Regional Science and Urban Economics xxx (xxxx) xxx–xxx

Contents lists available at ScienceDirect

Regional Science and Urban Economics

journal homepage: www.elsevier.com/locate/regsciurbeco

A new method for dealing simultaneously with spatial autocorrelation and

spatial heterogeneity in regression models☆
⁎
Ghislain Geniaux , Davide Martinetti
UR Ecodevelopment, INRA, Avignon, France

A R T I C L E I N F O A BS T RAC T

Keywords: Although spatial heterogeneity and spatial dependence are two cornerstones of spatial econometrics, models
Local models and methods for dealing at the same time with both issues are still rare in the literature, with few notable
Geographically weighted regression exceptions. The same can be said for studies on the performance of spatial econometric models under
Mixed GWR misspecification of explanatory variables and unknown structure of the spatial weight matrix. In this article, we
SAR
introduce a new class of data generating processes (DGP), called MGWR-SAR, in which the regression
parameters and the spatial autocorrelation coefficient can vary over the space. For the estimation of these new
models, we resort to the Spatial Two-Stage Least Squares (S2SLS) technique. We rely on a Monte Carlo
experiment for testing the performance of classical models, such as OLS, GWR (Geographically Weighted
Regression), mixed GWR and SAR (Spatial AutoRegressive model), as well as our proposals, paying special
attention to simulated data under the realistic assumption that they suffer from multicollinearity/concurvity
problems and/or misspecification of the covariates. The results suggest that certain model specifications
amongst the newly proposed family MGWR-SAR are the more robust. Furthermore, to complete our proposal,
we also suggest a specification procedure to identify the correct spatial weight matrix for DGPs with spatial
heterogeneity and spatial autocorrelation of the endogenous. We conclude the article with an empirical study on
the Lucas County house price dataset, confirming the good performance of the proposed estimators.

1. Introduction metrics literature, and even scarcer when possible model misspecifica-
tion is considered. Tests to diagnose the join presence of spatial
Usual spatial-econometric estimation frameworks, based on models autocorrelation and spatial heterogeneity, or to test heteroskedasticity
with spatial autocorrelation and with a given spatial weight matrix are taking into account spatial autocorrelation, exist: JLM test of Anselin
sometimes unfeasible in the presence of model misspecification. In (1988), KR-SPHET test of Kelejian and Robinson (1998), spatial
fact, they are unable to disentangle between real spatial autocorrelation Breuch-Pagan test, spatial Chow test (Anselin, 1990b), LM test (Mur
and the different sources of violation of IID, such as non-linear et al., 2008). However, they are limited to specific forms of spatial
relationship of spatially-correlated independent variables, spatial het- variation of the parameters, such as spatial regimes that imply block
erogeneity through unobserved covariates and/or spatially varying heteroskedasticity and they are not suitable for more general forms of
relationships (see, amongst others, (Anselin, 1990a; Anselin and spatial heterogeneity of model parameters, i.e. when the spatial
Bera, 1998; Brunsdon et al., 1999)). Non-linearities and spatial variation of parameters is continuous (smooth) over the space and
heterogeneity can cause spatial dependence and the reverse is also depends on coordinates. Moreover, when there is high uncertainty on
true (see, e.g. (Fotheringham, 2009; Pace and LeSage, 2004))1. which spatial interaction matrix to choose or on the model specifica-
Studies that consider both spatial autocorrelation and spatial tion, like in the presence of non-linearities and/or unobserved spatial
heterogeneity are still scarce in spatial statistics and spatial econo- heterogeneity, the existing tests are insufficient for delivering a

☆
Davide Martinetti has received the support of the EU in the framework of the Marie-Curie FP7 COFUND People Programme, through the award of an AgreenSkills fellowship under
grant agreement 267196. The research reported in this work has been partially supported by projects URBANSIMUL and EPIDEC.
⁎
Corresponding author.
E-mail addresses: [email protected] (G. Geniaux), [email protected] (D. Martinetti).
1
In this text we use the term spatial autocorrelation to define the fact that the endogenous variable and/or the disturbances are not independent from their neighbouring: values at
locations nearer to each other being more or less similar than values at locations further apart. With the term spatially heterogeneous (or spatially varying, or non-stationary, or local) we
intend the fact that the value of a certain model parameter depends on its location on the space. Finally, we use the term spatially dependent or spatially correlated to refer to the fact that
the value of an explanatory variable depends on its location, for example as a function of its distance from a certain point in space.

http://dx.doi.org/10.1016/j.regsciurbeco.2017.04.001
Received 8 October 2016; Received in revised form 28 March 2017; Accepted 4 April 2017
0166-0462/ © 2017 Elsevier B.V. All rights reserved.

Please cite this article as: Geniaux, G., Regional Science and Urban Economics (2017), http://dx.doi.org/10.1016/j.regsciurbeco.2017.04.001
G. Geniaux, D. Martinetti Regional Science and Urban Economics xxx (xxxx) xxx–xxx

significant diagnosis of the presence of both spatial autocorrelation and without misspecification of covariate. In the absence of suitable tests
spatial heterogeneity of unknown form. for the joint presence of spatial heterogeneity of general form and
The recent work of Basile et al. (2014) is a nice attempt in this line spatial dependence, the aim is to asses which model would perform
to promote more flexible estimation frameworks that follow the better in case of unknown DGP and is more robust to model
recommendations of McMillen (2012) of using spatial smoother misspecification. The last experiment allows to assess the performance
technics in order to remove spatial heterogeneity while considering of the estimation technique for identifying the spatial weight matrix.
other potential non-linearities. The extension of geoadditve models Finally, in Section 4, we compare the performance of our proposed
with spatial autocorrelation proposed by Basile et al. (2014) allows to methodology against the classical SAR model on the Lucas County
consider spatially autocorelated error terms and/or endogenous vari- pricing data set by comparing their predictive accuracy using cross
able, and it provides a useful tool for spatial econometrics practitioners. validation techniques. Conclusions and future developments are
An alternative to geoadditve models for considering spatially- sketched in Section 5.
varying coefficients can be found amongst local regression models
(Fotheringham et al., 1999; Loader, 1999b; McMillen, 1989): they 2. Spatial econometrics models and estimators with spatial
constitute a simple framework for extending local models in order to dependence and spatial heterogeneity
account for spatial autocorrelation. For example, Brunsdon et al.
(1996) proposed a Spatial AutoRegressive model (a.k.a. SAR) esti- 2.1. Motivations for spatially varying coefficient models in urban
mated with local maximum likelihood in which all coefficients are economics
spatially varying. Cho et al. (2010) proposed an approach that
combines geographically weighted regression (GWR) and Spatial Spatial heterogeneity problems in regression models are, in our
Error Model (SEM), called GWR-SEM, using Generalized Method of point of view, inseparable from other issues such as non-linearities and
Moments (Kelejian and Prucha, 1998). SEM should allay spatial spatial autocorrelation. It is the case, for example, of the effects of land
autocorrelation, while GWR addresses spatial heterogeneity by allow- area on land price in hedonic price function estimation: empirical
ing the coefficients to vary across observations. Nevertheless, their evidence suggests that the relationship between the area and the price
model is over-specified, since it has a random intercept term in both of a plot of land is often non-linear, with threshold effects due to zoning
the GWR and the SEM part. In the same vein, Páez et al. (2002) regulation (wherever applicable). Furthermore, the spatial distribution
propose an estimation method in which the covariance is locally of land areas is usually linked to the distance to the city centre, the
varying and that can handle spatial autocorrelation of the error terms. closer to the city, the smaller the area, but other less-measurable
Another notable contribution accounting for both spatial autocorrela- variables can also have an influence on that distribution. Is then easy to
tion and non-stationarity of the regression parameters has been made prove that not accounting for non-linearities can introduce spatial
by Pace and LeSage (2004): they propose a spatial autoregressive local biases, under the form of spatial autocorrelation and/or unobserved
estimation based on a recursive approach for maximum-likelihood spatial heterogeneity Basile et al. (2014). On the contrary, models that
estimation of SAR that implies estimates on subsamples related to a explicitly consider spatial autocorrelation and spatial heterogeneity can
neighboring of each observation. help reducing the problems of non-linearities Le Gallo (2004). Several
The two weaknesses shared by these propositions are that they do authors have also proved that the misspecification of the spatial
not allow to consider the mixed case in which some parameters are autocorrelation (either on the form of the model or on the specification
constant over space and other are spatially varying and, secondly, that of the spatial weight matrix) can introduce spatial heterogeneity and
they are computationally intensive. Despite local regression provides a that, on the other hand, considering spatial heterogeneity can reduce
powerful tool to model spatial heterogeneity, it suffers some limitations the problems issued from spatial autocorrelation. It is then hard to
regarding multicollinearity Barcena et al. (2014); Páez et al. (2017/02); disentangle between spatial autocorrelation, spatial heterogeneity and
Wheeler and Tiefelsdorf (2005a) and extreme coefficients, including non-linearities within the calibration of the same model: whenever one
sign reversal Farber and Páez (2007). Particularly, the presence of of these aspects is ill-specified, it can lead to the creation, intensifica-
multicollinearity can lead to artificial spatial patterns of the coeffi- tion or sometimes the disappearance, of the effects of the others. Many
cients, so that other authors prefer to use extended bandwidth Barcena authors have discussed the relationship between these three subjects,
et al. (2014); Páez et al. (2017/02), to use ridge or lasso GWR Wheeler but, in practice, there is still no clear agreement and the vast majority
(2007, 2009) or a “mixed” GWR approach, with only a limited number of spatial econometric practitioners still resort to classical models that
of coefficients that can vary over space. Geniaux et al. (2011) proposed only consider spatial autocorrelation. We then highlight that the
a framework for a specific class of spatial econometrics models in which interest in regression models with spatial heterogeneity has not been
also the spatial autocorrelation coefficient is non-stationary and the sufficiently promoted and justified from the economic point of view, as
corresponding spatial autoregressive mixed GWR is estimated using well as the importance of considering spatial heterogeneity, spatial
local Two-Stage Least Square methods. We show in this paper that the autocorrelation and non-linearity jointly. The interest in using regres-
fact that some covariates could be spatially dependent is a key sion models with spatially-varying coefficients is then multiple. Firstly,
dimension of the problem and that “mixed” GWR approaches are a we can argue that models that only consider spatial autocorrelation are
way of solving the concurvity/multicollinearity problems caused by not capable of correcting all the problems related to non-observable
spatially dependent covariates. spatial heterogeneity. This has pushed several authors to consider a
The previous discussion highlights the need for more flexible non-stationary intercept term amongst the regression variables, for
models where both the regression and the spatial autocorrelation example, by means of a spline function of the space coordinates Wood
coefficients can either be constant or non-stationary. In Section 2 we (2011). Nonetheless, this same argument can be pushed even further to
provide an economic justification for our approach and we later consider a regression model with more spatially-varying coefficients.
introduce a set of local regression models that include all possible To keep the example of hedonic price function of land sales, it is
combinations of constant and spatially-varying coefficients. Each sometimes difficult to have access to detailed urban regulation; these
model is presented with its corresponding estimation technique and unobserved variables can sometimes introduce spatial heterogeneity
in Section 3 we present a Monte-Carlo simulation study. The first set of that can only be partially taken into account by SEM models or models
experiments allows to study the finite properties of the proposed with a spatially-varying intercept term. The resultant of these variables
estimation techniques, particularly when multi-collinearity/concurvity cannot be accounted as either a fixed or a random effect in reduced
problems are artificially introduced in the data. In the second set of form of micro-econometric models, but it should rather be understood
experiments, each models is estimated over all possible DGPs, with and as an indicator of a different economic behaviour, hence it demands a

