Spatial Statistics 10 (2014) 1–11

Contents lists available at ScienceDirect

Spatial Statistics
journal homepage: www.elsevier.com/locate/spasta

A quality assessment of eigenvector spatial

filtering based parameter estimates for the
normal probability model
Yongwan Chun ∗ , Daniel A. Griffith 1
School of Economic, Political and Policy Sciences, The University of Texas at Dallas, 800 West Campbell
Road, Richardson, TX 75080, United States

article info abstract

Article history: Eigenvector spatial filtering, which introduces a subset of eigen-
Received 28 March 2013 vectors extracted from a spatial weights matrix as synthetic con-
Accepted 19 April 2014 trol variables in a regression model specification, furnishes a
Available online 5 May 2014
solution to extraordinarily intricate statistical modeling problems
involving spatial dependences. It accounts for spatial autocorrela-
Keywords:
tion in standard specifications of regression models. But the quality
Eigenvector spatial filtering
Unbiasedness
of the resulting regression parameter estimates has yet to be ascer-
Efficiency tained. The estimator properties to establish include unbiasedness,
Consistency efficiency, and consistency. The purpose of this paper is to demon-
strate these estimator properties for linear regression parameters
based on eigenvector spatial filtering, including a comparison with
the simultaneous autoregressive (SAR) model. Eigenvector spatial
filtering methodology requires the judicious selection of eigenvec-
tors, whose number tends to increase with both level of linear re-
gression residual spatial autocorrelation and the number of areal
units. A logistic regression description of the number of eigenvec-
tors selected in a simulation pilot study suggests estimator consis-
tency.
© 2014 Elsevier B.V. All rights reserved.

∗ Corresponding author. Tel.: +1 972 883 4719; fax: +1 972 883 6436.
E-mail addresses: ywchun@utdallas.edu, ywchun@gmail.com (Y. Chun), dagriffith@utdallas.edu (D.A. Griffith).
1 Tel.: +1 972 883 4950; fax: +1 972 883 4967.

http://dx.doi.org/10.1016/j.spasta.2014.04.001
2211-6753/© 2014 Elsevier B.V. All rights reserved.
2 Y. Chun, D.A. Griffith / Spatial Statistics 10 (2014) 1–11

1. Introduction

Classical mathematical statistics avoid correlation amongst observations by assuming that they are
independent. Consequently, joint probability distributions are the products of marginal probability
distributions. But many datasets comprise observational units that are related to each other in some
way, such as being adjacent in a time sequence (e.g., repeated measures, time series), being linked
members in a subgroup of a population (e.g., family ties, social networks), or being neighbors in
a geographic distribution (e.g., spatial series). Consequently, joint probability distributions are the
products of conditional probability distributions. Autoregressive model specifications furnish one
popular modification to classical mathematical statistics to capture these observational dependences.
In a regression context, these specifications result in the response variable, Y , being on both sides of
an equation: the left-hand side contains Y , and the right-hand side contains some linear combination
of observed values (yi ), the n (i.e., sample size/number of observations) realizations of Y , such that
each yi is not a function of itself. In classical mathematical statistics, all of the linear combination
weights are zero. Another popular approach deals with the inter-correlations among observations
(e.g., geostatistics, geographic spectral analysis). Eigenfunction spatial filtering is a third approach.
Spatial autoregressive models include the conditional/spatial error (CAR), the simultaneous/spatial
error (SAR/SE), and the response/spatial lag (AR/SL) specifications, all of which describe Gaussian
random variables. Although the CAR process is conditionally specified based on a Markov random
field structure, the SAR and AR processes are defined with spatially lagged terms. A SAR model has
only a spatial autocorrelation component in its error term, and an AR model contains a spatially
lagged variable in its model specification. Auto-models exist for non-normal random variables
(Besag, 1974; Hardouin and Yao, 2007), too, including the: auto-logistic/binomial, auto-Poisson, auto-
exponential, auto-gamma, and auto-beta. An eigenvector spatial filter (ESF) model specification is
more flexible, and can describe Gaussian, Poisson, and binomial random variables containing positive
spatial autocorrelation (i.e., similar values tending to cluster together on a map), whereas auto-models
like the Poisson, exponential, and gamma schemes have integrability conditions that ensure that
they are able to describe only spatial competition between neighboring sites (i.e., negative spatial
autocorrelation). This latter restriction may be circumvented by transforming these types of variables
onto a bounded range, such as by employing truncation, allowing them to describe positive spatial
autocorrelation. Some comparisons between the SAR and ESF specifications appear in, for example,
Getis and Griffith (2002). Comparisons also exist between the ESF and the auto-logistic and the auto-
Poisson (e.g., Griffith, 2004) specifications. Relationship articulations between the ESF and both the
AR and SAR specifications appear in Tiefelsdorf and Griffith (2007).
Although eigenvector spatial filtering has become more popular in addressing spatial autocorre-
lation latent in georeferenced data, the quality of ESF-based estimators has not been thoroughly in-
vestigated. The statistical qualities of ESF-based estimators, including unbiasedness, efficiency, and
consistency, remain under- or un-explored. Such a quality assessment of ESF-based estimators can
bolster the efficacy of eigenvector spatial filtering methodology, documenting that it furnishes a solid
foundation for analyzing georeferenced data in linear and generalized linear regression analysis. The
goal of this paper is to summarize selected empirical results in order to begin to fill this gap in the
literature. It utilizes an SAR data generating process and a judiciously selected purposeful sample
(criterion: span a wide range of n) of readily available geographic landscape datasets to execute an
exploratory pilot study.

