Pace (2009), there was a relative dearth of treatments

that provided a comprehensive review of methods and models, this is no

longer the case. In recent years, several new texts were published, providing

ample access to the breadth of the field, in terms of theoretical results, new

methods and a range of applications. Recent examples include Anselin and

Rey (2014), Arbia (2014), Dubé and Legros (2014), Elhorst (2014a), Kelejian

and Piras (2017), and Chi and Zhu (2020).

In addition, there are several extensive reviews of the state of the art,

both for a cross-sectional setting as well as for spatial panels. Examples

include Anselin and Bera (1998), Anselin (2001, 2006, 2021), Anselin et al.

(2008), Lee and Yu (2010, 2011, 2015), Elhorst (2012, 2014b), and Bai et al.

(2016).

Due to space constraints, it is impossible to do justice to the breadth

of topics covered by spatial econometric research. Specifically, a detailed

review of technical issues associated with specialized estimation methods and

specification tests is outside the scope (see the review articles cited above for

2
details). Instead, in this essay, the focus is on three important aspects that

distinguish the spatial models from standard econometric approaches. First,

the specification of spatial dependence is considered and the main models

in use briefly reviewed. Next, attention is paid to two particular features of

spatial models, the specification of the structure of interaction through the

device of spatial weights, and the specialized asymptotics required to obtain

desirable properties for estimators and specification tests. The focus is solely

on spatial dependence, with spatial heterogeneity outside the current scope.

The essay closes with some concluding remarks.

2 The Specification of Spatial Interaction

Spatial interaction is incorporated into a regression specification by including

observations of variables at other locations on the right hand side of the model

equation for the dependent variable y at a given location i, yi . Typically, this

is achieved by means of a so-called spatially lagged variable, or an average

of the values observed at neighboring locations (Anselin 1980). Formally,

P
a spatially lagged variables is constructed as j wij zj , where wij are so-

called spatial weights, expressing the presence of a neighbor relation between

locations i and j. In matrix notation, the spatial weights are contained in

a n × n spatial weights matrix, W, with n as the number of cross-sectional

(spatial) observations. By convention, the diagonal elements of the weights

matrix are zero and the rows are standardized such that their sum equals

3
one (a detailed discussion of spatial weights is given below). The spatially

lagged variable then takes on the form Wz, where z is a n × 1 vector of

observations.

Spatially lagged variables can be considered for the dependent variable,

Wy, for the explanatory variables, WX, or for the error terms, We. The

combination of these terms with additional exogenous and endogenous vari-

ables as well as random effects gives rise to a wide range of model specifica-

tions.

Three broad classes of models are reviewed in what follows, i.e., the basic

cross-sectional model, static spatial panels, and dynamic spatial panels. The

section closes with a brief overview of some recent extensions.

2.1 The Cross-Sectional Spatial Model

In the recent literature, the point of departure of a spatial specification search

is a so-called general nesting spatial model, which encompasses several tradi-

tional models as special cases. As outlined by Elhorst (2014a), such a model

combines a spatial Durbin model with a spatial autoregressive specification

for the error term. Specifically, using matrix notation, the unconstrained

4
spatial Durbin model is expressed as (LeSage and Pace 2009, Elhorst 2010):1

y = αι + ρWy + Xβ + WXγ + e,

where W is the spatial weights matrix, y a n × 1 vector of cross-sectional

observations on the dependent variable, X is a matrix of observations on

(exogenous) explanatory variables, α is the coefficient of the constant term,

ρ is the spatial autoregressive coefficient, and β and γ are k × 1 vectors

of regression coefficients (with k not including the constant term). In this

specification, the two sets of spatial variables are the spatially lagged depen-

dent variable, Wy, and the spatially lagged explanatory variables, WX (not

including the constant term).

Endogenous variables may be included as well, but are ignored here in

order not to complicate the notation (the main effect of including additional

endogenous variables is increased complexity of the estimation methods).

In addition, the error term is specified to follow a spatial autoregressive

process:

e = λWe + u,

with λ as the spatial autoregressive parameter and u as a vector of idiosyn-

1
Initially, this specification was used to assess the so-called common factor hypothesis,
and referred to as a spatial Durbin model in analogy to its time series counterpart, as a
spatial lag specification (with autoregressive parameter λ) equivalent to a spatial autore-
gressive error model (Burridge 1981, Anselin 1988). The common factor hypothesis boils
down to a nonlinear constraint on the coefficients, such that, for each explanatory variable,
the coefficients of WX equal the negative of the product of λ and the corresponding β.

5
cratic error terms, which could be heteroskedastic.

The main attractiveness of this encompassing approach is that simpler

models are obtained by imposing zero constraints on specific parameters,

although upon closer examination, this is not always that straightforward.

For example, by setting γ = 0 and λ = 0, the so-called spatial lag model is

obtained (Ord 1975, Anselin 1988), arguably the most studied spatial model:2

y = ρWy + Xβ + e.

A spatial lag model is often viewed as expressing the effect of the neigh-

bors on the right hand side of the equation through the spatial lag term Wy.

However, this must be interpreted with caution. In a conditional perspective,

the values for those neighbors are observed, although this is not realistic in

a true cross-sectional setting. In fact, in a simultaneous view, the complete

pattern is determined jointly, since the values for the neighbors in turn de-

pend on the value at i and cannot occur independently. This is clearly seen

from the reduced form for the spatial lag model, obtained by moving the

spatial lag term to the left hand side and “solving” for y by means of the

matrix inverse (I − ρW)−1 . In full:

y = (I − ρW)−1 Xβ + (I − ρW)−1 e

Under very general conditions that are typically satisfied in practice, a

2
An early formulation of the spatial autoregressive model on a regular lattice was given
by Whittle (1954).

6
power expansion can be applied to the inverse term in the reduced form:3

(I − ρW)−1 = I + ρW + ρ2 W2 + . . . ,

This gives the conditional expectation of the dependent variable as:

E[y|X] = Xβ + ρWXβ + ρ2 W2 Xβ + ρ3 W3 Xβ + . . .

This expression clearly demonstrates that the value for y at a location i

is in fact not determined by the values for y at the neighboring locations,

but instead by a series of diminishing effects (through the powers of the

autoregressive parameter) of the observations for the exogenous variables at

increasing orders of neighbors (powers of the spatial weights matrix).

The reduced form also leads to the notion of a spatial multiplier, and the

interpretation of direct and indirect effects. The latter show the effect of a

change in an explanatory variable at any location on the dependent variable

at any location. Hence, this involves a n × n matrix of effects, relating each

location of y to every location of each x. As a result, the matrix of effects

includes not only ∂yi /∂xh,i (for a given explanatory variable xh at i), but

also ∂yj /∂xh,i , as well as ∂yi /∂xh,j . These complex effects follow from the

reduced form and how the locations are interconnected through the elements

of the inverse matrix (I − ρW)−1 (see, e.g., LeSage and Pace 2009, for a

detailed discussion).

With ρ = 0 and γ = 0, the spatial error model results (Anselin 1988),

with a standard non-spatial regression for the substantive model. Using the
3
See Anselin and Bera (1998) for an early example of this principle.

7
result of the reduced form for the error term yields:

y = Xβ + (I − λW)−1 u.

This results in a particular structure for the error variance-covariance matrix:

E[ee0 ] = Σ = (I − λW)−1 E[uu0 ](I − λW0 )−1 .

When the number of neighbors contained in the spatial weights matrix is

not constant across observations (a typical case in practice), the error term

e becomes heteroskedastic, a complication unique to the spatial context.

Other forms of spatial dependence in the error term include a direct para-

metric specification of the error covariance (see, e.g., Dubin et al. 1999, for a

review), or a non-parametric approach in the form of a heteroskedastic and

autocorrelation consistent covariance, or HAC (Kelejian and Prucha 2007).

Alternative spatial process models include a spatial moving average speci-

fication (Haining 1978, Fingleton 2008, Baltagi and Liu 2011), and factor

models (further reviewed in the discussion of spatial panels).

Note that the constraints on the general specification of ρ = 0 and γ = 0,

which would suggest a spatial error model, may contradict such a speci-

fication. In fact, in the original formulation of the spatial Durbin model

as a lag specification for the spatial autoregressive error model (Burridge

1981, Anselin 1988), ρ = λ 6= 0. Consequently, these constraints must be

interpreted with caution as the models are not nested in the usual sense.

