Journal of Financial Management of Property and Construction

Spatial variation as a determinant of house price: Incorporating a geographically

weighted regression approach within the Belfast housing market
M. McCord P.T. Davis M. Haran S. McGreal D. McIlhatton
Article information:
To cite this document:
Downloaded by FLINDERS UNIVERSITY OF SOUTH AUSTRALIA At 10:44 15 February 2016 (PT)

M. McCord P.T. Davis M. Haran S. McGreal D. McIlhatton, (2012),"Spatial variation as a determinant of

house price", Journal of Financial Management of Property and Construction, Vol. 17 Iss 1 pp. 49 - 72
Permanent link to this document:
http://dx.doi.org/10.1108/13664381211211046
Downloaded on: 15 February 2016, At: 10:44 (PT)
References: this document contains references to 91 other documents.
To copy this document: [email protected]
The fulltext of this document has been downloaded 600 times since 2012*
Users who downloaded this article also downloaded:
Chyi Lin Lee, (2009),"Housing price volatility and its determinants", International Journal of Housing
Markets and Analysis, Vol. 2 Iss 3 pp. 293-308 http://dx.doi.org/10.1108/17538270910977572
Paul De Vries, Peter Boelhouwer, (2005),"Local house price developments and housing supply", Property
Management, Vol. 23 Iss 2 pp. 80-96 http://dx.doi.org/10.1108/02637470510589968
Marcelo Bianconi, Joe A. Yoshino, (2013),"House price indexes and cyclical behavior", International Journal
of Housing Markets and Analysis, Vol. 6 Iss 1 pp. 26-44 http://dx.doi.org/10.1108/17538271311305995

Access to this document was granted through an Emerald subscription provided by emerald-srm:272736 []
For Authors
If you would like to write for this, or any other Emerald publication, then please use our Emerald for
Authors service information about how to choose which publication to write for and submission guidelines
are available for all. Please visit www.emeraldinsight.com/authors for more information.
About Emerald www.emeraldinsight.com
Emerald is a global publisher linking research and practice to the benefit of society. The company
manages a portfolio of more than 290 journals and over 2,350 books and book series volumes, as well as
providing an extensive range of online products and additional customer resources and services.
Emerald is both COUNTER 4 and TRANSFER compliant. The organization is a partner of the Committee
on Publication Ethics (COPE) and also works with Portico and the LOCKSS initiative for digital archive
preservation.

*Related content and download information correct at time of download.

 The current issue and full text archive of this journal is available at
www.emeraldinsight.com/1366-4387.htm

A determinant
Spatial variation as a determinant of house price
of house price
Incorporating a geographically weighted
regression approach within the Belfast 49
housing market
Downloaded by FLINDERS UNIVERSITY OF SOUTH AUSTRALIA At 10:44 15 February 2016 (PT)

M. McCord, P.T. Davis, M. Haran, S. McGreal and D. McIlhatton

School of the Built Environment, University of Ulster,
Newtownabbey, UK

Abstract
Purpose – Tobler’s law of geography states that things that are close to one another tend to be more
alike than things that are far apart. In this regard, the spatial pattern of price distribution is defined by
the arrangement of individual entities in space and the geographic relationships among them. The
purpose of this paper is to provide emerging findings of research analysing the salient factors which
impact on the sale price of residential properties using a spatial regression approach.
Design/methodology/approach – The research develops and formulates a geographically
weighted regression (GWR) model to incorporate residential sales transactions within the Belfast
Metropolitan Area over the course of 2010. Transaction data were sourced from the University of
Ulster House Price Index survey (2010, Q1-Q4). The GWR approach was then evaluated relative to a
standard hedonic model to determine the spatial heterogeneity of residential property price within the
Belfast Metropolitan Area.
Findings – This investigation finds that the GWR technique provides increased accuracy in
predicting marginal price estimates, in comparison with traditional hedonic modelling, within the
Belfast housing market.
Originality/value – This study is one of only a few investigations of spatial house price variation
applying the GWR methodology within the confines of a UK housing market. In this respect it
enhances applied based knowledge and understanding of geographically weighted regression.
Keywords Hedonic price models, House price, Housing, Pricing, Geographically weighted regression,
Spatial analysis, Ireland
Paper type Research paper