2. Background: eigenvector spatial filtering

Eigenvector spatial filtering methodology utilizes an eigenfunction decomposition of the

transformed spatial weights matrix, (I − 11T /n)C(I − 11T /n) = MCM, where I is the identity matrix, 1
is an n-by-1 vector of ones, T denotes the matrix transpose operator, C is a binary spatial weights
matrix, and M = (I − 11T /n). This decomposition generates n eigenvalues and their associated
n eigenvectors. The n eigenvalues can be denoted by the set {λ1 , λ2 , . . . , λn }, where λ1 > λ2 >
· · · > λn . Their corresponding n eigenvectors can be denoted as E = {E1 , E2 , . . . , En }, where each
eigenvector, Ej , is an n-by-1 vector. These eigenvectors furnish distinct map pattern descriptions of
 Y. Chun, D.A. Griffith / Spatial Statistics 10 (2014) 1–11 3

latent spatial autocorrelation in georeferenced variables, because they are mutually both orthogonal
and uncorrelated (see Griffith, 2003 for details). Also, the map patterns by the eigenvectors portray
a specific level of spatial autocorrelation. The Moran Coefficient (MC) for a mapped eigenvector is a
direct function of its corresponding eigenvalue (Griffith, 1996; Tiefelsdorf and Boots, 1995). These
eigenvectors also can be extracted from a truncated inter-location distance matrix (Borcard and
Legendre, 2002; Borcard et al., 2004; Dray et al., 2006), rather than a binary C matrix. Griffith (2010)
furnishes an overview of various other spatial filtering methodologies.
ESF methodology utilizes the eigenvectors to account for spatial autocorrelation by adding a linear
combination of these eigenvectors. In linear regression, the ESF model specification may be written
as Y = Xβ + Ek βE + ε, where X is an n-by-(p + 1) matrix containing p covariates (including a vector of
1s for the intercept term), β is the corresponding (p + 1)-by-1 vector of regression parameters, Ek is
an n-by-K matrix containing K eigenvectors, βE is the corresponding vector of regression parameters,
and ε ∼ N 0, Iσ 2 . Because the linear combination of the eigenvectors, Ek βE , accounts for spatial
 
autocorrelation, the ESF linear regression specification does not suffer from spatially autocorrelated
residuals.
An ESF model specification is implemented by identifying a feasible set of eigenvectors. The
identification has been achieved through a two-stage process. In the first stage, a candidate set
of eigenvectors, which is a noticeably smaller subset (i.e., K ≪ n) of the entire set of n
eigenvectors, is constructed by eliminating eigenvectors that generally do not account for significant
spatial autocorrelation. These include eigenvectors portraying negligible (i.e., near zero) spatial
autocorrelation (i.e., their MCs are close to the expected value for zero spatial autocorrelation) and
eigenvectors portraying negative spatial autocorrelation when positive spatial autocorrelation needs
to be addressed (or eigenvectors portraying positive spatial autocorrelation when negative spatial
autocorrelation needs to be addressed). The overwhelming number of empirical cases to date contain
positive spatial autocorrelation. One criterion for identifying a candidate set is employing a threshold
minimum MC of 0.25, which relates to roughly 5% of the variance in a response variable being
attributable to positive spatial autocorrelation. Another criterion devised by Griffith and Chun (2009)
builds on the level of spatial autocorrelation detected in a response variable, Y , or residuals when a
specification contains covariates, and may be summarized as follows:

MCj ≥ 2.9970 − 2.8805/(1 + e−0.6606−0.2525 zMC ), (1)

where MCj is the MC value corresponding to the jth eigenvector, and zMC denotes the z-score
of the MC for the response variable Y (or some transformed version of it, such as a Box–Cox
power transformation, if used), or regression residuals when covariates appear in a linear model
specification.2 In the second stage, a more concise set of eigenvectors is identified through a stepwise
procedure. Often, this identification is achieved with a conventional stepwise regression technique,
which is available in most statistical software packages.3 An inclusion of many eigenvectors with
near-zero coefficients would inflate the variance of estimates. Bonferroni correction would furnish
a suitable significance level for stepwise regression (e.g., Griffith and Chun, 2014). Alternatively,
minimizing spatial autocorrelation in residuals can be used to select eigenvectors in regression
(Tiefelsdorf and Griffith, 2007). Here, the stepwise (i.e., a hybrid mixture of the forward and backward)
regression technique is preferred because it minimizes inflation of the variance estimates due to the
inclusion of extraneous eigenvectors, which does nothing more than decrease the number of degrees
of freedom, inflating the mean squared error (Greene, 2008).