2
G. Geniaux, D. Martinetti Regional Science and Urban Economics xxx (xxxx) xxx–xxx

differentiation of the parameters of agent preferences (indirect utility term λW , with a known interaction matrix W that specify the relation-
function) that drives bidding process. Similarly, Brady and Irwin ship between each observation i and j, is very practical because it
(2011) invoked this same argument to promote the use of models that simplifies the estimation and the interpretation of the model's para-
integrate spatial heterogeneity in order to better account for the meters. However, this hypothesis is very restrictive. In fact, there is no
variability of willingness to pay with respect to environmental ame- definitive indication for the best choice of W in the econometric
nities, according to socio-spatial segregation of individuals: they affirm literature Bhattacharjee and Jensen-Butler (2006); Harris et al.
that the propensity of householders for green or open spaces can vary (2011). Moreover, even with a proper specification of W, when the
according to their socio-professional category or their level of revenue interaction between neighbors is changing over space, in intensity and/
that, on the other hand, are often clustered in space. Likewise, the or in structure, a true W does not really exist and the estimates of Eq.
willingness to pay of individuals for parking lots can vary strongly over (1) could be biased. For these reasons, we aim to improve the
the space according, for example, to urban density, type of neighbor- modelization of Eq. (2), in order to provide better estimators for the
hood or proximity to public transportation. These two examples fit parameters and for being able to capture a larger variety of spatial
better in the logic of a model with spatial differentiation of the interaction patterns.
regressors rather than with spatial interaction. To reduce the complexity of estimating the set of λij, and to keep the
The choice of a model with spatial autocorrelation, justified merely interpretation as a diffusion process of the spatial autocorrelation, we
as a microeconomic model with agents interactions (see (Brueckner, propose to relax one of the main hypothesis generally adopted by
2006)), is just a way to include a spatial structure to the specification existing estimators of SAR models, i.e. the parameter λ and the linear
and to allow the evaluation of the marginal effects. Nonetheless, the regression coefficients β are constant over the coordinates space. On
same can be achieved by means of other statistical methods that can the other hand, in Section 2.3.4 we will also propose a methodology to
include spatially-varying coefficients (McMillen, 2004, 2010). Another improve the estimates by adapting the structure of the spatial weight
way of justifying this approach is to write explicitly the underlying matrix W.
micro-economic model in which some variables of the price function In our first proposition, the value of λ and β depends on the
depend on some unobserved spatial covariates. For example, Geniaux coordinates (ui , vi ) (λ = λ (ui , vi ) and β (ui , vi )), where (ui , vi ) represents
et al. (2011) have applied an inter-temporal utility model for land price the longitude and latitude of observation i. The parameters λ (ui , vi ) and
with non-stationary coefficients to evaluate the anticipation of land β (ui , vi ) are only required to be spatially smoothed (continuity of the
owners on the likelihood of future development of their lands: the level second derivative). The corresponding class of DGP can be written as:
of anticipation varied over the space according to stability/uncertainty Yi = λ (ui , vi ; h ) WY + βc Xc + βv (ui , vi ; h ) Xv + ϵi , (3)
of land regulation in the neighborhood which is not directly observable.
In our propositions, we also consider the possibility of a non- where h is a bandwidth parameter that allow to define the local
stationary spatial autocorrelation parameter. This possibility is based subsample around the coordinates of each points (ui , vi ) using a given
on its theoretical soundness, even if some economic behavior can be distance kernel, Xc represents kc explanatory variables with spatially
evoked to justify it: when the spatial weight matrix W is unknown and stationary coefficients and Xv contains kv explanatory variables with
spatial locations are irregularly distributed over space, the choice of a non-stationary coefficients.
neighboring scheme based only on distance or first nearest neighbours From Eq. (3) we can derive nine different DGPs as a combination of
can be tricky. Choosing one weighting scheme instead of the other can fixed and spatially-varying coefficients (λ , βc , βv ):
lead to a spatial interaction matrix that is too dense (resp. too y = βc Xc + ϵi (OLS)
dispersed) in the heterogeneous parts of the space, resulting in under y = βv (ui , vi ) Xv + ϵi (GWR)
or overestimation of the spatial parameter. There exist at least two
y = βc Xc + βv (ui , vi ) Xv + ϵi (MGWR)
possible solutions to mitigate the effect of the spatial weight matrix
y = λWy + βc Xc + ϵi (SAR)
misspecification with cross section data: either to adapt the structure of
y = λWy + βv (ui , vi ) Xv + ϵi (MGWR−SAR(0, 0, k ))
W in order to improve the quality of the fit ((Ertur and Koch, 2006;
Meen, 1996; Mur et al., 2008), see Section 2.3.4) keeping a stationary y = λWy + βc Xc + βv (ui , vi ) Xv + ϵi (MGWR−SAR(0, kc, k v ))
spatial parameter or to use of a non-stationary spatial autocorrelation y = λ (ui , vi ) Wy + βc Xc + ϵi (MGWR−SAR(1, k , 0))
parameter like in our proposition. This highlight the link between a y = λ (ui , vi ) Wy + βv (ui , vi ) Xv + ϵi (MGWR−SAR(1, 0, k ))
suitable choice of W when considering both spatial autocorrelation and y = λ (ui , vi ) Wy + βc Xc + βv (ui , vi ) Xv + ϵi (MGWR−SAR(1, kc, k v ))
spatial dependence, so we also propose in Section 2.3.4 a method for (4)
choosing W in such settings.
The first triplet of equations contains, respectively, a model with
only constant regression coefficients, only spatially-varying coefficients
2.2. DGPs with spatial autocorrelation and spatially varying or a combination of constant and spatially-varying coefficients. The
coefficients same pattern is used in the second and third triplets of equations, i.e.
the ones with constant or spatially-varying spatial autocorrelation
In spatial econometric literature, a regression model that considers coefficient, respectively.
spatial autocorrelation of the endogenous variable Y is formally written We introduced the term MGWR-SAR in order to identify the DGPs
as: of the second and third triplets of Eq. (4). To simplify the notation, we
Yi = λW Y + Xi β + ϵi , ∀ i ∈ {1, …, n}, (1) suppose that the number of regression parameters is always the same,
k, and that it can include both constant (kc) and spatially-varying (kv)
where Y is the n-vector of the continuous dependent variable, X is a
coefficients in such a way that k = kc + k v . The values in parenthesis of
matrix of k exogenous explanatory variables, ϵi is a IID error vector and
the MGWR-SAR(iλ, kc, k v ) notations denote:
W is a n × n spatial weight matrix. Assuming linear spatial dependence
of Yi with his neighbors simply implies that there exists a set of non-
iλ, the fact that λ is constant (0) or spatially varying (1);.
null λij such that:
k c, the number of constant βs (any integer between 0 and k);.
Yi = ∑ λij Yj + Xi β + ϵi ∀ i ∈ {1, …, n}, k v, the number of spatially varying βs (any integer between 0
j (2) and k).
The hypothesis that we can capture the effect of the λij by using a single