Similarly, imposing λ = 0 in the general model may suggest a spatial Durbin

8
specification, but with the constraint γ = −ρβ in place, the spatial Durbin

model in fact corresponds with a spatial autoregressive error specification,

contradicting λ = 0.

A final special model results for ρ = 0 and λ = 0 as the so-called

SLX model (for spatially lagged X), originally suggested as a spatial cross-

regressive specification by Florax and Folmer (1992). This specification

has gained greater prominence since its reintroduction by Vega and Elhorst

(2015):

y = Xβ + WXγ + e.

The SLX model differs from the reduced form of the spatial lag not only

in that higher order (global) effects are truncated, but also by the lack of

restrictions on the parameter vector γ. This allows for greater flexibility

in the specification of distance decay. In addition to using a spatial lag,

more complex ways can be included to bring in the effect of “variables at

other locations,” such as a potential measure (Anselin 2002) or nonlinear

specifications (Vega and Elhorst 2015).

The SLX model has received less attention in the theoretical literature

since it does not require specialized estimation methods. In the absence of

other complicating factors (e.g., endogenous variables), ordinary least squares

remains an unbiased estimator for γ.

9
2.2 Static Spatial Panels

When observations are available across space as well as over time, a rich

array of spatial panel specification is possible, incorporating various forms

of spatial and space-time correlation (for extensive reviews, see, e.g. Anselin

et al. 2008, Elhorst 2014a, Lee and Yu 2010, 2015). In static spatial panel

models, the focus is on cross-sectional dependence, and temporal (serial)

dependence is precluded or relegated to the error terms.

The basic cross-sectional specifications can readily be extended to the

panel setting by considering a separate cross-section for each period of time.4

In contrast to the non-spatial panel literature, the observations are thus

ordered as cross-sections for each time period, and not as (short) time series

for each cross-sectional observation.

To illustrate this point, consider a spatial lag model expressed as a vector

of n observations for time period t:

yt = ρWyt + Xβ + et ,

or the standard expression for a cross-sectional spatial lag model, with a

subscript t to identify the time period. Stacked over all time periods, the

model becomes:

y = ρ(IT ⊗ Wn )y + Xβ + e,

where y and e are vectors of dimension nT × 1, X is a nT × k matrix, and

4
In most of the literature, the cross-sectional units are the same over time, resulting in
so-called balanced panels.

10
the subscripts on I and W indicate their dimensions, respectively T × T

for the time dimension and n × n for the cross-sectional dimension. For

the sake of simplicity, the spatial autoregressive coefficient ρ and the slope

coefficients β are constant over time (the extension to ρt and βt , or even βi is

straightforward). This model (and its counterpart spatial Durbin or spatial

error specifications) is a direct extension of the cross-sectional case.

A form of error dependence can be introduced as a spatial seemingly un-

related regression (SUR), in which the different cross-sections are connected

through an unspecified error covariance that contains the serial (time) cor-

relation (see Anselin 1988, Anselin et al. 2008, Baltagi and Pirotte 2011)

A more interesting set of specifications results when, in addition to the

spatial lag term, cross-sectional and/or time-specific fixed or random effects

are introduced. With a n × 1 vector of cross-sectional effects µ and a T × 1

vector of time specific effect η, the stacked spatial lag model becomes:

y = ρ(IT ⊗ Wn )y + Xβ + (ιT ⊗ In )µ + (IT ⊗ ιn )η + e.

Due to the stacking by cross-section, this form differs from the standard

textbook expression.

As is well-known, a time random effect, since it is equally shared by all

cross-sectional units, induces a peculiar form of spatial correlation, with equal

strength among all pairs of observations. A more explicit specification for

spatial correlation is through a spatial autoregressive error process. With

a random cross-sectional effect, there are two avenues to introduce spatial

11
error correlation.

A first approach was outlined in Anselin (1988) (see also Baltagi et al.

2007), with the spatial autoregressive process applied to the error term in

each cross-section, as

et = λWn et + ut ,

with λ as the spatial autoregressive coefficient (assumed to be constant across

time periods for simplicity, but this is easily relaxed), and ut as a n×1 vector

of idiosyncratic errors (possibly allowed to be heteroskedastic). The complete

error term also includes a n × 1 vector of spatial random effects µ, such that

for each time period t,

vt = (In − λWn )−1 ut + µ.

A second specification was suggested in Kapoor et al. (2007), with the

spatial autoregressive process applied to both the spatial random effects and

the idiosyncratic error term:

vt = (In − λWn )−1 (µ + ut )

This results in a simpler expression for the overall variance-covariance matrix,

which facilitates estimation.

It should be noted that the two models imply different spatial spillover ef-

fects. In the first model, these spillovers are inherently time-varying, whereas

in the second model they are time invariant.

Baltagi et al. (2013) introduce an encompassing model that includes both

12
specifications as special cases. They refer to the first model as the Anselin

model and the second one as the KKP model.

The Baltagi et al. (2013) specification includes two error components for

a cross-section at time t, for ease of exposition, illustrated for an otherwise

non-spatial model:

yt = Xt β + u1 + u2t ,

with the first component as time-invariant, whereas the second component

is indexed by time period.

A spatial autoregressive process is introduced for each error component.

For the first component, this is time-invariant, which gives the fully stacked

specification as:

(IT ⊗ In )u1 = (IT ⊗ In )(λ1 Wu + µ),

yielding a vector of length nT × 1, with µ as the vector of spatial random

effects.

The second component yields a stacked spatial autoregressive process:

u2 = λ2 (IT ⊗ W)u2 + ν,

where ν is a nT × 1 vector of idiosyncratic errors.

Combining the two (uncorrelated) random components, with respective

variances σµ2 and σν2 , and using A = In − λ1 W and B = In − λ2 W, yields

the overall error variance-covariance matrix as:

Σ = σµ2 [ιT ι0T ⊗ (A0 A)−1 ] + σν2 [IT ⊗ (B0 B)−1 ],

13
with ιT as a T × 1 vector of ones. As Baltagi et al. (2013) show, setting

λ1 = 0 yields the Anselin model, whereas setting λ1 = λ2 results in the KPP

specification. With λ1 = λ2 = 0, the standard non-spatial panel variance

expression is found.

In addition to spatial autoregressive specifications, spatial moving average

and HAC models are other ways to incorporate spatial correlation in the panel

error terms. These error terms can be combined with spatial lag or spatial

Durbin specifications, potentially with heterogeneous coefficients across time

periods, to yield a wide range of spatial panel models.

An alternative view on cross-sectional correlation in panel models is of-

fered by the multifactor perspective, developed by Pesaran (2006) (see also

Pesaran and Tosetti 2011, Kapetanios et al. 2011, Chudik and Pesaran 2015,

among others). In this approach, the error term is considered to consist of

unobserved common factors with each cross-sectional unit having a distinct

factor loading. This leads to cross-sectional correlation. Depending on the

structure of the factors, such cross-sectional correlation can be classified as

strong or weak (Chudik et al. 2011), and the strength of cross-sectional cor-

relation can be quantified by means of a so-called exponent of cross-sectional

dependence (Bailey et al. 2016). However, while such models deal with cross-

sectional correlation, they are essentially aspatial in that they ignore the or-

dering of spatial observations, or a so-called structural interactions assump-

tion (Bhattacharjee and Holly 2011). In addition, multifactor models rely on

large T for estimation and inference.

14
2.3 Dynamic Spatial Panels

An additional degree of complexity is added to the spatial panels when in-

teraction over time is introduced, in so-called dynamic panels (Elhorst 2012,

2014a). A general specification, as outlined in Elhorst (2014a, p. 96), consists

not only of a contemporaneous spatially lagged dependent variable, Wyt ,

and/or spatially lagged explanatory variables, WXt , but also time-lagged

versions of these variables. This can include a time-lagged dependent vari-

able, yt−1 , and/or explanatory variables, Xt−1 , but also time-lagged versions

of the spatial lags, such as Wyt−1 and WXt−1 . In addition, several time lags

can be incorporated in the error component structure as well.