2 Griffith and Chun (2009) discuss that an ESF model with stepwise eigenvector selection can over- or under-correct for
spatial autocorrelation when a candidate set size is too large or too small, respectively. They develop the equation through
a simulation experiment with combinations of different spatial autocorrelation levels in a response variable and different
landscape and candidate set sizes.
3 Blanchet et al. (2008), Bini et al. (2009), and Peres-Neto and Legendre (2010) discuss how to select eigenvectors using the
stepwise forward selection technique. However, because the eigenvectors are uncorrelated with and orthogonal to each other,
these two approaches produce a near-identical result in an empirical data analysis.
4 Y. Chun, D.A. Griffith / Spatial Statistics 10 (2014) 1–11

3. Unbiasedness in the presence of spatial autocorrelation

One useful statistical property indexing the quality of an estimator is that its sampling distribution
average equals its corresponding population parameter. This property relates to the notion of
accuracy. The lack of bias is an attractive and desired, but not necessarily a fundamental or required,
estimator feature. It is a pragmatic feature in that the lack of systematic deviation from a population
parameter allows quality assessment to focus on other properties of an estimator.
Pace et al. (2013) show, utilizing the Frisch–Waugh–Lovell theorem, that an ESF produces unbiased
estimators for coefficients of independent variables when a SAR model specification is used as the
data generating process. Conclusions about biasedness are sensitive to which terms contain spatial
autocorrelation in a model specification (see Griffith, 1976). Nevertheless, Pace et al. (2013) conclude
that an ESF reduces bias found in ordinary least squares (OLS) estimators, thus improving upon OLS
results.
For an SAR model data generating process, the variable Y can be expressed as Y = Xβ + V−1/2 ε,
where V is the inverse covariance matrix (e.g., (I − ρ C)T (I − ρ C) or (I − ρ W)T (I − ρ W), where
W is a row-standardized spatial weights matrix, and ρ is a spatial autocorrelation parameter), and
ε ∼ N 0, Iσ 2 . The OLS, generalized least squares (GLS), and ESF estimators are unbiased:


OLS: E(β̂) = (XT X)−1 XT E(Y) = (XT X)−1 XT (Xβ + 0) = β (2)

GLS: E(β̂) = (X VX)

T −1 T
X VE(Y) = (X VX) T
X V(Xβ + 0) = β
−1 T
(3)

ESF: E(β̂) = (XT X − XT Ek ETk X)−1 XT E(Y) − (XT X − XT Ek ETk X)−1 XT Ek ETk E(Y)
= (XT X − XT Ek ETk X)−1 (XT − XT Ek ETk )(Xβ + 0) = β (4)
where E(·) denotes the calculus of expectations operator. However, the unbiasedness of the OLS
estimator does not hold for AR model process. The remainder of this paper evaluates the SAR case.
Evaluation of the AR case will be undertaken in future research.

4. Efficiency in the presence of spatial autocorrelation

The degree of sample-to-sample estimator stability is a quality property often more important
than unbiasedness. Relatively efficient estimators have smaller variances, and hence require smaller
sample sizes to achieve a given level of precision. The preceding sampling variance of the SAR model
specification implies that the expected value of the OLS and GLS constant mean specification (i.e.,
X = 1) variance is
OLS: E[(XT X)−1 XT V−1/2 εεT V−T/2 X(XT X)−1 ]
= (XT X)−1 XT V−1 X(XT X)−1 σ 2 ̸= (XT X)−1 σ 2 , V ̸= I.
When X = 1,
(XT X)−1 XT V−1 X(XT X)−1 σ 2 = [1T V−1 1/n2 ]σ 2 = [1T [(I − ρ W)T (I − ρ W)]−1 1/n2 ]σ 2 . (5)
E(σ̂ 2 ) = E[(Y − Xβ)T V−1 (Y − Xβ)/n] = TR(V−1 )σ 2 /n.
Note that E(σ̂ 2 ) ̸= (Y − Xβ)T (Y − Xβ)/n,
where TR denotes the matrix trace operator,
GLS: E[(XT VX)−1 XT VV−1/2 εεT V−T/2 VX(XT VX)−1 ] = (XT VX)−1 σ 2 .
When X = 1,
(XT VX)−1 σ 2 = (1T V1)−1 σ 2
= [1T (I − ρ W)T (I − ρ W)1]−1 σ 2 = σ 2 /[n(1 − ρ)2 ]. (6)
E(σ̂ ) = E[(Y − Xβ) V(Y − Xβ)/n] = σ .
2 T 2

The unbiased OLS variance estimator is σ̂ 2 /[nTR(V−1 )], with the denominator term TR(V−1 ) adjusting
for variance inflation in the presence of positive spatial autocorrelation. The GLS estimator is the
 Y. Chun, D.A. Griffith / Spatial Statistics 10 (2014) 1–11 5

Table 1
Summary spatial autocorrelation and ESF results for selected geographic landscapes.
s2GLS s2ESF
Landscape Attribute zMC n nselected ncandidate
s2OLS s2GLS