1. Introduction
Housing markets are unique as “no two houses are the same” (Miles, 1994). Price
determination is a result of the interaction between buyers and sellers in a defined
geographic region. O’Sullivan and Gibb (2003) identify that housing as a commodity is
distinct, with the defining characteristic is that it is locationally fixed with price a
function of location. Theoretically the price of housing characteristics are embedded in
the spatial heterogeneity of large housing markets due to demand and supply
imbalance (Goodman and Thibodeau, 1998; Dubin and Sung, 1987; Bitter et al., 2007). Journal of Financial Management of
Within large metropolitan locales, housing is not substitutable due to the complicated Property and Construction
Vol. 17 No. 1, 2012
patterns of consumer behaviour, choice, and budget constraints. House price is pp. 49-72
therefore considered a function of a bundle of inherent attributes which yield utility q Emerald Group Publishing Limited
1366-4387
maximisation for purchasers. Locational, accessible and structural attributes as well as DOI 10.1108/13664381211211046
 JFMPC neighbourhood characteristics result in differentials between house price and influence
17,1 demand-supply elasticity within particular submarkets. In this regard, the estimation
of spatial dynamics and the structure of markets are necessary to accurately examine
and explain house price coefficients (Bitter et al., 2007).
Metropolitan housing markets are inherently characterised by interrelated
submarkets, as a consequence price is considered a determinant of many important
50 attributes within a spatial dynamic. In the opinion of Fotheringham et al. (1998) it is
reasonable to assume that relationships vary over space, indeed they advocate that it is
reasonable to assume that parameter estimates display significant spatial variation
Downloaded by FLINDERS UNIVERSITY OF SOUTH AUSTRALIA At 10:44 15 February 2016 (PT)

due to preference, taste as well as administrative and political issues which manifest
within different housing locals. Indeed, it is widely acknowledged that spatial
phenomena vary across markets, therefore, location is an amalgam of several factors
and widely recognised as an important determinant of house price.
Traditional empirical analysis of house prices has customarily applied hedonic
regression models to explain the structural and locational attributes. These approaches
estimate the implicit price of each property characteristic by regressing transaction prices
of properties on corresponding property features (Choy et al., 2007). Conventional hedonic
modelling typically imposes a constant price structure on housing characteristics
throughout an entire market area (Bitter et al., 2007), thus, conceptualising metropolitan
locales as a single unified market. Therefore, whilst hedonic models incorporate
neighbourhood features, the neglect of spatial characteristics can detract from accurate
evaluation of housing submarkets and therefore house price determination. Furthermore,
standardised regression models are considered to be limited in that they can overestimate
prices in some neighbourhoods, whilst concomitantly underestimating prices in other
jurisdictions (Fotheringham et al., 2002).
This paper sets out to undertake an empirical comparative analysis of two distinct
modelling techniques that specifically address the marginal price of residential
property. A geographically weighted regression (GWR) model is applied to examine
the spatial variability of regression coefficients across the selected study area. GWR is
conceptually different from both spatial auto-regression and ordinary least squares
(OLS) techniques, as the GWR model provides a more localised insight into how
coefficients are determined by specific variables. Nonetheless, the application of GWR
within the UK remains somewhat underdeveloped with only a small number of studies
(Brunsdon et al., 1999; Farber and Yates, 2007; Bitter et al., 2007; Crespo et al., 2007)
applying this technique for explaining the spatial distribution of house prices. This
investigation therefore has the potential to offer fresh insight and understanding of
spatial dynamics within house price modelling.

2. Housing and spatial dimensionality: a review of the literature

Residential property is a multi-dimensional commodity normally observed as a bundle
of utility-bearing attributes commanding a distinct price (Choy et al., 2007). These
principal attributes of housing are characterised by their physical inflexibility, high
durability, multi-dimensional heterogeneity, and spatial immobility (Rothenberg et al.,
1991). As housing is perceived to be a multi-dimensional heterogeneous good, it is
reasonable to argue that each housing unit is unique when considered in terms of
structural attributes and locational and neighbourhood characteristics (Malpezzi,
2003). House prices are considered to be normally determined by accessibility to
 employment hubs and amenities, and also by environmental attributes of location, A determinant
consequently, housing differs in numerous characteristics and is a parcel of salient of house price
attributes which command a price premium. Indeed, these unique bundles of attributes
have been examined in a number of studies spanning more than half a century (Muth,
1960; Ridker et al., 1967; Stegman, 1969; Kain and Quigley, 1975; Evans, 1973; Follain
and Malpezzi, 1980; Bover et al., 1989; So et al., 1997; Malpezzi, 1998; Bover and Velilla,
2002; Song and Shon, 2007) suggesting that many of these attributes are spatially 51
related in the form of location hierarchy.
The spatially immobile aspect of housing means that location is an intrinsic attribute
Downloaded by FLINDERS UNIVERSITY OF SOUTH AUSTRALIA At 10:44 15 February 2016 (PT)

and inherent determinant of housing market value. As a result, the concepts of spatial
dependence and housing submarkets are closely related as submarket analysis relies on
the notion of substitutability (Bourassa et al., 2003). Traditional conjecture from
Straszheim (1975) implies that the housing market is best understood as a cluster of
functionally independent geographic submarkets, differentiated by the characteristics
of their housing units and or the locations of the supposed submarkets. Applying this
view, systematic geographic variation of housing characteristics and prices within an
urban area is a fundamental characteristic of the housing market. In a similar vein,
Grigsby (1963) classifies submarkets in terms of close substitutability, perceiving each
submarket as a collection of possible dwellings that a household can reasonably expect
to consume within a larger market. Significantly, Gribsby contends that this close
substitutability for dwellings tends to fit within a narrow price band. This view is
endorsed by Galster (1996) who suggests that the salient dynamic translating a larger
housing market into these price categories is the quality of housing. Thus, Galster (1996)
identifies that each submarket functions as an independent market equilibrating to a
price dictated by the conditions existing within that submarket. Due to these aspects,
determinants of house prices are complex and problematic.