Clearly, not all parameters in such a model can be identified, but many

simpler specifications can be found by selecting subsets of these spatially and

temporally lagged variables (see also Anselin et al. 2008, for a discussion of

identification concerns).

The main issues associated with dynamic panel models are determining

appropriate estimation methods (maximum likelihood or GMM), and the

interpretation of the associated space-time dynamics (for technical overviews,

see Lee and Yu 2010, 2011, 2015, among others)

2.4 Model Extensions

The basic cross-sectional model has been extended into a number of dif-

ferent directions, introducing new specifications for spatial dependence, and

15
including spatial interactions to a range of models beyond the standard linear

regression.

For example, LeSage and Pace (2007) proposed a matrix exponential

specification, replacing the geometric decay of influence of the original spatial

lag specification with an exponential pattern of decay.

Spatial lag and error models have been applied to discrete dependent

variable specifications, such as probit and tobit (e.g., Pinkse and Slade 1998,

Fleming 2004, Smirnov 2010, Calabrese and Elkink 2014, Xu and Lee 2015,

Baltagi et al. 2016, Xu and Lee 2018), as well as count models (Lambert

et al. 2010, Bhat et al. 2014, Liesenfeld et al. 2017).

A slightly different type of dependent variable is when the model per-

tains to flows between origins and destinations. While still fundamentally

cross-sectional in nature, such models encounter complications in terms of

the specification of the weights and the interpretation of direct and indirect

effects (a recent overview of the various methodological issues is contained

in the papers in the edited volume by Patuelli and Arbia 2016).

A final set of extensions pertains to the inclusion of spatial terms into

models other than the classic linear regression specification. Examples, in-

clude spatial quantile regressions (e.g., Kostov 2013, McMillen 2013), and

spatial stochastic frontier models (e.g., Fusco and Vidoli 2013, Orea and

Àlvarez 2019).

16
3 Spatial Weights

As mentioned, the way in which observations for variables at other locations

are incorporated into the regression specification is through the device of a

spatially lagged variable, i.e., the combination of a spatial weights matrix

with a vector of observations.

The specification of the spatial weights has been the subject of consid-

erable research and debate in spatial econometrics. Most of the early lit-

erature considered the weights as given and exogenous, but more recently,

data-driven estimated weights as well as weights that are determined endoge-

nously have received interest. In addition, the extent to which the weights

are stable over time has been an important concern. Each of these topics is

considered in turn.

3.1 Exogenous weights

Exogenous weights can be based on a number of different criteria to determine

with pairs of observations are neighbors, i.e., for which pairs i − j, wij 6= 0.

A range of measures have been employed to define such neighbor relation,

including contiguity (sharing a common geographical border), distance cri-

teria (distance bands or k-nearest neighbors), social network connectivity

and even economic distance. Reviews of the different options, operational

concerns and interpretation can be found in Bavaud (1998), Getis (2009),

Stakhovych and Bijmolt (2009), Harris et al. (2011), and Anselin and Rey

17
(2014, Chapters 3 and 4), among others.

Most exogenous spatial weights used in practice have a limit on the

number of neighbors (typically a small value), and use the matrix in row-

standardized form, such that the elements of each row sums to one. As a

result, the sum of all the elements in the weights matrix (an important ele-

ment in several test statistics) equals the number of observations. However,

the resulting weights matrix is not symmetric.

Row-standardization results in a number of desirable properties, such as

boundedness of the parameter space, and easy interpretation of the notion

of a spatially lagged variable (as an average of neighboring observations) and

the associated spatial parameters. However, this approach is not without

critics, and alternatives have been proposed that scale the full matrix and

retain its symmetry (e.g., Kelejian and Prucha 2010, Plümper and Neumayer

2010, Kelejian and Piras 2017).

A rarely used form of spatial weights matrix imposes equality of the

weights, i.e., wij = 1 for all i 6= j. While this approach facilitates some esti-

mation issues, it is otherwise problematic (for further discussion, see Kelejian

and Prucha 2002, Lee 2002, Martellosio 2011). So-called block weights have

equal weights in subsets of the data, resulting in a block-diagonal structure

with all observations in the same subset (e.g., counties in a state) consid-

ered to be neighbors (Case 1991, 1992, Kelejian et al. 2006, Anselin and

Arribas-Bel 2013).

One common source of confusion is between the structure of the weights

18
and the corresponding variance-covariance matrix. As shown in the reduced

form for the spatial autoregressive process, the range and strength of interac-

tion is driven by multiple orders of neighbors in combination with powers of

the autoregressive coefficient. In other words, the weights matrix is not the

covariance matrix. The actual structure of the covariance is quite complex

and sometimes seen as counterintuitive (Wall 2004, Rodrigues et al. 2014,

Strauß et al. 2017). A proper interpretation should be based on the induced

graph structure (Assunção and Krainski 2009, Martellosio 2012).

Opinions about importance of weights structure are divided, with most of

the literature arguing they are critical to the model specification (e.g., Harris

et al. 2011), with others stressing the interpretation of direct and indirect

effects, in which the actual weights specification plays a less important role

(LeSage and Pace 2014).

In practice, there are often several candidate weights, without clear the-

oretical guidance as to which is most appropriate. Selection can be based on

model fit criteria, such as AIC (Anselin 1988, Stakhovych and Bijmolt 2009),

or an extension of Mallow’s Cp to spatial autoregressive models (Zhang and

Yu 2018). Alternatively, pairwise comparison of candidate weights can be

carried out by means of spatial J-tests on non-nested hypotheses, an active

area of recent research (Anselin 1986, Kelejian 2008, Kelejian and Piras 2011,

Piras and Lozano-Gracia 2012, Han and Lee 2013, Jin and Lee 2013, Wang

et al. 2013, Delgado and Robinson 2015, Debarsy and Ertur 2019). Spatial

J tests have also been suggested to select the optimal number of neighbors

19
for k-nearest neighbor weights (Gerkman and Ahlgren 2014).

An alternative to selecting a single weights matrix is to combine different

weights, commonly implemented as an application of Bayesian model averag-

ing (examples are LeSage and Parent 2007, LeSage and Pace 2009, Cotteleer

et al. 2011, Piribauer and Fischer 2015, Debarsy and LeSage 2018, among

others). Yet a different approach considers stochastic, but exogenous spatial

weights in higher order spatial autoregressive models (Gupta 2019).

Finally, modern techniques of penalized regression and model boosting

have been applied to select the best (combination) from among a set of

candidate weights (e.g., Kostov 2010, Seya et al. 2013).

3.2 Estimated weights

An alternative to specifying the spatial interaction by scaling the weights

matrix with a single autoregressive parameter, e.g., ρW, is to embed a pa-

rameter in the weights themselves. Specifically, wij = f (θ, dij ), where f is

a distance decay function (i.e., decreasing with increasing distance) in some

distance metric dij between observations i and j, and θ consists of one or

more parameters. Typical functional forms are wij = e−θdij or wij = d−θdij

(see, e.g., Bodson and Peeters 1975, Anselin 1980, Benjanuvatra and Bur-

ridge 2015). Care must be taken that the parameter space does not violate

the regularity conditions required to obtain proper statistical inference.5

5
In addition, there may be identification problems if a scaling coefficient is applied to
the parameterized weights matrix.

20
 A second strand of literature focuses on estimating the elements of the

weights matrix from the covariance structure in the data. This is typically

couched in a panel data setting, with T  n, where the time dimension

provides a way to estimate a covariance matrix. Subject to additional sym-

metry or sparsity restrictions, the potentially n(n − 1) elements of W can be

estimated from the covariance matrix. Originally suggested by Meen (1996),

this idea was further elaborated in a number of papers by Bhattacharjee and

Holly (2011), Bhattacharjee et al. (2012), Beenstock and Felsenstein (2012),

and Bhattacharjee and Jensen-Butler (2013).6

The key aspect of these approaches is the equality between the residual

covariance matrix and an expression in the spatial weights and potentially

heteroskedastic (across locations) variances of the idiosyncratic error term,

the n × n matrix:

E[uu0 ] = (I − W)−1 Σ(I − W0 )−1 ,

which follows from the definition of the error SAR process.7 Together with

some additional constraints, this equality leads to a nonlinear optimization

procedure or a set of moment equations that allow for the elements of W to

be extracted.8
6
A cross-sectional application is given in Bhattacharjee et al. (2012), where the covari-
ance between housing submarkets is exploited to provide identification of the weights.
7
There is no separate spatial autoregressive coefficient, since it would not be jointly
identified with the elements of the weights matrix.
8
An alternative based on entropy maximization is given in Fernández-Vázquez et al.
(2009).