Columbus, OH Crime 6.2507 49 5 12 4.7431 0.0705

North Carolina SIDS 10.9617 100 13 24 0.4733 0.5574
Murray superfund site Arsenic 8.1821 253 22 64 0.7112 0.8889
Mercer-Hall Yield 14.4023 500 60 155 1.0773 0.3284
Toronto Population density 28.9400 731 115 308 3.3166 0.0881
High Peak Biomass 36.3366 900 244 349 0.0613 0.4801
Wiebe Yield 37.6297 1500 138 462 1.6719 0.1349
China Population density 61.1059 2379 248 561 0.3613 0.7543
Note: the ratios are for the sums of the variances (see Kramer and Donninger, 1987).
Note: nselected denotes the number of selected eigenvectors, ncandidate denotes the number of eigenvectors in a candidate
set, s2OLS denotes the simulated standard deviation of the OLS estimator, s2GLS denotes the simulated standard deviation of
the GLS estimator, and s2ESF denotes the simulated standard deviation of the ESF estimator.
Note: these datasets have been analyzed in the literature. Details about 5 datasets can be found in Griffith and Layne (1999):
Columbus (pp. 222–227), North Carolina (pp. 397–405), Mercer-Hall (pp. 311–316), High Peak (pp. 292–302), and Wiebe
(pp. 316–320). Details of the other datasets can be found in: Griffith (2002) for the Murray superfund site, Griffith (1997)
for Toronto, and Griffith (2006) for China.
Note: results in this table differ from those in Griffith and Chun (2014) because no covariates are included in the analyses
for this table.

more efficient of the two (Griffith, 1988; Cordy and Griffith, 1993) in part because the OLS estimator
includes a variance inflation factor (VIF) due to the presence of positive spatial autocorrelation.
Most of the efficiency is lost through the variance estimator (i.e., using (XT X)−1 in place of
(XT X)−1 XT V−1 X(XT X)−1 ) of the regression coefficients.
Meanwhile, the ESF estimator is
ESF: (XT X − XT Ek ETk X)−1 XT (Ik − Ek ETk )V−1
× (Ik − Ek ETk )X(XT X − XT Ek ETk X)−1 σ 2 = (XT X − XT Ek ETk X)−1 σ 2
 
 ∞
+ (X X − X Ek Ek X) X
T T T −1 T
ρ En−k 3n−k En−k X(XT X − XT Ek ETk X)−1 σ 2
h h T
(7)
h=1

where 3 is a diagonal matrix containing the corresponding eigenvalues, λj . If X is orthogonal to En−k ,

then the variance is (XT X − XT Ek ETk X)−1 σ 2 . The standard ESF variance estimator is

σ̂ 2 = (Y − Xβ − Ek βE )T (Y − Xβ − Ek βE )/(n − K − 1), (8)

implying its unbiased estimator is {(n − K − 1)/[TR(V ) − (  −1
)]}σ̂ . The denominator
TR ETk V−1 Ek 2

ρ λ /( ρ λ ) 2 ρλj /(1 − ρλj )2 . In

n 2 2 2 n
of this expression may be rewritten as j =K +1 j 1 − − j j =K +1

j=K +1 ρ λ /( − ρ λ ) > 0, and because the

n 2 2 2
the presence of positive spatial autocorrelation, j 1 j
eigenvectors Ek are associated with positive eigenvalues denoting positive spatial autocorrelation,
j=K +1 2ρ λj /(1 − ρ λj ) < 0, with the magnitude of this negative number tending to increase with
n 2

increasing K .
Employing the unbiased variance estimators, the constant mean sampling variances become

OLS : (1T V−1 1/n2 )σ 2 (9)

GLS for the SAR model: (n/1 V1)σ = σ /(1 − ρ)
T 2 2 2
(10)
ESF: (X X − X
T T
) σ = σ /n.
Ek ETk X −1 2 2
(11)
The expectation, then, is that the ESF estimator is more efficient than its GLS estimator counterpart
(e.g., see Table 1), as well as its OLS estimator counterpart.
Fig. 1 summarizes results from a simulation experiment with a 30-by-30 square tessellation
illustrating this improved efficiency. This simulation assumed an SAR data generating process, and
independent and identically distributed normal random variables into which spatial autocorrelation
6 Y. Chun, D.A. Griffith / Spatial Statistics 10 (2014) 1–11

Table 2
Constant mean model specification results based upon Puerto Rico data (n = 73).
Box–Cox transformed Y = ρ CY + 1β0 + ε Y = ρ MCMY + 1β0 + ε Y = Eγ + 1β0 + ε
variable
ρ̂ Pseudo- VIF ρ̂ Pseudo- VIF # R2 VDF*
R2 R2 Vectors

Farms 0.1475 0.4314 2.1551 0.1384 0.3862 1.5959 11 0.6546 0.9974

Farm areas 0.1239 0.3151 1.4927 0.1200 0.2886 1.3414 9 0.5602 0.9200
Note: details for the Puerto Rico farm dataset can be found in Chun and Griffith (2013).
*
VDF denotes variance deflation factor.

a 1.0 b
0.8
efficiency

0.6

0.4

0.2

0.0
0.0 0.2 0.4 0.6 0.8 1.0

Fig. 1. Relative efficiency results for the mean: the ESF versus SAR mean estimate. Left (a): simulation results (asterisk)
superimposed on theoretical results (open circle). Right (b): ESF simulation results using different sized candidate eigenvector
sets.