Approaches to modelling housing markets

The hedonic price method is a well established analytical technique within housing
markets and is premised on Lancaster’s (1966) theory of consumerism which was
subsequently applied by Rosen (1974) within the confines of housing market analysis.
Multiple-regression analysis (MRA) models are considered a classical and primary
technique for explaining and predicting property values within which locational factors
can potentially be taken into account. Indeed, MRA has been used to estimate residential
property values in the US since the 1960s and in the UK since the 1980s (Pendleton, 1965;
Greaves, 1984; Adair and McGreal, 1988). In the opinion of Michaels and Smith (1990)
these approaches are nonetheless inconsistent with the theory that prices of housing
attributes exhibit spatial heterogeneity within housing markets.
A continued limitation of these approaches however is the absence of applied
locational dummies to explain spatial variation. Casetti (1972, 1997) employed an
expansion method to permit parameter estimates to vary based on location attributes,
namely neighbourhood and accessibility attributes. Similarly, Can (1990) applied an
expansion method to account for price shifting with associated neighbourhood
interaction effects finding several variables to be significant. An enhanced interactive
approach has more recently been applied (Thériaut et al., 2003) to facilitate the housing
attribute price to vary over space, based on neighbourhood and accessibility
characteristics. Thériaut et al. (2003) furthered the spatial expansion method by
 JFMPC incorporating accessibility and neighbourhood parameters using GIS. This approach is
17,1 limited as it utilises average house price estimates over a geographic area thereby
neglecting to accurately explain a continuous price gradient.
Alternatively, Eckert (1990) suggested that, based on the assumption that subsets are
characterised by a lower variance, models generated for housing submarkets should
yield greater explanatory power (and predictive accuracy) than those computed at the
52 overall market level. Paradoxically, numerous localised modelling techniques have been
proposed to capture spatial heterogeneity in housing markets. Cliff and Ord (1981)
produced a simultaneous autoregressive model, similar to an OLS approach, examining
Downloaded by FLINDERS UNIVERSITY OF SOUTH AUSTRALIA At 10:44 15 February 2016 (PT)

the expected value of y, E( y) and the x variables. Goodman and Thibodeau (1998) also
introduced the concept of hierarchical linear modelling, concluding that dwelling
characteristics, neighbourhood characteristics and submarkets interact to influence
house prices. A more recent study by Xu (2008) applies a hedonic model incorporating
polynomial expansion and spatial characteristics using specific X, Y coordinates. These
expansion models successfully examine the heterogeneities within housing attribute
prices’ providing strong evidence that spatial non-stationarity varies with location.
In the opinion of Yu et al. (2007) these techniques are fundamentally flawed in that
they fail to adequately account for spatial autocorrelation, the coincidence of selling
price with location and spatial heterogeneity. Previous research by Fotheringham et al.
(1998) highlighted that failure to account for spatial autocorrelation contributes to high
standardised errors and less efficient coefficient estimates. In this regard, spatial
autocorrelation violates the regressive assumptions that the housing market acts in
relative equilibrium (Moore and Myers, 2010). Moreover, traditional models remain
inherently global in diagnosis, at best relying on broadly applied submarket dummy
variables, typically based upon census or electoral sub divisions. Consequently, to
accurately explain spatial variation in housing price, absolute location using {X, Y}
coordinates is fundamental to better explain house price determination.
Existing literature, notably Brunsdon et al. (1996), Fotheringham et al. (1998, 2002)
have all advocated for GWR as a local variation modelling technique for analysing house
price determinants, as it allows for the exploration of the variation of the parameters as
well as the testing of the significance of this variation. The GWR approach incorporates
the application of housing parcel centroid coordinates (Brunsdon et al., 1996), which is
identified by McMillen and Redfearn (2010) as a special case of the locally weighted
regression modelling as it is applied to geographic space. The approach is also
non-parametric, thus not requiring any assumptions to be made regarding the
underlying distributions of values of the predictor variables. Furthermore, this has the
ability to handle highly skewed and categorical predictors. As a technique GWR has
expanded substantially across many disciplines. Within the property context, it has two
core uses; as an exploratory tool to understand varying taste and preferences for
different property attributes and as a statistical technique to enhance estimation of
property price with a given set of attributes taking into account the effects of location.
Pertinently GWR as an analytical technique for price estimation has assumed
greater prominence in recent years. Brunsdon et al. (1999) in a study of house prices in
the town of Deal in south-east England examined the determinants of house price with
GWR and found that the relationship between house price and size varied significantly
through space. Meanwhile, Pavlov (2000), Fotheringham et al. (2002) and Yu (2006)
have all applied the GWR or GWR-similar methodology to housing market analysis.
 Nonetheless, the application of GWR to housing sub-markets in the UK remains A determinant
largely underdeveloped, additionally there appears to be little examination evaluating of house price
the GWR model relative to the common hedonic assessment methodologies. Given the
potential of these approaches to better explain housing markets for academic, lending
and property taxation purposes, there appears to be considerable merit in undertaking
further empirical analysis.
53
3. The Belfast housing market
Downloaded by FLINDERS UNIVERSITY OF SOUTH AUSTRALIA At 10:44 15 February 2016 (PT)