21
 In practice, this approach has only been implemented for rather small n.

For example, in empirical applications, n ranges from 4 and 7 (Bhattachar-

jee et al. 2012), to 5 (Bhattacharjee and Holly 2011) and 9 (Meen 1996,

Beenstock and Felsenstein 2012, Bhattacharjee and Jensen-Butler 2013).

Penalized estimation, based on versions of the LASSO (Least Absolute

Shrinkage and Selection Operator, Hastie et al. 2009) has recently been ap-

plied to this context as well. In Ahrens and Bhattacharjee (2015), and Lam

and Souza (2016, 2020), a two-step LASSO estimation is applied to both the

elements of the weights matrix and the model parameters by exploiting spar-

sity, symmetry and other constraints to provide identification. For example,

the sparsity condition must be such that the number of non-zero elements in

a row of the weights matrix, together with the number of exogenous variables

is much less than the time dimension.

Alternatively, one can avoid to specificy the spatial weights by taking

a semi-parametric approach, first suggested by Pinkse et al. (2002). For

example, the spatial lag model can be specified as:

y = α + W(d)y + Xβ + e,

where the weights matrix is a general unspecified function of a distance met-

ric, wij = g(dij ). The function g is approximated by an expression in basis

functions of the pairwise distances, such as a Fourier or polynomial basis

function. Estimation is carried out by means of instrumental variables, with

P
instruments replacing yj in the spatial lag term j g(dij )yj . This general

22
idea is extended to a functional coefficient spatial autoregressive model with

nonparametric spatial weights in Sun (2016) and Koroglu and Sun (2016),

allowing for both exogenous and endogenous variables to drive the spatial

interaction (spatial lag term).

3.3 Endogenous weights

A different specification issue related to the spatial weights pertains to their

potential endogeneity, i.e., when the (known) elements of the weights matrix

cannot be assumed to be exogenous, but are jointly determined with the

dependent variable in the model (and/or error terms). Examples of such

endogenous weights are the use of economic distance, trade shares, relative

prices or transportation costs to construct the elements of W. Even though

the weights are known (and thus not estimated), their endogeneity should be

accounted for in the estimation process.

An initial approach, suggested by Kelejian and Piras (2014) (see also

Kelejian and Piras 2017, Chapter 13) consists of constructing additional in-

struments for the weights matrix elements. An estimate of the weights ma-

trix, Ŵ replaces the non-zero weights by a predicted value from an auxiliary

regression of the weights on a number of polynomials in instrumental vari-

ables. The estimation of the spatial lag specification utilizes the predicted

weights to construct the usual spatially lagged exogenous variables, e.g., Ŵ X

in an instrumental variables estimation.

A general framework is outlined in Qu and Lee (2015). The endogene-

23
ity of the spatial weights matrix is formalized by expressing the weights

as a function of a set of endogenous variables, say z1 , z2 , . . . , zp , such that

wij = f (zi , zj ). These variables are themselves a linear function of a set of

q exogenous variables x2h for h = 1, 2, . . . , q and an error term. In matrix

notation:

Z = X2 Γ + E,

with Z as the n × p matrix with observations on the variables that define

the spatial weights, X2 as the n × q matrix of observations on the auxiliary

(exogenous) variables, with associated q × p coefficient matrix Γ. The n × p

matrix E consists of error vectors, each corresponding with a column of Z.

The spatial lag model takes on the customary form:

y = ρW(Z)y + X1 β + u,

where the notation W(Z) indicates that the spatial weights are constructed

from the variables z.

The source of the endogeneity comes from the joint distribution of the

error terms in E and u, which are zero mean, but have a covariance matrix:
 
2 0
 σu σue 
Σ= .
σue Σe

For σue 6= 0, the weights matrix is endogenous. As a result, the spatial lag

specification can be reformulated to include a matrix of control variables

from the residuals in the auxiliary regression, Z − X2 Γ, to control for the

24
endogeneity of the weights matrix. Estimation strategies are outlined in Qu

and Lee (2015) and Qu et al. (2016).

This general framework can also be exploited to develop tests for the

endogeneity of the weights, as in Qu and Lee (2015), Cheng and Lee (2017),

and Bera et al. (2018).

Alternative perspectives on the endogeneity of the spatial weights or adja-

cency structure, based on a Bayesian approach, are outlined in Chakraborty

et al. (2019), Gao and Bradley (2019), and Han et al. (2021).

3.4 Stability of weights

A final issue, pertinent to a spatial panel data setting, is whether the spa-

tial interaction is constant over time. More specifically, this pertains to the

interaction term ρt Wt . Clearly, when both ρt and Wt are allowed to vary

over time, this creates an identification problem.

Two different perspectives can be taken. In one, the weights are fixed,

Wt = W ∀ t, and the spatial autoregressive coefficient ρt is allowed to vary

by time period. For example, this can be specified as a spatial seemingly

unrelated regression specification. The stability of the coefficient can be

tested by means of a Chow test (Anselin 1988).

In another approach, the autoregressive coefficient is fixed, ρt = ρ ∀ t, and

the weights Wt are allowed to vary by time period. Examples of estimation

with time-varying weights in spatial dynamic panel models are Lee and Yu

(2012), Qu et al. (2017), and Shi and Lee (2018).

25
 Tests for structural breaks in the context of parameterized weights have

been suggested by Angulo et al. (2017, 2018). The underlying idea is to to

compare a model with a constant parameter to one with two different values,

before and after the breakpoint. In practice, a series of break points in a

search interval are considered.

4 Asymptotics

Inference in spatial (and space-time) models crucially depends on asymptotic

properties to establish the consistency (through laws of large numbers) and

asymptotic normality (through central limit theorems) of estimators and sta-

tistical tests. In order to obtain these properties, bounds need to be imposed

on the degree of dependence and heterogeneity that is allowed in a spatial

stochastic process. In practice, these regularity conditions lead to constraints

on the structure of the spatial weights and the allowed parameter space for

spatial coefficients.

The asymptotics for spatial models are not straightforward generaliza-

tions of results obtained for dependent processes in the time domain. In

fact, they differ in a number of fundamental ways. The overall framework

is based on properties of spatial random fields, i.e., stochastic processes that

appear at locations in 2-dimensional space, or, in general, in r-dimensional

space, where r is a finite integer (for an overview of the general principles,

see Guyon 1995).

26
4.1 General Concepts

In space, the mode of convergence, i.e., the manner in which the sample size

moves to the limit, or n → ∞, is not unambiguous (with n as the sample

size). On the one hand, one can envisage a fixed spatial domain from which

an increasingly dense sample of points (locations) is drawn. In the limit,

the so-called infill asymptotics form a continuous surface. This is consistent

with the field view of spatial data typically adopted in the physical sciences

and the associated methodology of geostatistics (see, e.g., Cressie 1993, Stein

1999, for an early discussion).

In contrast, in the so-called expanding domain perspective, a sample is

conceived as growing at the edges, such as an ever-expanding lattice (e.g.,

10 × 10, 12 × 12, 14 × 14, etc.). This is the approach typically taken in

econometrics (e.g., as outlined in Anselin 2002). An important consequence

of this sampling design is that the number of neighbors that interact directly

with a given observation is constrained to a (typically small) finite value.

This has major implications for obtaining the desired asymptotic properties.

Before going into some further discussion, it is important to note that

properties derived for one type of spatial asymptotics do no necessarily carry

over to the other type, something noted by Lahiri (1996), among others.

The expanding domain approach taken in econometrics brings with it

three additional complications.

First, unlike what generally holds for time series, the spacing of observa-

27
tions in two-(or three) dimensional space is uneven. Of course, regular lattice

structures can and have been considered, but these are of limited relevance

for the sampling situations typically considered by econometricians (the main

application of regular lattices would be in physics and image analysis). The

uneven spacing implies that many results obtained for dependent stochastic

processes in the time domain do not transfer directly to the spatial domain.