3
G. Geniaux, D. Martinetti Regional Science and Urban Economics xxx (xxxx) xxx–xxx

Model MGWR-SAR(1, 0, k ) corresponds to the mixed GWR model of 2.3.2. Geographically Weighted Regression (GWR) estimation
Brunsdon et al. (1998); Fotheringham et al. (1999); Páez et al. (2002), For each observation i ∈ {1, …, n} we deal with a different vector of
while the other four MGWR-SAR models in Eq. (4) are our original local coefficients β (ui , vi ). Consider the following weight matrix M, with
contribution. size n×n, such that mii ′ = K (dii ′, h ), for any i′ ∈ {1, …, n}, with K () a
kernel function based on the distance dii′ between coordinates (ui , vi )
and the (ui ′, vi ′) and a bandwidth h. Let define the n×n diagonal matrix
2.3. Estimators for models with spatial autocorrelation and spatial Mi that has on the diagonal the i-th row of M. Hence, the estimation of
heterogeneity the coefficients in each point i is given by

βl (ui , vi ) = (X ′Mi X )−1X ′Mi Y . (6)

Although most of the models involving local parameters are
unidentifiable because they suppose more parameters than observa-
tions, we found in the literature different ways to approximate these 2.3.3. Mixed GWR (MGWR) estimation
local coefficients by introducing conditions on local continuity and Geographically Weighted Regression can appear inadequate for
smoothness. The first approach in this field consisted in parametric many socioeconomic variables that have global effects and are inde-
models incorporating the spatial coordinates of observations as for pendent from individual location. Moreover, GWR is inappropriate for
example the “Trend Surface Analysis” method Ripley (2005) or local categorical variables, since spatially-varying coefficients asso-
“Variable Expansion” method proposed by Casetti (1972, 1997). ciated with such polytomous variables may have no meaning. For an
Nevertheless, while these parametric methods allowed to identify a adequate treatment of these problems, mixed models or MGWR were
spatial trend valid throughout the sample, they were unable to detect developed. This kind of model can be expressed as:
local variations Fotheringham et al. (1999). After that, the main kc kv
contributions considered semi-parametric methods like the “Locally yi = ∑ βj xij + ∑ βl (ui , vi ) xil + ϵi i = 1, 2, …, n .
Weighted Regression” (LWR) McMillen (1996) and the j =1 l =1 (7)
“Geographically Weighted Regression” (GWR) Brunsdon et al. (1996)
who provided a methodological framework to estimate local values of The intercept can be set amongst spatially varying variables or not. The
the parameters from local regressions. The idea was to estimate as MGWR was already introduced by Fotheringham et al. (1999) with a
many regressions as focal points (data coordinates or other sets of seven-steps estimation, requiring 2 + k v GWR estimations; somewhat
coordinates) by weighting the local regressions with respect to the intensive in terms of computation. Mei et al. (2004) proposed a two-
observations' distance to these focal points. The Mixed GWR intro- steps methodology, based on partial linear models developed in
duced by Fotheringham et al. (1999) and extended by Mei et al. (2004) Speckman (1988) and Bowman and Azzalini (1997). We will use this
and Geniaux et al. (2011) allows to combine spatially stationary and methodology that seems to be better adapted to empirical studies and
non-stationary coefficients. large samples. To reduce excessive notation and for inference purposes,
A notable contribution that takes into account the spatial auto- the estimation of the GWR in terms of a Hat matrix Hoaglin and
correlation in a local regression framework has been made by Pace and Welsch (1978) can be rewritten as:
LeSage (2004). They proposed a “Spatial Autoregressive Local ⎛ (Xv )1 [Xv ′ M1 Xv ]−1Xv ′ M1 ⎞
Estimation” based on a recursive approach for maximum-likelihood ⎜ ⎟
⎜ (Xv )2 [Xv ′ M2 Xv ]−1Xv ′ M2 ⎟
estimation of SAR, implying estimates on subsamples related to a Sv = ⎜ ⋮ ⎟,
neighboring of each observation. Despite the fact that they used local ⎜⎜ ⎟
(X ) [X M X ] Xv ′ Mn ⎟
−1
estimates of the spatial parameter, they did not really exploited that ⎝ v n v′ n v ⎠
possibility and only focused on the spatial variability of the parameter where Mi follows the same notation used in Section 2.3.2. It is then
β. Moreover, their proposition did not allow to consider the mixed- possible to define a two-steps estimation for βc and βv (ui , vi ) as follows:
GWR case and it was computationally very intensive.
Amongst the DGPs that implies both spatial autocorrelation and Step 1: l
βc = [Xc ′ (I − Sv )′(I − Sv ) Xc]−1Xc ′ (I − Sv )′(I − Sv ) Y .
spatial heterogeneity, the MGWR-SAR(1, 0, k ) has been studied in the Step 2: βv (ui , vi ) = [Xv ′ Mi ′ Xv ]−1Xv ′ Mi ′ (Y − Xc l
m βc ).
spatial econometric literature Fotheringham et al. (1999); Páez et al.
(2002) (known there as simply “mixed GWR”). The MGWR-
SAR(1, k , 0 ) model was already introduced in Brunsdon et al. (1998). 2.3.4. MGWR-SAR estimation
While these authors proposed local maximum likelihood estimation, if All MGWR-SAR models include a spatially-lagged dependent vari-
we want to rely only on linear regression to keep reasonable computa- able (either constant λWy , if iλ = 0 , or spatially varying λ (ui , vi ) Wy , if
tion time, the parameters of such DGP should be estimated using the iλ = 1) that introduces a source of endogeneity. The spatially-lagged
“Spatial Two-Stage Least Squares” (S2SLS) Anselin (1988); Kelejian dependent variable is usually correlated with the error term ϵ. In order
and Prucha (1998). We also explored another estimation technique, to get rid of the endogenous part, all MGWR-SAR models are estimated
notably the “Best Spatial Two-Stage Least Squares” (BS2SLS) of Lee by means of a Spatial Two-Stage Least Square (a.k.a. S2SLS, see
Lee (2003), that can be relevant here in case of misspecification of the amongst others (Anselin, 1988; Kelejian and Prucha, 1998, 1999;
true model DGP but the results were not convincing and are not Kelejian et al., 2004)) that uses the following set of instruments
presented in this version of the paper (although they are available upon H = [X , WX−1, W2X−1, W 3X−1], where X−1 denotes the matrix of covari-
request). ates without the column of ones that corresponds to the intercept term.
All the estimators and methods proposed in the following sections The advantages of using S2SLS are that it allows to formulate easily the
have been coded in C++ and embedded into R through the Rcpp solution of the parameters of the mixed version of the GWR, and also to
package Eddelbuettel et al. (2011). They are the core functions of a reduce computation time (particularly helpful for iterated computation
forthcoming R package, called gwrsar. of local regressions).
Kernel and Bandwidth selection. In this paper, the estimators that
imply local regression usually adopt a bisquare kernel, namely
2.3.1. OLS estimation ⎧ ⎛ ⎞2
It is the classical “Ordinary Least Square” estimation: ⎪ 15 di ′ i d
mi ′ i = K (di ′ i, h ) = ⎨ 16 ⎜⎝ h ⎟⎠
if hi′ i < 1,
,
⎪
βl = (X ′X )−1X ′Y . (5) ⎩0 otherwise