was embedded with the spatial autoregressive operator (I − ρ W)−1 , varying ρ from 0 to 0.95 (see
Fig. 1); we generated 10,000 sets of pseudo-random numbers. The graphical part of Fig. 1 suggests that
TR(V−1 )− TR(ETk V−1 Ek ) ≈ n − K − 1. The tabular part of Fig. 1 suggests that efficiency does not improve
very much by increasing the number of eigenvectors available for constructing an ESF. In other words,
once the prominent eigenvectors are selected in an ESF construction, marginal vectors add little to the
efficiency of the ESF estimator. Table 1 reports corroborating empirical analysis covariate standard
error ratios.
Table 2 furnishes empirical examples illustrating this preceding contention. Ignoring spatial
autocorrelation inflates the variance with a loss of efficiency. The impact of removing the principal
eigenvector by replacing matrix C with matrix MCM, yielding the SAR specification Y = ρ MCMY +
Xβ + ε (see Griffith and Chun, 2014, pp. 1496–1497), is a decrease in the VIF attributable to spatial
autocorrelation. The VIF for the GLS estimator is 1; i.e., employing a spatial autoregressive model
specification regains efficiency. But employing an ESF model specification improves efficiency beyond
that gain by adopting an autoregressive model (see Table 2).

4.1. Spatial autocorrelation contributions to multicollinearity

Spatial autocorrelation impacts upon the pairwise bivariate correlation coefficients for a set of
georeferenced variables (Richardson and Hémon, 1981; Haining, 1991). Deconstructing each variable
into an ESF and a residual term reveals how these impacts materialize. Eigenvectors common to
constructed ESFs tend to inflate, whereas eigenvectors unique to each ESF (UESF) tend to deflate,
bivariate correlations. Especially the VIFs can affect efficiency.
Table 3 illustrates this situation with the georeferenced variables of wages, population, and
employment for each of the 439 German Kreis (roughly equivalent to US counties) in 2000. The
 Y. Chun, D.A. Griffith / Spatial Statistics 10 (2014) 1–11 7

Table 3
Correlation coefficient decomposition summary statistics for georeferenced random variables.
Component Transformed unemployment given Wages Population Employment
Wages Population Employment

R2 for common eigenvectors 0.6115 0.5833 0.59788 0.1208 0.1811 0.1205

R2 for unique eigenvectors 0.2214 0.2496 0.23497 0.0384 0.0709 0
  
R2 for all selected eigenvectors 0.8329 0.1592 0.2519 0.1205
Stepwise level of significance* 10% 5% 0.2% 7%
Residual MC 1.2 −0.1 −0.1 −0.4
Shapiro–Wilk (SW)** 0.9795 0.9168 0.9401 0.8484
ESF MC 1.1032 0.8574 0.8986 1.1428
*
Levels of significance were adjusted, in part to account for different candidate set sizes, so that the absolute value of the
residual MC is as close as possible to zero.
**
All SW statistics are significant at the 0.01 level.

Table 4
Bivariate correlations between unemployment and selected covariate components.
Covariate Wages Population Employment

reX ,eY : underlying attribute 0.0148 −0.4468 −0.1038

rX̂ ,Ŷ : spatial autocorrelation
C C
−0.8405 −0.8182 −0.9129
re ,Ŷ : cross correlation
X U
−0.0848 −0.1059 −0.1024
rX̂ ,e : cross correlation
U Y
0.0366 −0.0551 0
rX̂ ,Ŷ : unique, uncorrelated 0 0 0
U U
Observed −0.2565 −0.4757 −0.3314
Note: X̂C and ŶC are eigenvectors common to the X and Y variables, X̂U and ŶU respectively are unique eigenvectors for the X
and Y variables, and eX and eY are aspatial random errors.

wages ESF contains 9 eigenvectors, of which 5 are in common with the UESF. The employment ESF
contains 3 eigenvectors, all of which are in common with the UESF. And, the population ESF contains
18 eigenvectors, of which 9 are in common with the UESF.
Bivariate correlations for each of the components appear in Table 4. In all three cases, the spatial
autocorrelation component increases the observed correlation coefficient in absolute value. For
wages, spatial autocorrelation changes the nature of the underlying relationship. For employment,
spatial autocorrelation changes a marginally significant underlying relationship to a markedly
significant one. And, for population, spatial autocorrelation only modestly changes the degree of
the underlying relationship. These coefficients may be calculated by combining the information in
Tables 3 and 4. For example,
 
wages : − 0.2565 = 0.0148 (1 − 0.8329)(1 − 0.1592) − 0.8406 (0.6115)(0.1208)
 
− 0.0848 (1 − 0.1592)(0.2214) + 0.0366 (0.0384)(1 − 0.8329) + 0. (12)

In this case, spatial autocorrelation changes both the sign and the magnitude of the bivariate
correlation coefficient.