In addition, in a time series context, growing the sample with new obser-

vations does not affect earlier observations. However, this does not hold for

spatial samples, due to the feedback from neighbors.

More specifically, in a time series y1 , y2 , . . . yT , the addition of yT +1 does

not alter or affect yi for i = 1, 2, . . . , T . In a spatial setting, the situation

is more complex since the dependence structure impacts (almost) all other

observations as well.

Consider, for example, a pure spatial autoregressive model

X
yi = ρ wij yj + ei ,
j

with wij as the elements of the spatial weights matrix and ei as a random

error term (typically assumed to be i.i.d.). For a sample of size n, the reduced

form can be expressed as:

y(n) = (I − ρW(n) )−1 e(n) ,

where the subscript n refers to the sample size. In other words, each yi is

not only affected by ei , but also by the error terms at all other locations

28
in the system.9 In a typical times series context, the addition of s new

error terms en+1 , en+2 , . . . , en+s would simply yield s new values for yi for

i = n + 1, n + 2, . . . , n + s. However, due to the simultaneity in the spatial

process, the values for yi for i ≤ n are affected by the introduction of the new

observations as well, since the inverse terms in the reduced form will ensure

that the s additional error terms change all the elements in the vector y. As

a result, a double indexing scheme, or so-called triangular array is needed to

conceptualize the sample expansion (see also Robinson 2011, p.6, for further

illustration). In this, an observation is not only indexed by its sequence in

the sample (as is customary), but also by the sample size, as in yi(n) . In

matrix notation, we can express increasing samples as column vectors that

grow in size, stacked next to each other in an triangular shape, as in

 
y1(1) y1(2) y1(3) . . . y1(n) 
 

 y2(2) y2(3) . . . y2(n) 

 
3(3) . . . y3(n)  ,
 

 y 

 .
.. 

 
 
yn(n)

hence the designation as a triangular array (see Kelejian and Prucha 1998,

for an early discussion).

9
The inverse matrix induces a distance decay effect, such that the influence of further
neighbors declines. So, even though in principle yi would be affected by all ej , j = 1, . . . , n,
in practice, the influence of far away ej will be negligible. This will turn out to be important
in deriving asymptotic properties.

29
 The consideration of the data generation process as following a triangular

array structure introduces additional variability in the asymptotics which

needs to be accounted for with the proper LLN and CLT.

A final point pertains to the heteroskedasticity induced in a spatial au-

toregressive process for the dependent variable y whenever the number of

neighbors is not constant (Anselin 1988, 2002), even when the error terms

are i.i.d. This also follows from the reduced form. Consequently, LLN and

CLT need to account for this additional heterogeneity, even for intrinsically

stationary processes (for the error terms). In addition, more complex forms

of non-stationarity may need to be accommodated as well, such as varying

degrees of dependence across space.

4.2 Asymptotics in Linear Spatial Models

The presence of irregularly spaced observations, triangular arrays and the

induced heteroskedasticity imply that the standard approaches to obtain

asymptotic properties of estimators in time series analysis do not directly

transfer to spatial models. Instead, specialized LLN and CLT need to be

developed, an aspect largely ignored in the early literature.

In Kelejian and Prucha (2001), a general framework is outlined that es-

tablishes asymptotic normality based on a CLT for a linear quadratic form in

the error terms. The framework can be applied to a range of tests for spatial

effects in linear models, such as Moran’s I and Lagrange Multiplier (LM)

tests, as well as several estimation methods for the lag and error models, in-

30
cluding maximum likelihood (ML), quasi-maximum likelihood (QMLE), two

stage least squares (2SLS), general method of moments (GMM), and gen-

eralized empirical likelihood (GEL) (see also Kelejian and Prucha 2010, Xu

and Lee 2019, for extensions and a detailed technical discussion).

The key result is that these tests and estimators can be written as special

cases of a linear-quadratic form in the error terms:

n X
X n n
X
Qn = aij,n ei,n ej,n + bi,n ei,n ,
i=1 j=1 i=1

or, in matrix notation,

Qn = e0 n An en + b0 n en .

In these expressions, the subscript n refers to the variability with the sample

size as a result of the triangular array property. The error terms ei,n are

assumed to be zero mean and independent, but they can be heteroskedastic

of unknown form. In addition, there are some general limits on the variability

of the error term imposed as bounds on higher order moments.

The CLT of Kelejian and Prucha allows for the diagonal elements of

the matrix A to be non-zero, although for most spatial models this is not

needed, since, by convention, the diagonals of the spatial weights matrix

wii = 0. The matrix A is assumed to be symmetric with the row/column

sums bounded (i.e., their maximum sum for each n is finite). The symmetry

requirement is achieved for the typically non-symmetric spatial weights by

utilizing A = (1/2)(W + W0 ).

31
 Kelejian and Prucha (2001) invoke a CLT for martingale difference arrays

(MDA) to prove that the standardized random variable (Qn − µQn )/σQn

converges in distribution to N(0, 1). The use of an MDA imposes milder

restrictions on the memory of the system than does independence.

The concept of a martingale difference array CLT is central to deriving

the properties of estimators in more complex linear spatial models as well,

such as in multivariate and panel data settings, e.g., Keursteiner and Prucha

(2013), and the review in Xu and Lee (2019). Central limit theorems for

linear spatial models with long range dependence are derived in Lahiri and

Robinson (2016).

In practice, the requirements to achieve the asymptotic results from the

CLT boil down to restrictions on the spatial weights. For example, in the

spatial lag model, the weights matrix W should be both row- and column-

bounded. For the typical row-standardized weights, row-boundedness is

guaranteed by construction (the row sum is 1), and column boundedness

follows in most cases as well, as long as the number of neighbors is con-

strained. This is satisfied in most instances in practice. An additional con-

straint pertains to the inverse matrix (I − ρW)−1 , which should also be row-

and column-bounded. This is satisfied for a parameter space with |ρ| < 1.

4.3 Asymptotics in Nonlinear Spatial Models

For nonlinear spatial models, i.e., when the dependent variable is not a linear

function of the error terms, such as spatial probit and tobit, the framework

32
of MDAs is no longer sufficient to establish asymptotic properties.

Jenish and Prucha (2009) develop the basic theory for asymptotic prop-

erties of nonlinear spatial M-estimators (e.g., ML and GMM) by deriving a

CLT and ULLN (uniform LLN) for random spatial fields that satisfy mix-

ing and continuity conditions. The random fields allow for non-stationarity,

unevenly spaced observations and triangular arrays, as required for spatial

models.

The familiar notions of α and φ mixing are broadened to account for the

growing size of the data subsets to which they pertain (more precisely, the

associated σ-algebras). Informally, the mixing coefficients correspond to a

weak concept of dependence, i.e., the α mixing coefficient is:

α = sup(|P (A ∩ B) − P (A)P (B)|),

and the φ mixing coefficient is:

φ = sup(|P (A|B) − P (A)|),

where A and B are elements of two sub σ-algebras of a proper probability

space.

In general terms, the α measure is the difference between the joint prob-

ability and the product of the marginal probabilities, which are equal under

independence. The φ measure is the difference between a conditional and an

unconditional probability, again equal under independence.

The contribution of these weak notions of dependence is to tie them to a

measure of distance between distinct subsets in a spatial random field, such

33
that the mixing coefficient goes to zero as the distance goes to infinity.10 In

combination with some smoothness and continuity conditions, the required

CLT and ULLN can be formulated.

However, the mixing conditions as such are not sufficient to deal with

some common spatial processes, such as spatially autoregressive models. In

Jenish and Prucha (2012), new CLT and LLN are derived that broaden the

range of models with spatial dynamics that can be accommodated. The key

to this generalization is the extension of the concept of near epoch dependence

(NED) from time series analysis to the spatial domain.

Following the definition in Jenish (2012), a random field ξ is NED in a

norm q on another random field , or Lq − N ED if

sup ||ξi − E(ξi |Fi (s)||q ≤ ψ(s),

for ψ(s) such that lims→∞ ψ(s) = 0. In this expression, Fi (s) is a σ-field

on the random field j for those j within a neighborhood (radius) s from

location i. In other words, in general terms, the difference between the

value of a random variable at a location and its conditional expectation

conditional upon neighboring random variables, becomes negligible as the

distance that defines the neighborhood becomes infinitely large. With the

 process itself conforming to α mixing, this spatial version of near epoch

dependence covers a wide range of nonlinear and nonparametric models with

10
The formal conditions are more complex, and include requirements tied to the size of
the subsets as well. See Jenish and Prucha (2009) for details. For an early discussion of
the concept of mixing coefficients in the spatial domain, see Anselin (1988, pp. 45-46).