4
G. Geniaux, D. Martinetti Regional Science and Urban Economics xxx (xxxx) xxx–xxx

where di ′ i denotes the Euclidean distance between locations i and i′ and panel structure of the data, without requiring the pre-specification of W
h a bandwidth or an adaptive bisquare kernel (see Eq. (9)). This when N is small, relative to T (time). Finally, a third approach is to use
bandwidth is selected by minimizing the following Cross Validation non parametric approaches for measuring spatial correlation of a single
(CV) criteria: variable. For example, López et al. (2010) use the concept of symbolic
n entropy as a measure of spatial dependence, while other authors
hl = min ∑ ( y − yl (h)−i )2 , propose the use of Moran I (see (Ord and Getis, 1995)) or local
h (8)
i =1 statistics (see (Aldstadt and Getis, 2006)) to identify the most suitable
where yl (h )−i represents the vector of estimated yl (h ) as a function of W.
the bandwidth h, from which the i-th observation has been removed. Our proposal is to use a parametric approach that is at the frontier
The reason for removing the i-th observation is that otherwise the of the approaches discussed earlier. The estimation procedure uses a
minimum would converge with a null bandwidth h. The choice of the family of distance kernels with a single parameter hw (the bandwidth or
bandwidth is more critical than the choice of kernel (see (McMillen and the number of neighbors, according to the kernel). To identify the
Redfearn, 2010) for a recent discussion on the issue of bandwidth matrix W, we use a moment estimator that tries to minimize the
selection). Nevertheless, minimizing the Cross Validation criteria may residual sum of squares (RSS) of the model estimation with respect to
be unsuccessful when kernel is based on distance with data irregularly W (h w ). In this paper, we consider only three types of kernel: k-nearest
distributed over space. When the bandwidth becomes too small for neighbors with rectangular kernel, bisquare or adaptive bisquare. The
some local points, the continuity of cross validation values can not be search is performed for each kernel, in order to find the best value of
guaranteed and some singularity appears. Then, in order to avoid hw. The final choice correspond to the pair of kernel/bandwidth that
numerical problems due to the presence of unconnected observations assures the best RSS for the model.
or too small local subsamples, it could be useful to use an adaptive
bisquare kernel that allows to have at least 2k + 1 observations for each 3. Monte Carlo experiments
local regression as follows:
In this section we present three Monte Carlo experiments. The first
⎧ ⎛ ⎞2 set of experiments focuses on the finite properties of the proposed
⎪ 15 ⎜ di ′ i ⎟ if #{mi ′. > 0} > 2k + 1
estimation techniques, particularly when multi-collinearity/concurvity
⎪ 16 ⎝ h ⎠
⎪
mi ′ i = K (d i ′ i , h ) = ⎨ ⎛ problems are artificially introduced in the simulated data. In the
⎪ 15 ⎜ di ′ i ⎞⎟ otherwise, where h∼ is such that #{m > 0}
2
second set of experiments, each models is estimated over all possible
⎪ 16 ⎝ h∼ ⎠ i ′.
⎪ DGPs, with and without misspecification of covariates, to assess the
⎩ ≥ 2k + 1.
robustness of the estimators. The last experiment allows to assess the
(9)
performance of the estimation technique for finding the right specifica-
Another issue concerning bandwidth selection is related to models with tion of W.
spatially-varying λ estimated via S2SLS. For such models, the S2SLS
method could lead to an estimation of λ that is out of the domain 3.1. Monte Carlo settings
interval [−1, 1] for those bandwidths h that are too small. To avoid such
issue, we use a penalized CV criteria with an arbitrary large penalty for All data are simulated over a coordinate space contained in the unit
bandwidth values that lead to an out-of-support λ. The chosen kernel square [0, 1]2 with n=1000 observations. We consider four CBDs,
will be the one showing the smallest penalized CV criteria. positioned around the four points (0.25, 0.25), (0.25, 0.75), (0.75, 0.25)
In search of W . As suggested in Section 2.1, the issue of estimating and (0.75, 0.75) respectively, and a set of explanatory variables
λij presents several analogies with the estimation procedures that X = [X0 , X1, X2 , X3], where X0 is the intercept term, X1 ∼ 5 (4, 8) and
propose to adapt the structure of W by iteratively modifying wij to X3 ∼ 5 (1, 2). The covariate X2 is constructed as a function of the
improve the fit Bhattacharjee and Jensen-Butler (2006); Ertur and coordinates space (based on the distance from nearest CBD) plus a
Koch (2006); Jennrich (2001); Meen (1996); Mur et al. (2008). A zero-mean noise:
choice of W that corresponds to the true data generating process is then
X2 (ui , vi ) = dist(ui , vi , CBD) + ϵ. (10)
crucial, since otherwise we will observe dramatic changes of parameter
estimates and explanatory variables specification Bhattacharjee and Hence, the covariate X2 is spatially dependent. The regression coeffi-
Jensen-Butler (2006). Harris et al. (2011) identified three main cients β and the spatial autocorrelation parameter λ can either be
approaches for choosing W. First, it is common practice to compare constant or spatially varying, according to the chosen DGP. When they
pre-specified versions of W using “goodness of fit statistics”, like AIC, are constant, they will take the values β0 = 0 , β1 = 1, β2 = −1, β3 = 1
to choose the best specification of W. LeSage and Fischer (2008) use and λ = 0.4 . Otherwise, they will follow these spatial patterns (see
Bayesian technics for choosing between non-nested competing models, Fig. 1):
while Stakhovych and Bijmolt (2009) show through Monte Carlo
experiments that information criteria could also be a valid option for • β (u , v ) is mono-centric w.r.t. the origin of the coordinate space
0 i i
choosing W. However, such approaches provide a local maximum for (0, 0),
which substantial bias of regression estimates may still remain. A • β (u , v ) is mono-centric w.r.t. the center of the coordinate space
1 i i
second approach starts with an unspecified spatial weight matrix W (0.5, 0.5),
and try to fit it in such a way that it is consistent with observed patterns • β (u , v ) corresponds to the (−x, y)-plane,
i i
• β (u , v ) corresponds to the (x,y)-plane,
2
of spatial dependence. Conley (1999) proposes a method to estimate i i
• λ (u , v ) is mono-centric w.r.t. the North-Est CBD.
3
the spatial auto-covariance matrix involving imperfect measures of i i
economic distance. In the same vain, Pinkse et al. (2002) develop a
framework with uncertainty on distance measures and the possibility of The spatial weight matrix W (whenever applicable) used for
spatial non-stationarity. Mur et al. (2008) proposed the zoom estima- simulating data is based on the four first neighbors. The error term ϵ
tion based on local ML estimates that adapt the number of neighbors to is drawn from a normal distribution with zero mean and a level of
consider in W for each observation. When observations over time are variance that allows to achieve a proxy of signal-to-noise ratio equal
available, the original proposal of Meen (1996) has been extended by either to 0.95 or 0.75. We use the formula proposed by Kelley Pace
Beenstock and Felsenstein (2012); Bhattacharjee and Jensen-Butler et al. (2012) to estimate this signal-to-noise ratio in presence of spatial
(2006); Harris et al. (2011) to construct W by taking benefits of the dependence.

5
G. Geniaux, D. Martinetti Regional Science and Urban Economics xxx (xxxx) xxx–xxx

Fig. 1. Spatial pattern of spatially varying coefficients and X2 covariate.

When simulating data with spatial heterogeneity, our three Monte • a number of neighbors between {2, 4, 6, 10, 16}, if k-nearest neigh-
Carlo experiments consider always DGPs in which all coefficients are bors or adaptive bisquare kernel was chosen in the first step;
non-stationary. Said differently, we do not simulate mixed model that • a real value drawn from the interval [0.045, 0.075] with uniform law,
have both stationary and non-stationary coefficients and we only used otherwise3.
mixed model for estimation. The only exception to this rule is in the
first set of experiments, where some mixed GDP are simulated for A simulated data set for each DGP is generated with the spatial weight
exploring the source of bias related to concurvity/multicollinearity matrix W constructed according to the random choice of the kernel and
problems in GWR. In this experiment, to study the effects of introdu- bandwidth.
cing concurvity between β0 and β2, we simulate data with β0 and/or β2
stationary, and X2 with and without spatial dependence.
In the second experiment, each of the five considered estimators 3.2. Experiment 1: multicollinearity/concurvity problems
(OLS, SAR, GWR, MGWR-SAR(0, kc, k v ), MGWR-SAR(1, kc, k v )) is
repeatedly tested (1000 repetitions) over all the five corresponding Before testing our family of estimators (MGWR-SAR), we first focus
DGPs (OLS, SAR, GWR, MGWR-SAR(0, 0, k v ), MGWR-SAR(1, 0, k v )) on the potential sources of bias estimation when the data are analyzed
in order to assess their robustness2. Moreover, we consider the case via a standard GWR model. To do that, we simulate different datasets,
where X2 can be misspecified, in the sense that in the DGP it is where the introduction of a spatially dependent variable X2 leads to
constructed as in Eq. (10), while we suppose that its observed value is concurvity problems between the β0 (ui , vi ) (intercept) and the β2 (ui , vi )
as follows: parameters.
This problem has been already partially addressed for GWR
X2′ (ui , vi ) = (1 − ψ ) X2 (ui , vi ) + ψX2 (ui , vi ) α (i , CBD), (11) estimators through collinearity analysis of certain variables in local
samples Páez et al. (2017/02); Wheeler and Tiefelsdorf (2005b), but it
where ψ ∈ [0, 1] represents the amount of misspecification and makes the estimation of the parameters not reliable. For example,
α (i , CBD) is the angle of the ray joining observation i and the closest Wheeler (2007, 2009) proposed to use ridge or LASSO GWR, while
CBD point. The form of misspecification follows the suggestions of Brunsdon et al. (2012) and Barcena et al. (2014) used “extended
McMillen (2012) and McMillen and Soppelsa (2014). kernels” (whenever collinearity is detected, the kernel extends to
Finally, in the third experiment, we will test our algorithm to find consider a bigger local neighbourhood). On the other hand, Páez
the best spatial weight matrix W. Obviously, the experiment make et al. (2017/02) and Fotheringham and Oshan (2016) showed that
sense only for those DGPs where W is present, i.e. those of the MGWR- using sufficiently large local samples in the GWR could improve the
SAR family and SAR. We proceed by choosing randomly a kernel quality of the estimated parameters. Nevertheless, the finite-sample
function between k-nearest neighbors, bisquare and adaptive bisquare properties of estimators adopting these strategies have only been
kernel. Then a bandwidth value is also randomly selected amongst: studied on simulated data where all variables where generated

3
Note that we have chosen the interval for the bandwidth of continuous kernels in
2
The MGWR-SAR(1, kc, 0 ) DGP has not been included because it can be considered as such a way that it will result in a number of neighbours comparable to the case of discrete
a special case of the MGWR-SAR(1, kc, kv ). kernels.