4.2. Covariate spatial autocorrelation impacts on VIFs

Variance inflation of standard errors, indexed with a traditional VIF, is one impact of
multicollinearity in linear regression analysis that affects the precision of estimated regression
coefficients, and hence efficiency. It is defined as VIFj = 1/(1 − Rj 2 ), where Rj 2 is the squared multiple
correlation coefficient obtained by regressing Xj on the set of remaining covariates. Moreover, VIFj
measures the linear dependence among the covariates on the variance of variable Xj . A complete lack
of multicollinearity results in Rj 2 = 0, and hence VIFj = 1. Applied statisticians consider VIFj values
between 5 and 10 to indicate problematic, and exceeding 10 to indicate serious, variance inflation,
and hence unstable regression estimates (Montgomery et al., 2001).
8 Y. Chun, D.A. Griffith / Spatial Statistics 10 (2014) 1–11

Table 5
Summary standard errors and VIFs for an illustrative subset of the selected geographic landscapes with 2 ESF residual (i.e.,
spatially unautocorrelated) covariates.
Y Estimator Residual Intercept X1 X2 (pseudo-) R2
MC se se VIF se VIF

Columbus, Crime OLS 0.5630 2.1382 0.6071 1.1541 0.1628 1.1541 0.2332
OH
GLS NA 4.9610 0.3738 NA 0.1007 NA 0.6292
ESF NA 1.2357 0.3538 1.1739 0.0982 1.2583 0.7717

Toronto Population OLS 0.6701 0.0221 0.0067 1.0262 0.0281 1.0262 0.0445
density
GLS NA 0.0637 0.0037 NA 0.0162 NA 0.6089
ESF NA 0.0116 0.0037 1.0996 0.0154 1.1215 0.7797
China Population OLS 0.7531 0.0213 0.1635 1.1935 0.1554 1.1935 0.0857
density
GLS NA 0.0873 0.0736 NA 0.0715 NA 0.7999
ESF NA 0.0108 0.0849 1.2536 0.0814 1.2764 0.7899
Note: se denotes standard error.
Note: X1 and X2 for the different geographic landscapes are: income and house price for Columbus, distance from the CBD and
areal unit size for Toronto, and the logarithm of birth and death percentages for China.

In the absence of covariates, the eigenvectors contained in an ESF constructed for response variable
Y have VIFs of 1 (the eigenvectors are orthogonal and uncorrelated). Introduction of covariates into a
model specification tends to result in VIFs > 1, depending upon the eigenvectors that constitute ESFs
for each of the individual covariates, and in part because of the degree of multicollinearity remaining
in the ESF residuals for these covariates. Table 5 summarizes autocorrelation-free covariate results for
a subset of the set of empirical landscapes studied in this paper.
R2 and pseudo-R2 (the squared correlation between Y and Ŷ ) results summarized in Table 5
indicate that much of the variance accounted for in the response variables is attributable to spatial
autocorrelation. A comparison of the residual MC values in Tables 1 and 5 indicates that much of
this spatial autocorrelation is shared by a response variable and its pair of covariates. Standard error
results reported in Table 5 indicate that GLS estimators tend to be more efficient than OLS estimators,
although not always (also see Cordy and Griffith, 1993), and that ESF estimators are far more efficient
than either OLS or GLS estimators.

5. Consistency in the presence of spatial autocorrelation

An extremely important quality feature of an estimator is that its sampling distribution becomes
increasingly more concentrated near its corresponding unbiased parameter value with increasing
n. Linear regression parameter estimators are consistent when the covariates are orthogonal and
uncorrelated with the error term in a model specification. But the context for this property to hold is
that n increases while the number of covariates remains unchanged. An ESF tends to have an increasing
number of eigenvectors with increasing n (as well as with increasing spatial autocorrelation—see Eq.
(1)). Based on a selected set of empirical data analyses – where n ranges from 49 to 2379 (see Table 1)
– the following logistic equation describes this increase extremely well (see Fig. 2)4 :

nselected ≈ ncandidate /(1 + 3.2508e−0.1995 z MC +0.0047 n ) (13)

where nselected is the number of selected eigenvectors for constructing an ESF, ncandidate is the number
of candidate eigenvectors for a given surface partitioning (e.g., those with a MCj /MC1 > 0.25, or
those identified by Eq. (1)), and zMC is the z-score for the response variable’s residual MC. Even when

4 This equation differs from that reported in Griffith and Chun (2014) because the analysis summarized here contains no
covariates.
 Y. Chun, D.A. Griffith / Spatial Statistics 10 (2014) 1–11 9

a b

50 100 150 200 250

0

0 50 100 150 200 250

Fig. 2. Left (a): scatterplot of the predicted (from Eq. (2)) and the observed number of selected vectors. Right (b): the number
of vectors.

ncandidate = n, MC = 1.5 (an excessively large value that is quite unlikely) and for a completely
connected planar graph (i.e., 1T C1 = 6(n − 2)), the asymptotic limit of Eq. (13) divided by n (i.e.,
nselected /n) is 0:
√
limit[n/(1 + 3.2508e−0.1995(1.5000+1/(n−1))/ 2/[6(n−2)]+0.0047n
)]/n = 0. (14)
n→∞

Because ncandidate < n, this limiting result implies that nselected described by Eq. (13), increases more
slowly than n, a finding that is in keeping with those reported by Portnoy (1984). In other words, the
covariate parameter estimates for an ESF model specification appear to be consistent. This finding
needs to be corroborated with a simulation experiment, and exercise constituting future research.