The analysis undertaken by this research focuses on a case study of the Belfast Local
Government District (LGD) area[1] contained in the Belfast metropolitan area (BMA)
within Northern Ireland. The rationale behind using the LGD as the spatial extent of
the research is due to the high sample size of property transactions throughout the
defined urban area (Figure 1). House price patterns are generally differentiated
throughout BMA, however, on the urban fringe and within enclave pockets of
affluence, the market generally commands a higher house price, due to enhanced
structural housing characteristics and associated neighbourhood effects.
The Belfast housing market has undergone rapid house price acceleration and
depreciation during the ten year period 2000-2010. Prices remained generally stable
during the start of the decade; however, the period between 2003 and 2007 witnessed an
increase in the average property price of 178 per cent (UU HPI, 2008). Post 2007, figures
compiled by the University of Ulster indicate that the average house price decreased by
35.6 per cent from the peak of the market in 2007Q3 to the end of 2009Q4[2]. Overall, the
market in 2010 appeared to stabilise and demonstrated marginal fluctuation.

Figure 1.
The Belfast housing
market area
© Crown copyright 2012. Permit Number 120055
 JFMPC The Belfast market is the largest urban agglomeration in NI. In 2009 the
population of NI was estimated at 1.78 million[3], with the BMA accounting for
17,1 approximately 650,000. The projected population of the City of Belfast in 2009 stood
at 268,000. In relation to housing stock, the BMA comprises around 39 per cent of the
total number of houses in NI2, with Belfast making up 17 per cent of the overall NI
figure and 42 per cent of the BMA total (Table I).
54
4. Methodology
Unlike the global perspective offered by MRA, which has a one-model wide set of
Downloaded by FLINDERS UNIVERSITY OF SOUTH AUSTRALIA At 10:44 15 February 2016 (PT)

regression coefficients, the GWR technique produces a differing set of coefficients for
every property processed through a series of weighted least squares (WLS) regressions
(Moore and Myers, 2010). This weighting is determined by nearest neighbours distance
analysis (bandwidth), which results in a combination of many small weighted MRAs
that are performed around each subject property. Significantly, this results in the GWR
approach being considered as a function of location as it is a local modelling approach
that explicitly allows parameter estimates to vary over space (Brunsdon et al., 1996;
Fotheringham and Brunsdon, 1999; Fotheringham et al., 2002). These separate
estimates are constructed through the incorporation of the dependent and explanatory
variables falling within the bandwidth of each target feature allowing for the
production of “local” regression results. The output is a set of spatial statistics which
denote local relationships (Fotheringham et al., 1998).
In GWR the sample points within a fixed distance of a given point are included in
the regression, with all others excluded. Nonetheless, the regression points do not
receive equitable weighting; instead, the weight attributed is a function of location
which diminishes with the distance from the regression point. This weighting
function is premised on the (X, Y) coordinates of the regression point and the data
points in adjacent proximity. Therefore, the peak of the surface is the regression point
with sample points under the surface being attributed a weight based on the height of
the surface at that specific point (Figure 2).

Belfast BMA NI
Table I. Number of households 116,000 273,000 700,000
Total number of
housing stock Source: NISRA 2010

h
dij wij
Figure 2. X
Weighting function based
on the coordinates of the Y
Sample property
regression point and the Subject property (regression point)
data points
Source: Adapted from Borst and McCluskey (2008)
 As shown in Figure 2, the kernel density function shows the spatial dimension along A determinant
the X-axis, with the weighting function represented by the vertical height of the curve. of house price
The height of the curve at point j, given by wij, is the weight applied to point j when
point i is the regression point and dij is the distance between the regression point i and
data point j (Borst and McCluskey, 2008). The bandwidth of the spatial kernel, h, is a
parameter that affects how the weight is computed as the distance between the
regression point and other sample point’s increase. The delineation between 55
bandwidth selection is an important consideration as the differential between a
fixed or adaptive bandwidth can result in insufficient nodes being in(ex)cluded within
Downloaded by FLINDERS UNIVERSITY OF SOUTH AUSTRALIA At 10:44 15 February 2016 (PT)

the parameter estimation.