34
weak spatial dependence. Formal conditions and the application of these

concepts to derived asymptotic properties of GMM estimators can be found

in Jenish and Prucha (2012) and Xu and Lee (2019).

A slightly different perspective on formal asymptotics for nonparametric

spatial models is offered in Robinson (2009, 2011).

Deriving results that provide conditions that can readily be verified in

practice remains an active area of research in spatial econometrics.

5 Concluding Remarks

The explicit introduction of spatial effects and especially spatial interaction

into econometric models creates a number of methodological challenges that

are increasingly being taking up in econometric research. In this essay, the

focus has been on “what is special about spatial” in terms of model specifi-

cation. In particular, a range of spatial models has been reviewed and their

differences with standard (non-spatial) practice highlighted.

Two items received particular attention, since they are essentially absent

in standard econometrics: the use of a spatial weights matrix and the spe-

cial asymptotics required due to unequal spacing of observations, triangular

arrays, and non-stationarity.

While these topics provide a taste of the special features of spatial mod-

els, many other econometric aspects could not be considered within the cur-

rent scope. This includes specialized estimation methods (e.g., maximum

35
likelihood, GMM), diagnostic tests for spatial correlation (e.g., Lagrange

Multiplier tests), Bayesian spatial models, as well as the treatment of spatial

heterogeneity. The texts and review articles cited in the introduction provide

an effective entry into this literature.

References

Ahrens, A. and Bhattacharjee, A. (2015). Two-step Lasso estimation of the

spatial weights matrix. Econometrics, 3:128–155.

Angulo, A., Burridge, P., and Mur, J. (2017). Testing for a structural break

in the weight matrix of the spatial error or spatial lag model. Spatial

Economic Analysis, 12:161–181.

Angulo, A., Burridge, P., and Mur, J. (2018). Testing for breaks in the

weighting matrix. Regional Science and Urban Economics, 68:115–129.

Anselin, L. (1980). Estimation Methods for Spatial Autoregressive Structures.

Regional Science Dissertation and Monograph Series, Cornell University,

Ithaca, NY.

Anselin, L. (1986). Non-nested tests on the weight structure in spatial au-

toregressive models. Journal of Regional Science, 26:267–84.

Anselin, L. (1988). Spatial Econometrics: Methods and Models. Kluwer

Academic Publishers, Dordrecht, The Netherlands.

36
Anselin, L. (2001). Spatial econometrics. In Baltagi, B., editor, A Companion

to Theoretical Econometrics, pages 310–330. Blackwell, Oxford.

Anselin, L. (2002). Under the hood. Issues in the specification and interpre-

tation of spatial regression models. Agricultural Economics, 27(3):247–267.

Anselin, L. (2006). Spatial econometrics. In Mills, T. and Patterson, K., edi-

tors, Palgrave Handbook of Econometrics: Volume 1, Econometric Theory,

pages 901–969. Palgrave Macmillan, Basingstoke.

Anselin, L. (2010). Thirty years of spatial econometrics. Papers in Regional

Science, 89:2–25.

Anselin, L. (2021). Spatial econometrics. In Rey, S. and Franklin, R., ed-

itors, Handbook of Spatial Analysis in the Social Sciences. Edward Elgar

Publishing, Cheltenham, UK. forthcoming.

Anselin, L. and Arribas-Bel, D. (2013). Spatial fixed effects and spatial

dependence in a single cross-section. Papers in Regional Science, 92:3–17.

Anselin, L. and Bera, A. (1998). Spatial dependence in linear regression

models with an introduction to spatial econometrics. In Ullah, A. and

Giles, D. E., editors, Handbook of Applied Economic Statistics, pages 237–

289. Marcel Dekker, New York.

Anselin, L., Le Gallo, J., and Jayet, H. (2008). Spatial panel econometrics.

In Matyas, L. and Sevestre, P., editors, The Econometrics of Panel Data,

37
 Fundamentals and Recent Developments in Theory and Practice (3rd Edi-

tion), pages 627–662. Springer-Verlag, Berlin.

Anselin, L. and Rey, S. J. (2014). Modern Spatial Econometrics in Practice,

A Guide to GeoDa, GeoDaSpace and PySAL. GeoDa Press, Chicago, IL.

Arbia, G. (2014). A Primer for Spatial Econometrics. With Applications in

R. Palgrave Macmillan, Basingstoke, UK.

Assunção, R. and Krainski, E. (2009). Neighborhood dependence in Bayesian

spatial models. Biometrical Journal, 5:851–869.

Bai, J., Baltagi, B., and Pesaran, H. (2016). Cross-sectional dependence

in panel data models: a special issue. Journal of Applied Econometrics,

31:1–3.

Bailey, N., Kapetianos, G., and Pesaran, M. H. (2016). Exponent of cross-

sectional dependence: Estimation and inference. Journal of Applied Econo-

metrics, 31:929–960.

Baltagi, B. B., LeSage, J. P., and Pace, R. K. (2016). Spatial Econometrics:

Qualitative and Limited Dependent Variables (Advances in Econometrics,

Vol 37). Emerald Group Publishing, Bingley, UK.

Baltagi, B. H., Egger, P., and Pfaffermayr, M. (2013). A generalized spatial

panel data model with random effects. Econometric Reviews, 32:650–685.

38
Baltagi, B. H. and Liu, L. (2011). An improved generalized moments estima-

tor for a spatial moving average error model. Economics Letters, 113:282–

284.

Baltagi, B. H. and Pirotte, A. (2011). Seemingly unrelated regressions with

spatial error components. Empirical Economics, 40:5–49.

Baltagi, B. H., Song, S. H., Jung, B. C., and Koh, W. (2007). Testing for

serial correlation, spatial autocorrelation and random effects using panel

data. Journal of Econometrics, 140:5–51.

Bavaud, F. (1998). Models for spatial weights: A systematic look. Geograph-

ical Analysis, 30:153–171.

Beenstock, M. and Felsenstein, D. (2012). Nonparametric estimation of the

spatial connectivity matrix using spatial panel data. Geographical Analysis,

44:386–397.

Benjanuvatra, S. and Burridge, P. (2015). QML estimation of the spatial

weight matrix in the MR-SAR model. Discussion Papers in Economics

No. 15/24, University of York, York, UK.

Bera, A. K., Doğan, O., and Taşpinar, S. (2018). Simple tests for endogene-

ity of spatial weights matrices. Regional Science and Urban Economics,

69:130–142.

39
Bhat, C. R., Paleti, R., and Singh, P. (2014). A spatial multivariate count

model for firm location decisions. Journal of Regional Science, 54:462–502.

Bhattacharjee, A., Castro, E., and ao Marques, J. (2012). Spatial interactions

in hedonic pricing models: the urban housing market of Aveiro, Portugal.

Spatial Economic Analysis, 7:133–167.

Bhattacharjee, A. and Holly, S. (2011). Structural interactions in spatial

panels. Empirical Economics, 40:69–94.

Bhattacharjee, A. and Jensen-Butler, C. (2013). Estimation of the spatial

weights matrix under structural constraints. Regional Science and Urban

Economics, 43:617–634.

Bodson, P. and Peeters, D. (1975). Estimation of the coefficients of a linear

regression in the presence of spatial autocorrelation: An application to a

Belgian labour-demand function. Environment and Planning A, 7:455–72.

Burridge, P. (1981). Testing for a common factor in a spatial autoregressive

model. Environment and Planning A, 13:795–800.

Calabrese, R. and Elkink, J. A. (2014). Estimators of binary spatial au-

toregressive models: a Monte Carlo study. Journal of Regional Science,

54:664–687.

Case, A. C. (1991). Spatial patterns in household demand. Econometrica,

59:953–965.

40
Case, A. C. (1992). Neighborhood influence and technological change. Re-

gional Science and Urban Economics, 22:491–508.