6
G. Geniaux, D. Martinetti Regional Science and Urban Economics xxx (xxxx) xxx–xxx

Table 1 Table 3
Bias and RMSE of the regression coefficients estimated with GWR, for 4 DGPs (GWR Bias and RMSE of the regression coefficients on data simulated with a MGWR-SAR
with no spatial dependence of X2, GWR with spatial dependence of X2, MGWR with (0,0,4) DGP and estimated with a MGWR-SAR(0,0,4), a MGWR-SAR(0,1,3) where β0 or
spatial dependence of X2 and stationary β0 or β2). β2 are kept stationary and MGWR-SAR(0,2,2) where both β0 and β2 are stationary.

X2 IID X2 Spatially Autocorelated MGWR-SAR(0,0,4) MGWR-SAR(0, kc, kv )

DGP GWR GWR MGWR MGWR kc = {β0 , β2} k c = β0 k c = β2

β0 stat. β2 stat.
β0 (ui , vi ) BIAS 0.7009 0.1539 0.0810 −0.0046
β0 (ui , vi ) BIAS 0.0037 0.6945 0.6523 0.6467 RMSE 0.7046 0.1650 0.2176 0.1540
RMSE 0.5344 0.7048 0.6679 0.6592 β1 (ui , vi ) BIAS 0.0034 0.0028 0.0035 0.0060
β1 (ui , vi ) BIAS 0.0022 0.0028 0.0042 0.0043 RMSE 0.0064 0.0059 0.0063 0.0080
RMSE 0.0044 0.0047 0.0061 0.0055 β2 (ui , vi ) BIAS -0.6000 −0.1212 −0.0681 −0.0384
β2 (ui , vi ) BIAS −0.0000 −0.5993 −0.5674 −0.5587 RMSE 0.6075 0.1428 0.2326 0.1744
RMSE 0.0267 0.6087 0.5836 0.5713 β3 (ui , vi ) BIAS 0.0050 −0.0002 0.0007 0.0107
β3 (ui , vi ) BIAS 0.0039 0.0063 0.0037 0.0003 RMSE 0.0105 0.0123 0.0111 0.0154
RMSE 0.0164 0.0173 0.0157 0.0148

To conclude the first experiment we also consider the case of DGPs

independently from the space. We will show in this section that once with spatially varying coefficients and spatial autocorrelation of the
the spatial independence assumption fails to be verified, concurvity dependent variable, with the λ coefficient either fixed (MGWR-
problems become more puzzling, as already noticed by Fotheringham SAR(0,0,4)) or spatially varying (MGWR-SAR(1,0,4)). In Table 3 we
and Oshan (2016). This can be easily observed by looking at the first report the results of the estimation of the regression coefficients of
two columns of Table 1, where the introduction of spatial dependence different estimators: MGWR-SAR(0,0,4) with all spatially varying
of variable X2 (second column) causes the GWR estimates of β0 (ui , vi ) coefficients (as in the DGP), MGWR-SAR(0,1,3) with either β0 or β2
and β2 (ui , vi ) to be strongly biased. On the other hand, we observe no fixed and MGWR-SAR(0,2,2) with β0 or β2 fixed. Similarly, in Table 4,
deterioration of the quality of the estimation of β1 (ui , vi ) and β3 (ui , vi ). we report the result for the case of DGP and estimators with the
In order to identify the source of bias in the estimation of a GWR spatially varying coefficient of spatial autocorrelation λ (ui , vi ). We can
model, we also considered an alternative DGP with fixed intercept or observe that introduction of the spatial autoregressive parameter does
fixed β2 parameter. The behaviour of the GWR estimator did not not remove the bias on the parameters β0 (ui , vi ) and β2 (ui , vi ) as far as
changed ( β0 (ui , vi ) and β2 (ui , vi ) are still biased). Alternative DGPs have the estimators consider only spatially varying coefficients. Again, the
been tested, as well as a LASSO GWR estimation, but the results did solution seems to be using mixed models. In particular, for the DGP
not improved (results are not reported and are available upon request). with stationary λ (MGWR-SAR(0,0,4)), the ratio bias/RMSE suggests
The take-home message of this first experience is definitively that the that fixing the coefficient β0 can be sufficient, while for the DGP with
source of bias of β0 (ui , vi ) and β2 (ui , vi ) comes exclusively from the fact non-stationary λ, the best estimations are obtained through a mixed
that X2 is spatially dependent. This is particularly interesting, if we model with both β0 and β2 fixed.
consider that the vast majority of data that practitioners are faced to,
actually include independent variables that are strongly correlated with
3.3. Experiment 2: robustness of MGWR-SAR estimators
their localisation.
In order to correct for the bias introduced by the spatially
The first experiment taught us that it is preferable to use mixed models
dependent variable X2, we test the performance of mixed models
whenever we detect that certain independent variables are spatially
(MGWR) with different combinations of fixed and spatially varying
dependent, and that the coefficients corresponding to those variables and
coefficients (again on data generated with a GWR DGP and spatially
the intercept should be kept stationary. In empirical studies, we suggest to
dependent X2). We immediately observe in Table 2 that the introduc-
test spatial autocorrelation of all covariates in order to detect potentially
tion of fixed coefficients reduce the bias relative to the coefficients β0
problematic variables. It is worth noting that the problems of identification
and β2. Notably, fixing β2 reduces by one half the bias, while fixing the
of the parameters for the GWR, MGWR-SAR(0, 0, k v ) and MGWR-
intercept or the intercept and β2, leads to more significant improve-
SAR(1, 0, k v ) models makes complicated the use of asymptotic tests
ments. In some sense, fixing these coefficients coincides conceptually
Leung et al. (2000); Páez et al. (2002); Wei and Qi (2012) for the
with the idea of using extended kernels Brunsdon et al. (2012); Barcena
identification of the true DGP, even in case of applying a correction
et al. (2014): if we take the kernel extension to the extremes, we enlarge
adapted to the critical thresholds as a function of the d.o.f. Páez et al.
the local sample up to cover the entire space. Nonetheless, in contrast
(2002). Alternatively, bootstrap tests like the one proposed by Mei et al.
with the proposition of these authors, we only extend the parameters
(2016) can contribute to the identification of the stationary coefficients.
that correspond to the problematic variables.
Table 4
Table 2 Bias and RMSE of the regression coefficients on data simulated with a MGWR-SAR
Bias and RMSE of the regression coefficients on data simulated with a GWR DGP and (1,0,4) DGP and estimated with a MGWR-SAR(0,0,4), a MGWR-SAR(1,1,3) where β0 or
estimated with a GWR and a MGWR where β0, β2 or both are kept stationary. β2 are kept stationary and MGWR-SAR(1,2,2) where both β0 and β2 are stationary.

GWR MGWR MGWR-SAR(1,0,4) MGWR-SAR(1, kc, kv )

β0 and β2 stat. β0 stat. β2 stat. kc = {β0 , β2} k c = β0 k c = β2

BIAS 0.6631 0.1090 0.3344 0.1953 β0 (ui , vi ) BIAS 0.8996 0.1182 0.5476 0.2022
β0 (ui , vi )
RMSE 0.6816 0.1460 0.3656 0.2289 RMSE 0.9086 0.1672 0.5699 0.2336
BIAS 0.0044 0.0037 0.0039 0.0053 β0 (ui , vi ) BIAS 0.0079 0.0109 0.0087 0.0101
β0 (ui , vi )
RMSE 0.0068 0.0059 0.0063 0.0073 RMSE 0.0095 0.0119 0.0105 0.0114
BIAS −0.5843 −0.1300 −0.3559 −0.2547 β0 (ui , vi ) BIAS -0.7500 −0.1225 −0.4005 −0.2321
β0 (ui , vi )
RMSE 0.6011 0.1674 0.3876 0.2851 RMSE 0.7599 0.1846 0.4318 0.2680
BIAS 0.0019 0.0030 0.0019 0.0065 β0 (ui , vi ) BIAS −0.0051 0.0085 −0.0070 0.0045
β0 (ui , vi )
RMSE 0.0173 0.0164 0.0173 0.0185 RMSE 0.0175 0.0181 0.0185 0.0173

7
G. Geniaux, D. Martinetti Regional Science and Urban Economics xxx (xxxx) xxx–xxx

Table 5
Mean Bias and RMSE of the marginal effects of β0. Rows correspond to estimators and columns to DGPs.