Mechanics of GWR methodology

The study of the attributes influencing house price using GWR across the LGD market
level is unique for house price analysis in Northern Ireland. The study builds upon
existing hedonic literature (Adair, 1992) which identifies the general function of house
prices in the standard hedonic as the function of a set of housing characteristics:
X
y i ¼ a0 þ ak x i k þ 1 i ; ð1Þ

where xik represents the ith observation of the kth independent variable. The GWR
specification is analogous; however, the unique coefficients are estimated at each
specific observation point. As previously mentioned, the GWR model extends the
traditional regression framework by allowing the parameters to be estimated locally so
that the model can be rewritten as:
X
yi ¼ a0i ðui ; vi Þ þ ak ðui ; vi Þxki þ 1i ; ð2Þ

where (ui, vi) denotes the coordinates of the ith point in space and ak (ui, vi) is the
continuous function ak (ui, vi) at point i, indicating that the spatial variations between
the relationships provide a mechanism for measurement (Fotheringham et al., 1998).
Previous studies have provided a rationale for helping to calibrate equation (2) due to
the elevated level of unknown than observed variables. In this regard, this paper
mimics Fotheringham et al. (1998), where the coefficients are deterministic functions of
other variables (location in space). As conceptualised by Fotheringham et al. (1998),
this estimation process is a substitution between bias and standard error (For a full
explanation, see Fotheringham et al. (1998)). According to Fotheringham et al. (1998)
equation (2) implicitly assumes that data in close proximity to i influence the
estimation of ak (ui, vi) more so than data located further away, a notion accepted in
the research undertaken by Huang et al. (2010). The result is that this measures the
relationships inherent in the model around each node i, using weighted least squares.
Hence, an observation is weighted in accordance to proximity to point i, resulting in it
no longer remaining constant. In matrix notation the model parameters are:

aðui ; vi Þ ¼ ðX T W ðui ; vi ÞXÞ21 X T W ðui ; vi Þy; ð3Þ

where W(ui, vi) is the spatial weighting matrix. As observed in Bitter et al. (2007) the
Gaussian function specifies d to signify the Euclidian distance between the regression
node and observation point, with h denoting the bandwidth as follows:


JFMPC d 2
W i ðui ; vi Þ ¼ exp 2 ð4Þ
17,1 h
As highlighted by Fotheringham et al. (1998), and more recently by Bitter et al. (2007)
the findings of GWR are sensitive to bandwidth estimations, which can distort
variances in the estimators. Following the approach of Bitter et al. (2007), this paper
56 employs the same procedure selecting the spatial kernel function which permits the
bandwidth to modify in relation to transaction price at each regression node. The
bandwidth selected was the adaptive approach as it permits nearby properties to be
Downloaded by FLINDERS UNIVERSITY OF SOUTH AUSTRALIA At 10:44 15 February 2016 (PT)

afforded more weight than properties further away, following Tobler’s (1970) first
principle of geography and the backbone of the comparative appraisal method. Indeed,
this approach observes the weight function applied to all sale properties as a sliding
neighbourhood around each property in which a local regression specific to a number
of sale properties is variable (Moore and Myers, 2010). In essence, distances differ and
the allocation of nearest neighbours remains the same. The most common adaptive
bandwidth weighting approach as suggested by Brunsdon et al. (2002) is the bi-square
function given as:

2 !2
dij
vij ¼ 1 2 dij # b vij ¼ 0; otherwise ð5Þ
b

where dij is geographic distance between the ith subject property and its jth
neighbouring sale property, and b is equal to the bandwidth.

5. Data
The analysis is conducted using data for 2010 (n ¼ 676) with these observations
highlighting the period of market correction post housing boom. The sample size
accounts for 20 per cent of all house price transactions in Northern Ireland[4]. This
stationary point in time (2010) was selected within the modelling phase as it was
considered that the high volatility recently witnessed in the housing market would not
accurately capture true purchaser sentiment and behaviour. Furthermore, the
utilisation of observed house price values for this study adds significant weight to
the research as such an approach is considered to reduce levels of bias in the
methodological design which would be commonly associated with analysis involving
asking price (Sirmans et al., 2005).
The achieved selling price is used as the dependent variable. In total, seven
explanatory variables based on the structural attributes of each property were used in
the research (Table II). The variables used include property type, age, gross floor area
(GFA), properties with a garage present, heating type, number of bedrooms and
number of reception rooms (Table II). According to Orford (1999) it is impossible to
accurately specify all locational attributes that affect housing price. Therefore, location
variables incorporated into the study were; distance to transport amenities and hubs,
local services and the CBD, schools, and areas of open space. The distances were
calculated using ArcGIS 10 which permitted the spatial distance to be measured for
each {X, Y} coordinate and therefore each individual property transaction. The data
for the price, structural attributes and house type variables were made available by the
University of Ulster Northern Ireland House Price Index.
 A determinant
Variable Description
of house price
Price Selling price in pounds sterling (£)
Type Categorical variable indicating property type
Age Categorical variable indicating the period the property was built
GFA Property area in square feet (sq. ft)
Bedrooms Number of bedrooms within the property 57
Receptions Number of reception rooms within the property
Garage Dummy variable indicating the presence of a garage
Downloaded by FLINDERS UNIVERSITY OF SOUTH AUSTRALIA At 10:44 15 February 2016 (PT)