Chakraborty, S., Bhattacharjee, A., and Maiti, T. (2019). Structural factor-

ization of latent adjacency matrix, with an application to auto industry

networks. Sankhya B. doi:10.1007/s13571-019-001001-0.

Cheng, W. and Lee, L. (2017). Testing endogeneity of spatial and social

networks. Regional Science and Urban Economics, 64:81–97.

Chi, G. and Zhu, J. (2020). Spatial regression models for the social sciences.

Sage, Thousand Oaks, CA.

Chudik, A. and Pesaran, M. H. (2015). Large panel data models with cross-

sectional dependence: A survey. In Baltagi, B. H., editor, The Oxford

Handbook of Panel Data, pages 3–45. Oxford University Press, Oxford,

UK.

Chudik, A., Pesaran, M. H., and Tosetti, E. (2011). Weak and strong cross

section dependence and estimation of large panels. Econometrics Journal,

14:C45–C90.

Cliff, A. and Ord, J. K. (1981). Spatial Processes: Models and Applications.

Pion, London.

Cotteleer, G., Stobbe, T., and van Kooten, G. C. (2011). Bayesian model

41
 averaging in the context of spatial hedonic pricing: an application to farm-

land values. Journal of Regional Science, 51:540–557.

Cressie, N. (1993). Statistics for Spatial Data. Wiley, New York.

Debarsy, N. and Ertur, C. (2019). Interaction matrix selection in spatial au-

toregressive models with an application to growth theory. Regional Science

and Urban Economics, 75.

Debarsy, N. and LeSage, J. P. (2018). Flexible dependence modeling using

convex combinations of different types of connectivity structures. Regional

Science and Urban Economics, 69:48–68.

Delgado, M. A. and Robinson, P. M. (2015). Non-nested testing of spatial

correlation. Journal of Econometrics, 187:385–401.

Dubé, J. and Legros, D. (2014). Spatial Econometrics Using Microdata.

ISTE-Wiley, London - Hoboken.

Dubin, R., Pace, R. K., and Thibodeau, T. G. (1999). Spatial autoregression

techniques for real estate data. Journal of Real Estate Literature, 7:79–95.

Elhorst, J. P. (2010). Applied spatial econometrics: Raising the bar. Spatial

Economic Analysis, 5:9–28.

Elhorst, J. P. (2012). Dynamic spatial panels: Models, methods, and infer-

ences. Journal of Geographical Systems, 14:5–28.

42
Elhorst, J. P. (2014a). Spatial Econometrics, From Cross-Sectional Data to

Spatial Panels. Springer, Heidelberg.

Elhorst, J. P. (2014b). Spatial panel models. In Fischer, M. M. and Nijkamp,

P., editors, Handbook of Regional Science, pages 1634–1652. Springer Ver-

lag, Berlin, Heidelberg.

Fernández-Vázquez, E., Mayor-Fernández, M., and Rodrı́guez-Vález, J.

(2009). Estimating spatial autoregressive models by GME-GCE tech-

niques. International Regional Science Review, 32:148–172.

Fingleton, B. (2008). A generalized method of moments estimator for a

spatial model with moving average errors, with application to real estate

prices. Empirical Economics, 34:35–57.

Fleming, M. (2004). Techniques for estimating spatially dependent discrete

choice models. In Anselin, L., Florax, R. J., and Rey, S. J., editors, Ad-

vances in Spatial Econometrics, pages 145–168. Springer-Verlag, Heidel-

berg.

Florax, R. and Folmer, H. (1992). Specification and estimation of spatial

linear regression models: Monte Carlo evaluation of pre-test estimators.

Regional Science and Urban Economics, 22:405–432.

Fusco, E. and Vidoli, F. (2013). Spatial stochastic frontier models: control-

ling spatial global and local heterogeneity. International Review of Applied

Economics, 27:679–694.

43
Gao, H. and Bradley, J. R. (2019). Bayesian analysis of areal data with

unknown adjacencies using the stochastic edge mixed effects model. Spatial

Statistics, 31:100357.

Gerkman, L. M. and Ahlgren, N. (2014). Practical proposals for specifying

k-nearest neighbours weights matrices. Spatial Economic Analysis, 9:260–

283.

Getis, A. (2009). Spatial weights matrices. Geographical Analysis, 41:404–

410.

Gupta, A. (2019). Estimation of spatial autoregressions with stochastic

weight matrices. Econometric Theory, 35:417–463.

Guyon, X. (1995). Random Fields on a Network. Modeling, Statistics and

Applications. Springer-Verlag, NewYork, NY.

Haining, R. (1978). The moving average model for spatial interaction. Trans-

actions, Institute of British Geographers, 3:202–225.

Han, X., Hsieh, C.-S., and Ko, S. I. (2021). Spatial modeling approach for

dynamic network formation and interactions. Journal of Business and

Economic Statistics, 39:120–135.

Han, X. and Lee, L. (2013). Model selection using J-test for the autoregressive

model vs. the matrix exponential spatial model. Regional Science and

Urban Economics, 43:250–271.

44
Harris, R., Moffat, J., and Kravtsova, V. (2011). In search of ‘W’. Spatial

Economic Analysis, 6:249–270.

Hastie, T., Tibshirani, R., and Friedman, J. (2009). The Elements of Statis-

tical Learning (2nd Edition). Springer, New York, NY.

Jenish, N. (2012). Nonparametric spatial regression under near-epoch de-

pendence. Journal of Econometrics, 167:224–239.

Jenish, N. and Prucha, I. R. (2009). Central limit theorems and uniform laws

of large numbers for arrays of random fields. Journal of Econometrics,

150:86–98.

Jenish, N. and Prucha, I. R. (2012). On spatial processes and asymptotic

inference under near-epoch dependence. Journal of Econometrics, 170:178–

190.

Jin, F. and Lee, L.-F. (2013). Cox-type tests for competing spatial autore-

gressive models with spatial autoregressive disturbances. Regional Science

and Urban Economics, 43:590–616.

Kapetanios, G., Pesaran, M. H., and Yamagata, T. (2011). Panels with non-

stationary multifactor error structures. Journal of Econometrics, 160:326–

348.

Kapoor, M., Kelejian, H. H., and Prucha, I. (2007). Panel data models with

45
 spatially correlated error components. Journal of Econometrics, 140:97–

130.

Kelejian, H. H. (2008). A spatial J-test for model specification against a

single or a set of non-nested alternatives. Letters in Spatial and Resource

Sciences, 1:3–11.

Kelejian, H. H. and Piras, G. (2011). An extension of Kelejian’s J-test

for non-nested spatial models. Regional Science and Urban Economics,

41:281–292.

Kelejian, H. H. and Piras, G. (2014). Estimation of spatial models with

endogenous weighting matrices, and an application to a demand model for

cigarettes. Regional Science and Urban Economics, 46:140–149.

Kelejian, H. H. and Piras, G. (2017). Spatial Econometrics. Academic Press,

London, UK.

Kelejian, H. H. and Prucha, I. (1998). A generalized spatial two stage least

squares procedures for estimating a spatial autoregressive model with au-

toregressive disturbances. Journal of Real Estate Finance and Economics,

17:99–121.

Kelejian, H. H. and Prucha, I. (2001). On the asymptotic distribution of

the Moran I test statistic with applications. Journal of Econometrics,

104(2):219–257.

46
Kelejian, H. H. and Prucha, I. R. (2002). 2SLS and OLS in a spatial au-

toregressive model with equal spatial weights. Regional Science and Urban

Economics, 32(6):691–707.

Kelejian, H. H. and Prucha, I. R. (2007). HAC estimation in a spatial frame-

work. Journal of Econometrics, 140:131–154.

Kelejian, H. H. and Prucha, I. R. (2010). Specification and estimation of

spatial autoregressive models with autoregressive and heteroskedastic dis-

turbances. Journal of Econometrics, 157:53–67.

Kelejian, H. H., Prucha, I. R., and Yuzefovich, Y. (2006). Estimation prob-

lems in models with spatial weighting matrices which have blocks of equal

elements. Journal of Regional Science, 46:507–515.

Keursteiner, G. M. and Prucha, I. R. (2013). Limit theory for panel data

models with cross sectional dependence and sequential exogeneity. Journal

of Econometrics, 174:107–126.