6. Implications

ESF-based linear regression estimators are unbiased at least for an SAR data generating process.
Empirical evidence presented in this paper supports the contention that they tend to be more efficient
than either their OLS or their GLS counterparts. This efficiency can be compromised by the presence
of spatial autocorrelation in covariates, which can cause variance inflation in the ESF estimators. Fig. 3
portrays relationships between VIFs and relative efficiencies. Intercept terms do not have VIFs. Fig. 3(a)
displays a scatterplot of intercept standard error ratios versus average VIFs for the covariates, and
reveals no relationship. In contrast, Fig. 3(b) and (c) display scatterplots for the covariate standard
error ratios versus their corresponding VIFs, and suggest that efficiency calculations may need to be
adjusted for VIFs arising from common spatial autocorrelation in order to be meaningful. VIFs are not
available for the SAR model specification, preventing a similar analysis for the GLS estimators. Finally,
the ESF estimators appear to be consistent.
The empirical-based pilot study summarized in this paper suggests a number of worthwhile
future research projects. First, the efficiency of ESF estimators needs to be established across
the possible sources of spatial autocorrelation in a linear regression model specification (Griffith,
1976). Second, the consistency of ESF estimators needs to be established with properly designed
simulation experiments. Third, relationships between VIFs and spatial autocorrelation need to be
articulated. Completion of such research will help build a sound mathematical statistical foundation
for eigenfunction spatial filtering methodology.
The investigation of ESF methodology can be extended further in future research in several
different ways, too. First, a comparison of ESF methodology with geostatistical approaches can
establish links between the two approaches. This comparison essentially is missing from the literature,
although Bellier et al. (2007) provide a comparison of an eigenvector based approach, named
principal coordinates of neighbor matrices, and geostatistics methods. Second, the ESF estimator is
investigated here mainly for an SAR data generating process, but further investigation for other data
generating process, including the AR and spatial Durbin specifications, is needed to provide a more
comprehensive understanding of eigenvector spatial filtering methodology.
10 Y. Chun, D.A. Griffith / Spatial Statistics 10 (2014) 1–11

a 0.9 b c 0.9
Ratio of standard errors

Ratio of standard errors

Ratio of standard errors

1.1
0.8
for the intercepts

1.0
0.8 0.9 0.7

for X_1

for X_2
0.8 0.6
0.7 0.7
0.6 0.5
0.6 0.5 0.4
0.4
0.3 0.3
0.5
0.2 0.2
0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0
Ratio of averageVIFs Ratio of VIFs Ratio of VIFs

Fig. 3. Scatterplots of the ratio of spatially autocorrelated to spatially unautocorrelated relative efficiency of ESF estimators vis-
a-vis OLS estimators versus their corresponding VIF ratios. Left (a): intercepts. Middle (b): 1st covariate regression coefficients.
Right (c): 2nd covariate regression coefficients.

Acknowledgments

Daniel A. Griffith is an Ashbel Smith Professor.

Research for this article was supported, in part, by the US National Science Foundation under the
Geography and Spatial Sciences and Methodology, Measurement, and Statistics Programs grant no.
BCS-1229223; any opinions, findings, and conclusions or recommendations expressed in this material
are those of the authors, and do not necessarily reflect the views of the National Science Foundation.