Heating Categorical variable indicating the type of heating system in the property
Distance_Road Categorical variable indicating the distance to the nearest principal road
CBD_Distance Distance to the CBD in metres
GrammerSchool Distance to the nearest grammar school in metres
GP_Distance Distance to the nearest GP surgery in metres
GreenSpace_Dis Distance to the nearest area of green open space in metres
PrimarySchools Distance to the nearest primary school in metres
Train_Distance Distance to the nearest train halt in metres Table II.
Bus_Distance Distance to the nearest bus terminal in metres Variable descriptions

Model comparison criterion

The main criteria adopted for comparing the two models (OLS and GWR) are the
conventional goodness-of-fit measures using the adjusted R 2. Furthermore, and similar
to previous studies, in order to judge whether there is statistical significance in the
differences between the accuracies achieved by the models, McNamara’s test was
performed which is based on the standardized normal test statistic (Foody, 2004).

Results, comparison and analysis

Trend analysis within the study area data shows hotspots between property price and
the location density of transactions within the Belfast housing market in 2010. The
hotspot analysis reveals that the level of transaction density is generally higher around
the bands of higher price indicating that the property market is potentially spatially
clustered in terms of market demand. The distribution diagram suggests that areas
that have stronger house prices also have more buoyant markets. The density output
further highlights that more market activity occurs in the south-west of Belfast and
generally this sub-market is comprised of above average house prices. Furthermore,
the west and north of the metropolitan area shows pockets of high prices indicating
distinctive submarket areas (Figure 3).
The initial OLS regression model comprising the 15 property related attributes is
reported in Table III. The findings show the model to be statistically significant with
the adjusted R 2-value explaining 70.6 per cent of house price variation in which the
marginal-price estimates are held constant throughout the study area. The analysis of
variance (ANOVA) test also revealed that overall the model was statistically
significant. The preliminary analysis highlighted eight coefficients to be statistically
significant determinants of house prices in Belfast at the 95 per cent level. However,
the model diagnosis highlighted instances of near-collinearity amongst a number of
attribute coefficients.
This was identified within the correlation findings (Table IV) which showed the
GFA variable to comprise high correlation with BEDS (0.745, p , 0.001) and RECEP
 JFMPC
17,1

58
Downloaded by FLINDERS UNIVERSITY OF SOUTH AUSTRALIA At 10:44 15 February 2016 (PT)

Figure 3.
Density distribution of
transactions and house
price distributions

Variable Coefficient t-statistic 95% confidence interval Total VIF

Intercept 217,920.071 20.846 259,515.57 23,675.43

Type 9,055.138 6.311 * 6,237.80 11,872.47 0.556 1.797
Age 3,380.327 1.697 2 532.18 7,292.83 0.607 1.648
NFA 136.760 18.011 * 121.85 151.67 0.283 3.534
Bedrooms 1,606.334 0.402 26,242.14 9,454.81 0.379 2.635
Receptions 9,022.283 1.553 * * * 22,387.61 20,432.18 0.485 2.063
Garage 225,017.586 24.248 * 236,582.05 213,453.12 0.683 1.464
Heating 5,872.601 1.421 22,241.91 13,987.11 0.773 1.293
Distance_Road 22,588.231 21.142 27,037.22 1,860.75 0.878 1.139
CBD_Distance 2 11.994 23.675 * 2 18.40 25.58 0.088 11.375
GrammerSchool 2 11.509 23.915 * 2 17.28 25.73 0.397 2.517
GP_Distance 13.790 2.160 * * 1.25 26.32 0.469 2.133
GreenSpace_Dis 4.744 0.450 2 15.95 25.44 0.671 1.490
PrimarySchools 20.562 2.107 * * 1.40 39.72 0.475 2.107
Train_Distance 12.377 3.240 * 4.87 19.87 0.076 13.185
Bus_Distance 3.726 1.256 2 2.10 9.55 0.193 5.188
R2 0.713
Adjusted R 2 0.706
Table III. F-statistic 108.35 *
OLS attribute coefficients n 676
and collinearity
summaries Note: Significant at: *0.001, * *0.05 and * * *0.10 levels
 (0.683, p , 0.001), and the CBD_DIS variable to be severely correlated with both A determinant
TRAIN_DIS (0.946, p , 0.001) and (0.873, p , 0.001). In addition, the model also of house price
identified a number of outliers within the data displaying elevated standardized
residuals (. 2.5). In total, 31 outliers were removed from the study to reduce any bias or
skewness within the data.
In order to address these problems of multicollinearity, principal component
analysis (PCA) was performed on the data to reduce bias, variance inflation and 59
collinearity between the variables. This data dimensionality technique was selected as
it retains as much information as possible whilst suppressing the data into a reduced
Downloaded by FLINDERS UNIVERSITY OF SOUTH AUSTRALIA At 10:44 15 February 2016 (PT)