Table 6
Mean Bias and RMSE of the marginal effects of β1. Rows correspond to estimators and columns to DGPs.

Table 7
Mean Bias and RMSE of the marginal effects of β2. Rows correspond to estimators and columns to DGPs.

In the second experiment we focus on the robustness of the parameters and we rather turn to the bias of marginal spatial effects.
different estimators w.r.t. unknown DGP. Such experiment is indis- For example, if the true DGP is a SAR model and we use a OLS for the
pensable, since there exist no test in the literature that is capable of estimation, we are more interested in the ability of the OLS of
detecting simultaneously spatially heterogeneity and spatial depen- approximating (I − λW )−1β , instead that a comparison against the
dence, hence there is no way of choosing the model specification based initial parameters that does not account for the predictive capability
on solid assumptions. Since we plan to cross different DGP and of the estimator. In the following tables we then measure the mean bias
estimators, there is no point in looking at the bias of the estimators and RMSE of the spatial marginal effects (I − λ (ui , vi ) W )−1β (ui , vi ) on

8
G. Geniaux, D. Martinetti Regional Science and Urban Economics xxx (xxxx) xxx–xxx

Table 8
Mean Bias and RMSE of the marginal effects of β3. Rows correspond to estimators and columns to DGPs.

Table 9
Mean Bias and RMSE of the marginal effects of β2. The variable X2 is misspecified with ψ = 0.4 . Rows correspond to estimators and columns to DGPs.

the set of combinations of DGPs and estimators. (MGWR-SAR(0,2,2) and MGWR-SAR(1,2,2)) preferable. Whenever
From Tables 5–8, in which the darker shades of grey indicates the the expected adjustment of the model is poor, due to an increase of σ
smallest BIAS/RMSE and the lighter shades of grey indicates the and/or of the misspecification on the variable X2, we observe that the
second best, we can observe that: MGWR-SAR(0,2,2) and MGWR-SAR(1,2,2) estimators are no longer
the second best choice. Instead, it appears more important the ability
• the most severe bias are again on the parameters β0 and β2; to test if spatial heterogeneity and or spatial autocorrelation are
• OLS estimates are always biased as far as spatial heterogeneity or present. Nonetheless, both MGWR-SAR(0,2,2) and MGWR-
spatial autocorrelation are introduced in the DGP; SAR(1,2,2) remain at a reasonable distance from the best performing
• MGWR estimates are always biased as far as spatial autocorrelation model and still insure to avoid too high bias. The take-home message
are introduced in the DGP; from the second experiment is then the following: the more robust
• the MGWR-SAR(0,2,2) estimator with stationary β0 and β2 seems to estimators in case of lack of appropriate tests for model specification
be the more robust, since it is always the first or second best that encompass all DGPs considered here, are MGWR-SAR(0, kc, k v )
performing, independently on the true DGP, and it allows to avoid and MGWR-SAR(1, kc, k v ) (Table 10).
too high biases in the presence of spatial heterogeneity and/or
spatial autocorrelation. 3.4. Experiment 3: searching W

If we introduce misspecification on the variable X2 (Table 9), we In the third experiment we test the performance of the identifica-
observe that both bias and RMSE tend to increase, while the hierarchy tion procedure for the spatial weight matrix W detailed earlier in
of best estimators remains the same, except for the case of SAR DGP Section 2.3.4. For the three DGPs with spatial autocorrelation used in
where the estimation of λ is biased and it makes models without spatial the second experiment, we compute the percentage of correct identi-
autocorrelation (OLS and MGWR) preferable. This is in line with what fication according to the following criteria:
observed in Le Gallo and Fingleton (2012), where OLS estimators
prove better performance than standard techniques, in the context of • the kernel is correctly identified and
spatial models with measurement errors of the independent variable. • the estimated bandwidth is less than 5% apart from the true
Similar conclusion can be reached for the case of an increase of the bandwidth used for generating the true W.
signal-to-noise ratio (see (Kelley Pace et al., 2012), Eq. (62)) that
makes models with spatial autocorrelation and spatial heterogeneity In Table 11 we observe that in general our procedure is highly reliable:

9
G. Geniaux, D. Martinetti Regional Science and Urban Economics xxx (xxxx) xxx–xxx

Table 10
Mean Bias and RMSE of the marginal effects of β2. The signal-to-noise ratio is increased by generating errors with higher variance σ = 3. Rows correspond to estimators and columns to
DGPs.

chosen by our identification algorithm is constructed with an adaptive

Table 11
Percentage of correct identification of W, 1000 replications. kernel with 12 nearest neighbors (see Eq. (9)), where closest observa-
tions are assigned higher weights. Actually, we can observe that the
DGP SAR MGWR-SAR(1, 0, kv ) MGWR-SAR(0, 0, kv ) first three neighbours (out of 12) of each observation in Wopt absorb
almost 40% of the total weight for that observation, meaning that Wsoi
σ=1 96.54% 94.55% 87.75%
σ=3 79.75% 67.50% 76.50% and Wopt are, in principle, comparable.
By comparing the results of the SAR estimation with Wsoi and Wopt,
we notice that the estimation with Wopt is always better (in terms of both
for simulated data with a signal-to-noise ratio of 95% (σ = 1), the RSS and AIC) than the one with (Wsoi)6. In conclusion, we can say that
degree of right identification is around 90%, while it decreases to 70% the predictive power of the two methods is comparable, but that our
for simulated data with higher variance (σ = 3, SNR=75%). approach can be considered as an “off-the-shelf” solution that does not
requires to make assumptions on the true form of the spatial weight.
4. Empirical study on Lucas county pricing data For the MGWR-SAR(0, kc, k v ) estimator, the first issue is choosing
the parameters that will be spatially stationary. The variable ’‘baths” is
We conclude the experimentations of the proposed MGWR- almost an obligate choice, since discrete variables are very likely to
SAR(0, kc, k v ) method and the W-identification procedure by testing create problems in local samples. Since all the other continuous
their performance on a real data set. We consider the well-known Lucas variables are spatially dependent7, we chose to test the out-of-sample
county house-price dataset4, in line with what has been done by other predictive power for all combinations of spatially stationary and non-
authors such as LeSage and Pace (2009); Bivand (2010) and Basile stationary. Based on this criteria, we finally retained the intercept, the
et al. (2014), when treating with comparable spatial issues. number of bathrooms and the (log of the) lot size as spatially stationary
We plan to test the following models: variables, while age and (log of the) total living area are spatially non-
stationary.
• SAR model with Wsoi as in Basile et al. (2014); The performance of the MGWR-SAR(0, kc, k v ) model improves
• SAR model with Wopt issued from our identification procedure; considerably that of the SAR specification, in terms of RSS, AIC and
• MGWR-SAR(0, kc, k v ) with Wopt issued from our identification PMSE. Nonetheless, the remarkable drop of the RSS of the MGWR-
procedure. SAR(0, kc, k v ) model should be put into perspective, accounting for the
fact that this model has different degrees of freedom. On the other
In Table 12, we present the Residual Sum of Squares (RSS), the Akaike hand, both AIC, PMSE10 and PMSE20 criteria allow a direct compar-
Information Criterion (AIC) and the Predicted Mean Square Error for ison of the MGWR-SAR(0, kc, k v ) against the SAR model: we then
two train/test splitting strategies (PMSE10 for 10-fold cross validation observe that the decrease is substantial and favors the mixed model.
and PMSE20 for 20-fold cross validation, see Shao (1993)). For the Results for PMSE10 and PMSE20 are similar and highlight the good
latter, PMSE10 and PMSE20, the predictions on the out-of-sample data performance of MGWR-SAR(0, kc, k v ) model: the PMSE10 decreases
have been computed by means of the Best Linear Unbiased Predictor between 10% and 18% according to the year. We think that a 15%
(BLUP)5, that ensure a better performance than using directly the mean improvement in the predictive power of a model that has already
parameters estimated on the in-sample data. The first observation been the subject of several researches is undoubtedly an encouraging
concerns the choice of the spatial weight matrix W: the matrix Wsoi is result and proves once more the advantage of having more flexible
computed from a sphere-of-influence (soi) graph constructed from a models: in empirical situations, where the true DGP and W are
triangulation of the point coordinates of the houses after projection to unknown, the possibility of adjusting the structure of the spatial weight
the Ohio North NAD83 Lambert Conformal Conical specification. The matrix and of choosing amongst different combination of stationary
resulting matrix is relatively sparse, with less than three neighbors per and non-stationary parameters, can help reducing the unsolvable
observation Basile et al. (2014). On the other hand, the matrix Wopt identification problem of empirical spatial data.