References

Bellier, E., Monestiez, P., Durbec, J.-P., Candau, J.-N., 2007. Identifying spatial relationships at multiple scales: principal
coordinates of neighbour matrices (PCNM) and geostatistical approaches. Ecography 30, 385–399.
Besag, J., 1974. Spatial interaction and the statistical analysis of lattice systems. J. Roy. Statist. Soc. Ser. B 36 (2), 192–225.
Bini, L.M., Diniz-Filho, J.A.F., Rangel, T.F.L.V.B., Akre, T.S.B., Albaladejo, R.G., Albuquerque, F.S., Aparicio, A., Araújo, M.B., Baselga,
A., Beck, J., Bellocq, M.I., Böhning-Gaese, K., Borges, P.A.V., Castro-Parga, I., Chey, V.K., Chown, S.L., Marco Jr., P.d., Dobkin,
D.S., Ferrer-Castán, D., Field, R., Filloy, J., Fleishman, E., Gómez, J.F., Hortal, J., Iverson, J.B., Kerr, J.T., Kissling, W.D., Kitching,
I.J., León-Cortés, J.L., Lobo, J.M., Montoya, D., Morales-Castilla, I., Moreno, J.C., Oberdorff, T., Olalla-Tárraga, M.Á., Pausas,
J.G., Qian, H., Rahbek, C., Rodríguez, M.Á., Rueda, M., Ruggiero, A., Sackmann, P., Sanders, N.J., Terribile, L.C., Vetaas, O.R.,
Hawkins, B.A., 2009. Coefficient shifts in geographical ecology: an empirical evaluation of spatial and non-spatial regression.
Ecography 32, 193–204.
Blanchet, F.G., Legendre, P., Borcard, D., 2008. Forward selection of explanatory spatial variables. Ecology 89, 2623–2632.
Borcard, D., Legendre, P., 2002. All-scale spatial analysis of ecological data by means of principal coordinates of neighbour
matrices. Ecol. Model. 153, 51–68.
Borcard, D., Legendre, P., Avois-Jacquet, C., Tuosimoto, H., 2004. Dissecting the spatial structure of ecological data at multiple
scales. Ecology 85, 1826–1832.
Chun, Y., Griffith, D.A., 2013. Spatial Statistics and Geostatistics. SAGE Publications, London.
Cordy, C., Griffith, D.A., 1993. Efficiency of least squares estimators in the presence of spatial autocorrelation. Commun. Stat.
B-Simul. 22, 1161–1179.
Dray, S., Legendre, P., Peres-Neto, P.R., 2006. Spatial modelling: a comprehensive framework for principal coordinate analysis
of neighbour matrices (PCNM). Ecol. Model. 196, 483–493.
Getis, A., Griffith, D.A., 2002. Comparative spatial filtering in regression analysis. Geogr. Anal. 34, 130–140.
Greene, W., 2008. Econometric Analysis, seventh ed. Prentice Hall, Upper Saddle River, NJ.
Griffith, D.A., 1976. Spatial autocorrelation problems: some preliminary sketches of a structural taxonomy. East Lakes Geogr.
11, 59–68.
Griffith, D.A., 1988. Advanced Spatial Statistics. Martinus Nijhof, Dordrecht.
Griffith, D.A., 1996. Spatial autocorrelation and eigenfunctions of the geographic weights matrix accompanying georeferenced
data. Can. Geogr. 40, 351–367.
Griffith, D.A., 1997. Using estimated missing spatial data in obtaining single facility location-allocation solutions. l’Espace Géogr.
26, 173–182.
Griffith, D.A., 2002. The geographic distribution of soil lead concentration: description and concerns. URISA J. 14, 5–15.
Griffith, D.A., 2003. Spatial Autocorrelation and Spatial Filtering: Gaining Understanding through Theory and Scientific
Visualization. Springer-Verlag, Berlin.
Griffith, D.A., 2004. A spatial filtering specification for the auto-logistic model. Environ. Plann. A 36, 1791–1811.
Griffith, D.A., 2006. Hidden negative spatial autocorrelation. J. Geogr. Syst. 8, 335–355.
Griffith, D.A., 2010. Spatial filtering. In: Getis, A., Fischer, M. (Eds.), Handbook of Applied Spatial Analysis. Springer-Verlag,
Berlin, pp. 301–318 (Chapter B.5).
Griffith, D.A., Chun, Y., 2009. Eigenvector selection with stepwise regression techniques to construct spatial filters. In: Paper
Presented at the 105th Annual Association of American Geographers Meeting, Las Vegas, NV, March 25.
 Y. Chun, D.A. Griffith / Spatial Statistics 10 (2014) 1–11 11

Griffith, D.A., Chun, Y., 2014. Spatial autocorrelation and eigenvector spatial filtering. In: Fischer, M., Nijkamp, P. (Eds.),
Handbook of Regional Science. Springer-Verlag, Berlin, pp. 1477–1507.
Griffith, D.A., Layne, L.J., 1999. A Casebook for Spatial Statistical Data Analysis: A Compilation of Analyses of Different Thematic
Data Sets. Oxford University Press, New York.
Haining, R., 1991. Bivariate correlation and spatial data. Geogr. Anal. 23, 210–227.
Hardouin, C., Yao, J.-F., 2007. Multi-parameter auto-models with applications to cooperative systems. C. R. Math. 345, 349–352.
Kramer, W., Donninger, C., 1987. Spatial autocorrelation among errors and the relative efficiency of OLS in the linear regression
model. J. Amer. Statist. Assoc. 82, 577–579.
Montgomery, D.C., Peck, E.A., Vining, G.G., 2001. Introduction to Linear Regression Analysis. John Wiley & Sons, New Jersey.
Pace, K., LeSage, J., Zhu, S., 2013. Interpretation and computation of estimates from regression models using spatial filtering.
Spat. Econ. Anal. 8, 352–369.
Peres-Neto, P.R., Legendre, P., 2010. Estimating and controlling for spatial structure in the study of ecological communities.
Glob. Ecol. Biogeogr. 19, 174–184.
Portnoy, S., 1984. Asymptotic behavior of M estimators of p regression parameters when p2/n is large: I. Consistency. Ann.
Statist. 12, 1298–1309.
Richardson, S., Hémon, D., 1981. On the variance of the sample correlation between two independent lattice processes. J. Appl.
Probab. 18, 943–948.
Tiefelsdorf, M., Boots, B., 1995. The exact distribution of Moran’s I. Environ. Plann. A 27, 985–999.
Tiefelsdorf, M., Griffith, D.A., 2007. Semi-parametric filtering of spatial autocorrelation: the eigenvector approach. Environ.
Plann. A 39, 1193–1221.