number of variables. Significantly, Bitter et al. (2007) identify that this process also
serves to help the performance of GWR, due to its inherent computational limitations.
The PCA technique requires that the correlation matrix R ¼ {R[Rmm]} is obtained
through the transformation of the data matrix X ¼ {X[mn]} into a matrix of
standardised scores [Z] which are computed using the mean and standard deviation for
each row m of the data matrix where m is the number of elements (variables) and n is
the number of observations (column vectors) in the dataset.
To identify the number of principle components (PCs) to be retained, the analysis
applies a priori criteria in order to select the number of principal components that
explain the maximum amount of variance. The criterion employed is that each principal
component explains at least 5 per cent of the variance and cumulatively 75 per cent of
variance. Variables with absolute scoring coefficients . 0.4 were considered important
contributors to a pattern, a logic premised on Steven’s (2002) suggestion that loadings
greater than 0.4 represent substantive values which are appropriate for interpretative
purposes. The Varimax rotation method with Kaiser Normalization is applied in order to
furnish the most accurate and reliable inferences within the data.
The Bartlett test of sphericity and the Kaiser-Meyer-Olkin measure of sampling
accuracy were performed on the data. Based on the guidelines suggested by Kaiser
(1974) and Hutcheson and Sofroniou (1999), the statistic 0.718 falls into the range of
being “good”. Moreover, the Bartlett measure is highly significant ( p , 0.001)
indicating that the R-matrix is not an identity matrix. Overall, both tests show that
PCA is appropriate for the data. The PCA analysis retained two components with
eigenvalues greater than one, which cumulatively explain 84.78 per cent of the total
variance in the original six variables (Table V).
The extracted and rotated sums of squared loadings indicated that two components
should be retained, with the remaining discarded factors ignored. This is confirmed by
both the component score matrix and scree plot (Appendix) which suggest the
retention of two components as they meet all assumptions and guidelines advanced in
the criterion of extraction. The first extracted component loaded positively on distance
to the CBD, train halts and bus stations and represents the distance to transport nodes

NFA BEDS RECEPS CBD_Dis Train_Dis Bus_Dis

NFA 1 CBD_Dis 1
BEDS 0.745 * 1 Train_Dis 0.946 * 1
RECEPS 0.683 * 0.687 * 1 Bus_Dis 0.869 * 0.873 * 1 Table IV.
Highly correlated
Note: Significant at: *0.001, * *0.05, and * * *0.10 levels findings
 JFMPC and CBD amenities. The second component reveals a high positive loading between
17,1 GFA, the number of bedrooms and number of reception rooms thereby measuring
house dimension.
The OLS regression was re-run substituting the newly created variables
(House_Dim) and (Dist_Amen) for the original six variables. The Descriptive
statistics for the OLS model are presented in Table VI. The OLS findings show the
60 global stationary model to explain a high degree of variation in Belfast house prices as
indicated by an adjusted R 2-value of 0.683. All predictors except two neighbourhood
characteristics (GPs_Dist and Distance_Ro) proved to be statistically significant. In
Downloaded by FLINDERS UNIVERSITY OF SOUTH AUSTRALIA At 10:44 15 February 2016 (PT)

this instance, the results indicate that distance to GP surgeries and proximity to
principal arterial routes does not have a positive effect on house price estimates.
The findings indicate that the most significant variable as expected was the
House_Dim predictor displaying a t-value of 25.532, thereby illustrating that the larger
the living area, the higher the house price. Property type was the second most significant
predictor which is perhaps expected given that an increased level of detachment is
associated with enhanced privacy and improvement within the surrounding living
environment. In addition, the proximity to both grammar and primary schools are
significant at the 99 per cent level which highlight the importance of education
opportunities and catchment boundary areas in the decision making of house purchasers.

Component 1 Component 2

TRAIN_DIS 0.975
CBD_DIS 0.971
BUS_DIS 0.937
GFA 0.931
BEDS 0.841
RECEP 0.830
Table V. Notes: Extraction method: principal component analysis; rotation method: Varimax with Kaiser
Rotated components normalization; a. rotation converged in three iterations