6
The comparison in terms of PMSE does not highlight any statistical difference
4
See LeSage and Pace (2009) for a detailed explanation of the dataset. between the two approaches.
5 7
We adapted the BLUP to our estimators via the Golberger formula, see Kelejian and We tested spatial dependency by means of the Moran's I test and all variables
Prucha (2007) and Thomas-Agnan et al. (2013). resulted dependent from their location.

10
G. Geniaux, D. Martinetti Regional Science and Urban Economics xxx (xxxx) xxx–xxx

Table 12
Comparison of SAR and MGWR-SAR(0, kc, kv ) estimation for Lucas county house price dataset, 100 repeated cross validation (k-fold).

year 1995 1996

Est. W RSS AIC PMSE10 PMSE20 RSS AIC PMSE10 PMSE20

SAR Wsoi 152.002 7.020 0.193 0.197 224.616 7.411 0.218 0.223
SAR Wopt 150.520 7.011 0.195 0.198 216.116 7.373 0.226 0.234
MGWR-SAR(0, kc, kv ) Wopt 95.596 5.143 0.175 0.177 128.032 5.502 0.197 0.202

year 1997 1998

Est. W RSS AIC PMSE10 PMSE20 RSS AIC PMSE10 PMSE20
SAR Wsoi 269.684 7.594 0.221 0.219 220.117 7.391 0.264 0.294
SAR Wopt 245.287 7.500 0.221 0.218 217.397 7.379 0.265 0.290
MGWR-SAR(0, kc, kv ) Wopt 140.606 5.571 0.194 0.193 121.536 5.436 0.216 0.231

5. Conclusions and future works we have been able to slightly improve the results with respect to the
SAR estimation with the original W matrix.
The goal of this paper was to study the problem of spatial regression In order to improve the performance of our model specification
models with spatial autocorrelation of the dependent variable and procedure, that is burdened by an high number of iterations over
spatial heterogeneity of the parameters, a subject that has not received several different DGPs and possible specifications of the spatial
enough attention in the literature, but that it is crucial for spatial structure, we see the urgent need for asymptotic tests for:
econometric practitioners. In Section 2.2 we recalled existing techni-
ques to deal with either spatial autocorrelation of the dependent • non-stationarity of the spatial autocorrelation parameter (in our
variable (SAR model) or spatial heterogeneity of the regression notation, MGWR-SAR(0, ·,· ) vs. MGWR-SAR(1, ·,· ));
parameters (GWR model if all coefficients vary over the space or • detecting spatial autocorrelation for mixed models (MGWR vs
MGWR if some coefficients are stationary and others are not). We also MGWR-SAR(0, kc, k v ));
proposed four new models that simultaneously accounts for both • identifying mixed vs non-mixed models in the presence of spatial
autocorrelation of the endogenous variable and spatial heterogeneity autocorrelation (MGWR-SAR(0, 0, k ) vs. MGWR-SAR(0, kc, k v )).
of the parameters.
These models are more flexible, since they relax the hypothesis on In our contribution, scarce attention has been paid to the estima-
the nature of the spatial interactions. The results of a series of Monte tion time required by the different methods. Nevertheless, in the future
Carlo experiments show that they are also more robust to certain types we will look for improved estimators, capable of handling large samples
of misspecification of the model. In particular, it appeared that mixed in reasonable time. For this purpose, we plan to use the target-point
models with a combination of stationary and non-stationary regression technique for the family of MGWR-SAR models as already proposed by
coefficients are more reliable, since they reduce identification pro- Loader (1999a) for regression models and by McMillen and Redfearn
blems. In fact, MGWR-SAR(0, kc, k v ) and MGWR-SAR(1, kc, k v ) models (2010); McMillen (2012) for the GWR model. Other proposals for
are often amongst the best or the second best, especially when faced to improving the computational efficiency of the estimators are to use the
the presence of autocorrelation of the endogenous variable and spatial Estimated Generalized Least Squares (EGLS) and Maximum
heterogeneity of the regression coefficients. Likelihood methods proposed by Pace and Barry (1997a), provided
Another issue that arose during the Monte Carlo experiments is that that we can extend these techniques to the MGWR-SAR family of
the presence of spatially dependent covariates have a strong impact on models. Finally, the exponential matrix model proposed by LeSage and
the performance of certain models. In fact, we included in the DGPs a Pace (2007) can also be used to reduce computational requirements:
variable which value is dependent on its location and that caused this method implies the transformation of the spatial weight matrix
different models, especially those with non-stationary coefficients, to inverse by means of the so called matrix exponential. This allow to
underperform w.r.t. the case of non-spatially dependent covariates. avoid the computation of log-determinant of large matrices.
Although there exist tests for the detection of spatial dependence, we
found that specification tests in the literature never consider the case of References
a DGP with spatially dependent regressors. In view of the results of our
experiment, we foster the study of appropriate tests that also consider Aldstadt, J., Getis, A., 2006. Using amoeba to create a spatial weights matrix and identify
spatially dependent covariates, while the use of bootstrap tests can spatial clusters. Geogr. Anal. 38 (4), 327–343.
Anselin, L., 1988. Spatial econometrics: Methods and Models. Wiley Online Library.
provide a first solution, whenever the sample size and the number Anselin, L., 1990a. Some robust approaches to testing and estimation in spatial
covariates remain treatable. econometrics. Reg. Sci. Urban Econ. 20, 141–163.
Regarding the choice of stationary and non-stationary coefficients, Anselin, L., 1990b. Spatial dependence and spatial structural instability in applied
regression analysis. J. Reg. Sci. 30 (2), 185–207.
it seems that good performances are achieved by leaving only the Anselin, L., Bera, A.K., 1998. Spatial dependence in linear regression models with an
intercept term amongst the stationary coefficients. Further improve- introduction to spatial econometrics. Stat. Textb. Monogr. 155, 237–290.
ments can be obtained by adding some of those coefficients that Barcena, M., Menendez, P., Palacios, M., Tusell, F., 2014. Alleviating the effect of
collinearity in geographically weighted regression. J. Geogr. Syst. 16 (4), 441–466.
correspond to spatially dependent covariates to the set of stationary Basile, R., Durbán, M., Mínguez, R., Montero, J.M., Mur, J., 2014. Modeling regional
coefficients. Indeed, in empirical settings, we are likely to find a large economic dynamics: spatial dependence, spatial heterogeneity and nonlinearities. J.
number of strongly spatially-dependent regressors, and some of them Econ. Dyn. Control 48, 229–245.
Beenstock, M., Felsenstein, D., 2012. Nonparametric estimation of the spatial
can generate more or less concurvity problems than others: as rule of
connectivity matrix using spatial panel data. Geogr. Anal. 44 (4), 386–397.
thumb, our results suggest that out-of-sample cross-validation proce- Bhattacharjee, A., Jensen-Butler, C., 2006. Estimation of spatial weights matrix in a
dures should be prioritized for choosing the set of spatially varying spatial error model, with an application to diffusion in housing demand. University
coefficients. of St Andrews. Department of Economics.
Bivand, R., 2010. Computing the Jacobian in Spatial Models: an Applied Survey. Journal
We also tested a specification procedure for the identification of the of Statistical Software VV (II).
spatial weight matrix W. The results on the accuracy of the identifica- Bowman, A.W., Azzalini, A., 1997. Applied Smoothing Techniques for Data Analysis: The
tion procedure on simulated data are very encouraging. Furthermore Kernel Approach with S-Plus Illustrations 18. OUP, Oxford.
Brady, M., Irwin, E., 2011. Accounting for spatial effects in economic models of land use:
when we applied this strategy to the Lucas dataset using a SAR model,