Koroglu, M. and Sun, Y. (2016). Functional-coefficient spatial Durbin models

with nonparametric spatial weights: an application to economic growth.

Econometrics, 4:6.

Kostov, P. (2010). Model boosting for spatial weighting matrix selection in

spatial lag models. Environment and Planning B, 37:533–549.

47
Kostov, P. (2013). Empirical likelihood estimation of the spatial quantile

regression. Journal of Geographical Systems, 15:51–69.

Lahiri, S. (1996). On the inconsistency of estimators under infill asymptotics

for spatial data. Sankhya A, 58:403–417.

Lahiri, S. and Robinson, P. M. (2016). Central limit theorems for long range

dependent spatial linear processes. Bernoulli, 22:345–375.

Lam, C. and Souza, P. C. (2016). Detection and estimation of block structure

in spatial weight matrix. Econometric Reviews, 35:1347–1376.

Lam, C. and Souza, P. C. (2020). Estimation and selection of spatial weight

matrix in a spatial lag model. Journal of Business and Economic Statistics,

38:693–710.

Lambert, D. M., Brown, J. P., and Florax, R. J. (2010). A two-step estimator

for a spatial lag model of counts: theory, small sample performance and

an application. Regional Science and Urban Economics, 40:241–252.

Lee, L.-F. (2002). Consistency and efficiency of least squares estimation

for mixed regressive, spatial autoregressive models. Econometric Theory,

18(2):252–277.

Lee, L.-F. and Yu, J. (2010). Some recent developments in spatial panel data

models. Regional Science and Urban Economics, 40:255–271.

48
Lee, L.-F. and Yu, J. (2011). Estimation of spatial panels. Foundations and

Trends in Econometrics, 4:1–164.

Lee, L.-F. and Yu, J. (2012). QML estimation of spatial dynamic panel

data models with time varying spatial weights matrices. Spatial Economic

Analysis, 7:31–74.

Lee, L.-F. and Yu, J. (2015). Spatial panel data models. In Baltagi, B. H.,

editor, The Oxford Handbook of Panel Data, pages 363–401. Oxford Uni-

versity Press, Oxford, UK.

LeSage, J. P. and Pace, R. K. (2007). A matrix exponential spatial specifi-

cation. Journal of Econometrics, 140:190–214.

LeSage, J. P. and Pace, R. K. (2009). Introduction to Spatial Econometrics.

CRC Press, Boca Raton, FL.

LeSage, J. P. and Pace, R. K. (2014). The biggest myth in spatial economet-

rics. Econometrics, 2:217–249.

LeSage, J. P. and Parent, O. (2007). Bayesian model averaging for spatial

econometric models. Geographical Analysis, 39:241–267.

Liesenfeld, R., Richard, J.-F., and Vogler, J. (2017). Likelihood-based in-

ference and prediction in spatio-temporal panel count models for urban

crimes. Journal of Applied Econometrics, 32:600–620.

49
Martellosio, F. (2011). Nontestability of equal weights spatial dependence.

Econometric Theory, 27:1369–1375.

Martellosio, F. (2012). The correlation structure of spatial autoregressions.

Econometric Theory, 28:1373–1391.

McMillen, D. P. (2013). Quantile Regression for Spatial Data. Springer-

Verlag, Heidelberg.

Meen, G. (1996). Spatial aggregation, spatial dependence and predictability

in the UK housing market. Housing Studies, 11:345–372.

Ord, J. K. (1975). Estimation methods for models of spatial interaction.

Journal of the American Statistical Association, 70:120–126.

Orea, L. and Àlvarez, I. C. (2019). A new stochastic frontier model with

cross-sectional effects in both noise and inefficiency terms. Journal of

Econometrics, 213:556–577.

Patuelli, R. and Arbia, G. (2016). Spatial Econometric Interaction Modelling.

Springer-Verlag, Heidelberg.

Pesaran, M. H. (2006). Estimation and inference in large heterogenous panels

with a multifactor error structure. Econometrica, 74:967–1012.

Pesaran, M. H. and Tosetti, E. (2011). Large panels with common factors

and spatial correlation. Journal of Econometrics, 161:182–202.

50
Pinkse, J. and Slade, M. E. (1998). Contracting in space: An application

of spatial statistics to discrete-choice models. Journal of Econometrics,

85:125–154.

Pinkse, J., Slade, M. E., and Brett, C. (2002). Spatial price competition: A

semiparametric approach. Econometrica, 70(3):1111–1153.

Piras, G. and Lozano-Gracia, N. (2012). Spatial J-test: some Monte Carlo

evidence. Statistics and Computing, 22:169–183.

Piribauer, P. and Fischer, M. M. (2015). Model uncertainty in matrix expo-

nential spatial growth regression models. Geographical Analysis, 47:240–

261.

Plümper, T. and Neumayer, E. (2010). Model specification in the analysis of

spatial dependence. European Journal of Political Research, 49:418–442.

Qu, X. and Lee, L. (2015). Estimating a spatial autoregressive model with an

endogenous spatial weight matrix. Journal of Econometrics, 184:209–232.

Qu, X., Lee, L., and Yu, J. (2017). QML estimation of spatial dynamic

panel data models with endogenous time varying spatial weights matrices.

Journal of Econometrics, 197:173–201.

Qu, X., Wang, X., and Lee, L. (2016). Instrumental variable estimation of

a spatial dynamic panel model with endogenous spatial weights when T is

small. The Econometrics Journal, 19:261–290.

51
Robinson, P. M. (2009). Large-sample inference on spatial dependence.

Econometrics Journal, 12:S68–S82.

Robinson, P. M. (2011). Asymptotic theory for nonparametric regression

with spatial data. Journal of Econometrics, 165:5–19.

Rodrigues, E., Assunção, R., and Dey, D. K. (2014). A closer look at the

spatial exponential matrix specification. Spatial Statistics, 9:109–121.

Seya, H., Yamagata, Y., and Tsutsumi, M. (2013). Automatic selection of

a spatial weight matrix in spatial econometrics: application to a spatial

hedonic approach. Regional Science and Urban Economics, 43:429–444.

Shi, W. and Lee, L. (2018). A spatial panel data model with time varying

endogenous weights matrices and common factors. Regional Science and

Urban Economics, 72:6–34.

Smirnov, O. (2010). Modeling spatial discrete choice. Regional Science and

Urban Economics, 40:292–298.

Stakhovych, S. and Bijmolt, T. H. (2009). Specification of spatial models: A

simulation study on weights matrices. Papers in Regional Science, 88:389–

408.

Stein, M. L. (1999). Interpolation of Spatal Data, Some Theory for Kriging.

Springer-Verlag, New York.

52
Strauß, M. E., Mezzetti, M., and Leorato, S. (2017). Is a matrix exponential

specification suitable for the modeling of spatial correlation structures?

Spatial Statistics, 20:221–243.

Sun, Y. (2016). Functional-coefficient spatial autoregressive models with

nonparametric spatial weights. Journal of Econometrics, 195:134–153.

Vega, S. H. and Elhorst, J. P. (2015). The SLX model. Journal of Regional

Science, 55:339–363.

Wall, M. M. (2004). A close look at the spatial structure implied by the CAR

and SAR models. Journal of Statistical Planning and Inference, 121:311–

324.

Wang, S., Wang, S., and Smith, P. (2013). Two spatial non-nested tests for

weight structure in the spatial autoregressive model. Geographical Analy-

sis, 45:345–358.

Whittle, P. (1954). On stationary processes in the plane. Biometrika, 41:434–

449.

Xu, X. and Lee, L. (2015). Maximum likelihood estimation of a spatial

autoregressive Tobit model. Journal of Econometrics, 188:264–280.

Xu, X. and Lee, L. (2018). Sieve maximum likelihood estimation of the

spatial autoregressive Tobit model. Journal of Econometrics, 203:96–112.

53
 Xu, X. and Lee, L. (2019). Theoretical foundations for spatial econometric

research. Regional Science and Urban Economics, 76:2–12.

Zhang, X. and Yu, J. (2018). Spatial weights matrix selection and model av-

eraging for spatial autoregressive models. Journal of Econometrics, 203:1–

18.

54

View publication stats