11
G. Geniaux, D. Martinetti Regional Science and Urban Economics xxx (xxxx) xxx–xxx

recent developments and challenges ahead. Environ. Resour. Econ. 48 (3), 487–509. LeSage, J.P., Pace, R.K., 2009. Introduction to Spatial Econometrics (Statistics,
http://dx.doi.org/10.1007/s10640-010-9446-6. Textbooks and Monographs). CRC Press.
Brueckner, J.K., 2006. Strategic interaction among governments. Companion Urban Leung, Y., Mei, C., Zhang, W., 2000. Testing for spatial autocorrelation among the
Econ., 332–347. residuals of the geographically weighted regression. Environ. Plan. A 32, 871–890.
Brunsdon, C., Charlton, M., Harris, P., 2012. Living with collinearity in local regression Loader, C., 1999a. Local Regression and Likelihood 47. springer, New York.
models. Spatial Accuracy Conference Brazil, 2012. Loader, C.R., 1999b. Bandwidth selection: classical or plug-in? Ann. Stat., 415–438.
Brunsdon, C., Fotheringham, A., Charlton, M., 1996. Geographically weighted López, F., Matilla-García, M., Mur, J., Marín, M., 2010. A non-parametric spatial
regression: a method for exploring spatial nonstationarity. Geogr. Anal. 28 (4), independence test using symbolic entropy. Reg. Sci. Urban Econ. 40 (2–3), 106–115.
281–298. McMillen, D., 1996. One hundred fifty years of land values in Chicago: a nonparametric
Brunsdon, C., Fotheringham, A.S., Charlton, M., 1999. Some notes on parametric approach. J. Urban Econ. 40 (1), 100–124.
significance tests for geographically weighted regression. J. Reg. Sci. 39 (3), McMillen, D., 2004. Employment densities, spatial autocorrelation, and subcenters in
497–524. large metropolitan areas. J. Reg. Sci. 44 (2), 225–244.
Brunsdon, C., Fotheringham, S., Charlton, M., 1998. Spatial nonstationarity and McMillen, D., 2010. Issues in spatial data analysis. J. Reg. Sci. 50 (1), 119–141.
autoregressive models. Environ. Plan. A 30 (6), 957–973. McMillen, D., 2012. Perspectives on spatial econometrics: linear smoothing with
Casetti, E., 1972. Generating models by the expansion method: applications to structured models. J. Reg. Sci. 52 (2), 192–209.
geographical research*. Geogr. Anal. 4 (1), 81–91. McMillen, D., Redfearn, C., 2010. Estimation and hypothesis testing for nonparametric
Casetti, E., 1997. The expansion method, mathematical modeling, and spatial hedonic house price functions. J. Reg. Sci. 50 (3), 712–733.
econometrics. Int. Reg. Sci. Rev. 20 (1–2), 9–33. McMillen, D., Soppelsa, M., 2014. A conditionally parametric probit model of microdata
Cho, S.-H., Lambert, D., Roberts, R., Kim, S., 2010. Moderating urban sprawl: Is there a land use in chicago. J. Reg. Sci. 55 (3), 391–415.
balance between shared open space and housing parcel size? J. Econ. Geogr. 10 (5), McMillen, D.P., 1989. An empirical model of urban fringe land use. Land Econ. 65 (2),
763–783. 138–145.
Conley, T., 1999. Gmm estimation with cross sectional dependence. J. Econ. 92 (1), Meen, G., 1996. Spatial aggregation, spatial dependence and predictability in the uk
1–45. housing market. Hous. Stud. 11 (3), 345–372.
Eddelbuettel, D., François, R., Allaire, J., Chambers, J., Bates, D., Ushey, K., 2011. Rcpp: Mei, C., He, S., Fang, K., 2004. A note on the mixed geographically weighted regression
seamless r and c++ integration. J. Stat. Softw. 40 (8), 1–18. model. J. Reg. Sci. 44, 143–157.
Ertur, C., Koch, W., 2006. Regional disparities in the european union and the Mei, C.-L., Xu, M., Wang, N., 2016. A bootstrap test for constant coefficients in
enlargement process: an exploratory spatial data analysis, 1995-2000. Ann. Reg. Sci. geographically weighted regression models. Int. J. Geogr. Inf. Sci. 30 (8),
40 (4), 723–765. 1622–1643.
Farber, S., Páez, A., 2007. A systematic investigation of cross-validation in GWR model Mur, J., López, F., Angulo, A., 2008. Symptoms of instability in models of spatial
estimation: empirical analysis and Monte Carlo simulations. J. Geogr. Syst. 9 (4), dependence. Geogr. Anal. 40 (2), 189–211.
371–396. Ord, J., Getis, A., 1995. Local spatial autocorrelation statistics:distributional issues and
Fotheringham, A., 2009. the problem of spatial autocorrelation and local spatial an application. Geogr. Anal. 27 (4), 286–306.
statistics. Geogr. Anal. 41 (4), 398–403, (cited By 14). Pace, R., Barry, R., 1997a. Quick computation of spatial autoregressive estimators.
Fotheringham, A.S., Brunsdon, C., Charlton, M., 1999. Some notes on parametric Geogr. Anal. 29 (3), 232–247.
significance tests for geographically weighted regression. J. Reg. Sci. 39, 497–524. Pace, R., LeSage, J., 2004. Spatial autoregressive local estimation. Spat. Econ. Spat. Stat.,
Fotheringham, A.S., Oshan, T.M., 2016. Geographically weighted regression and 31–51.
multicollinearity: dispelling the myth. J. Geogr. Syst. 18 (4), 303–329. Páez, A., Farber, S., Wheeler, D., 2017/02/09 2011. A simulation-based study of
Geniaux, G., Ay, J.-S., Napoléone, C., 2011b. A spatial hedonic approach on land use geographically weighted regression as a method for investigating spatially varying
change anticipations. J. Reg. Sci. 51 (5), 967–986. relationships. Environment and Planning A, 43, 12, pp. 2992–3010.
Harris, R., Moffat, J., Kravtsova, V., 2011. In search of w. Spat. Econ. Anal. 6 (3), Páez, A., Uchida, T., Miyamoto, K., 2002. A general framework for estimation and
249–270. inference of geographically weighted regression models: 1. location-specific kernel
Hoaglin, D., Welsch, R., 1978. The hat matrix in regression and ANOVA. Am. Stat. 32 bandwidths and a test for locational heterogeneity. Environ. Plan. A 34 (4), 733–754.
(1), 17–22. Pinkse, J., Slade, M., Brett, C., 2002. Spatial price competition: a semiparametric
Jennrich, R., 2001. A simple general procedure for orthogonal rotation. Psychometrika approach. Econometrica 70 (3), 1111–1153.
66 (2), 289–306. Ripley, B., 2005. Spatial Statistics 575. Wiley-Interscience.
Kelejian, H., Prucha, I., 1998. A generalized spatial two-stage least squares procedure for Shao, J., 1993. Linear model selection by cross-validation. J. Am. Stat. Assoc. 88 (422),
estimating a spatial autoregressive model with autoregressive disturbances. J. Real 486–494.
Estate Financ. Econ. 17, 99–121. Speckman, P., 1988. Kernel smoothing in partial linear model. J. R. Stat. Soc. Ser. B 50,
Kelejian, H., Prucha, I., 1999. A generalized moments estimator for the autoregressive 413–436.
parameter in a spatial model. Int. Econ. Rev. 40 (2), 509–533. Stakhovych, S., Bijmolt, T., 2009. Specification of spatial models: a simulation study on
Kelejian, H., Prucha, I., Yuzefovich, Y., 2004. Instrumental variable estimation of a weights matrices. Pap. Reg. Sci. 88 (2), 389–408.
spatial autoregressive model with autoregressive disturbances: large and small Thomas-Agnan, C., Laurent, T., Goulard, M., 2013. About Predictions in Spatial
sample results. Adv. Econ. 18, 163–198. Autoregressive Models: Optimal and Almost Optimal Strategies. Working Paper.
Kelejian, H.H., Prucha, I.R., 2007. The relative efficiencies of various predictors in spatial Wei, C.-H., Qi, F., 2012. On the estimation and testing of mixed geographically weighted
econometric models containing spatial lags. Reg. Sci. Urban Econ. 37 (3), 363–374. regression models. Econ. Model. 29 (6), 2615–2620.
Kelejian, H.H., Robinson, D.P., 1998. A suggested test for spatial autocorrelation and/or Wheeler, D., Tiefelsdorf, M., 2005a. Multicollinearity and correlation among local
heteroskedasticity and corresponding monte carlo results. Reg. Sci. Urban Econ. 28 regression coefficients in geographically weighted regression. J. Geogr. Syst. 7 (2),
(4), 389–417. 161–187.
Kelley Pace, R., LeSage, J.P., Zhu, S., 2012. Spatial dependence in regressors and its Wheeler, D., Tiefelsdorf, M., 2005b. Multicollinearity and correlation among local
effect on performance of likelihood-based and instrumental variable estimators. In: regression coefficients in geographically weighted regression. J. Geogr. Syst. 7 (2),
30th Anniversary Edition. Emerald Group Publishing Limited, pp. 257–295. 161–187.
Le Gallo, J., 2004. Hétérogénéité spatiale. Econ. Prévision 162 (1), 151–172. Wheeler, D.C., 2007. Diagnostic tools and a remedial method for collinearity in
Le Gallo, J., Fingleton, B., 2012. Measurement errors in a spatial context. Reg. Sci. urban geographically weighted regression. Environ. Plan. A 39 (10), 2464–2481.
Econ. 42 (1), 114–125. Wheeler, D.C., 2009. Simultaneous coefficient penalization and model selection in
Lee, L.-f., 2003. Best spatial two-stage least squares estimators for a spatial geographically weighted regression: the geographically weighted lasso. Environ.
autoregressive model with autoregressive disturbances. Econ. Rev. 22 (4), 307–335. Plan. A 41 (3), 722–742.
LeSage, J., Fischer, M., 2008. Spatial growth regressions: model specification, estimation Wood, S.N., 2011. Fast stable restricted maximum likelihood and marginal likelihood
and interpretation. Spat. Econ. Anal. 3 (3), 275–304. estimation of semiparametric generalized linear models. J. R. Stat. Soc.: Ser. B Stat.
LeSage, J.P., Pace, R.K., 2007. A matrix exponential spatial specification. J. Econ. 140 Methodol. 73 (1), 3–36.
(1), 190–214.

12