Variable Range Min. Max. Mean SE SD Variance

price 756,000 49,000 775,000 160,577.2 3,776.4 97,824.34 9.570 £ 100

type 5.00 1.00 6.00 2.72 0.07 1.91 3.66
age 5.00 1.00 6.00 4.15 0.05 1.32 1.74
garage 1.00 1.00 2.00 1.77 0.02 0.42 0.18
heating 4.00 0.00 4.00 2.25 0.02 0.56 0.32
Distance_Ro 3.00 1.00 4.00 1.85 0.04 0.96 0.93
Grammar_Dis 6,936 94.37 7,030.87 1,378.12 42.68 1,105.53 1,222,199
GPs_Dist 3,304 18.87 3,323.03 709.39 18.09 468.57 2,195,565
GreenSpace 1,887 14.32 1,901.45 428.88 9.16 237.18 56,256
Primary_Di 2,863 35.34 2,898.51 464.97 11.77 304.80 92,902
Table VI. Dist_Amen 5.20 21.72 3.47 0.00 0.04 1.00 1.00
Descriptive summary House_Dim 7.47 21.58 5.89 0.00 0.04 1.00 1.00
statistics n 645
 The obvious utility of a garage for storage and other ancillary purposes is also significant A determinant
at the 0.01 level showing that in the study area it adds to house price. The new access of house price
location proxy is significant at the 0.01 level suggesting that the accessibility to the CBD
and key transport nodes also has a positive influence on house prices (Table VII).
A hedonic model was also analysed to identify the significance of both the structural
and locational parameters on house price within the study area. The model applies
terraced housing constructed pre-1919 with oil-fired central heating as the intercept. The 61
results by and large represent similar findings to the OLS model, increasing the overall
level of explanation by 1 per cent. In tandem with the OLS model, the House_Dim
Downloaded by FLINDERS UNIVERSITY OF SOUTH AUSTRALIA At 10:44 15 February 2016 (PT)

predictor emerged as the most significant attribute of house price (Table VIII). The
findings highlight some differentials within the coefficients such as property type, age
and heating type.

GWR model
The results emanating from the GWR model are illustrated in Table IX. The findings
show that the localised regression analysis improves the adjusted R 2 to 0.83[5]
illustrating that the explanation of variance has significantly increased and thereby
achieving a more efficient goodness-of-fit than that of OLS model. This enhanced level
of explanation signifies that there are important localised spatial influences affecting
the marginal price of housing within the Belfast market. Significantly, a consequence
of the local nature of the GWR parameter estimates is that they are extensive[6], hence
they are described by their distribution summaries which reveal the extent of their
heterogeneity and spatial non-stationarity across the Belfast housing market.
The model summaries show that throughout the study area the parameter estimates
for the independent variables are not constant but vary widely over space and are
therefore characterised by significant spatial non-stationarity. Nonetheless, similar to
other studies (Borst and McCluskey, 2008; Bitter et al., 2007) the findings do display
instances of counter-intuition. The estimates for the Garage_present parameter varies

Variable Coefficient b t-statistic 95% confidence interval

Intercept 106,912.539 10.409 * 86,741.979 127,083.099

type 10,944.620 0.276 9.554 * 8,695.113 13,194.128
age 4,714.034 0.080 2.908 * 1,530.542 7,897.526
heating 27,044.571 2 0.048 21.912 * * * 214,278.093 188.951
Distance_Ro 21,589.132 2 0.020 20.843 25,291.038 2,112.774
Grammar_Di 210.940 2 0.159 24.650 * 215.561 2 6.320
GPs_Dist 3.634 0.022 0.700 26.563 13.831
GreenSpace 26.996 0.082 3.115 * 9.977 44.015
Primary_Di 28.749 0.112 3.518 * 12.702 44.795
Dist_Amen 10,624.021 0.139 3.714 * 5,007.196 16,240.845
House_Dim 58,361.219 0.688 25.504 * 53,867.647 62,854.790
Garage_present 24,377.687 0.130 4.938 * 14,683.641 34,071.734
R2 0.689
Adjusted R 2 0.683
F-statistic 127.256 *
n 645
Table VII.
Notes: Significant at: *0.001, * *0.05, and * * *0.10 levels; dependent variable: house price Global OLS model
 JFMPC
Variable Coefficient b t-statistic 95% confidence interval
17,1
Intercept 119,509.74 18.089 * 106,535.231 132,484.266
Distance_Ro 22,774.246 2 0.035 21.471 26,478.322 929.831
Grammar_Dis 212.115 2 0.176 25.117 * 2 16.765 2 7.466
GPs_Dist 3.140 0.019 0.610 2 6.966 13.245
62 GreenSpace_Dis 24.765 0.075 2.862 * 7.774 41.756
Primary_Dis 26.267 0.102 3.229 * 10.292 42.242
Dist_Amen 10,196.822 0.134 3.533 * 4,529.723 15,863.920
Downloaded by FLINDERS UNIVERSITY OF SOUTH AUSTRALIA At 10:44 15 February 2016 (PT)

House_Dim 53,566.849 0.632 19.958 * 48,296.141 58,837.557

det 43,417.113 0.160 5.336 * 27,439.830 59,394.395
sdt 15,569.510 0.093 3.313 * 6,342.003 24,797.017
detb 46,403.234 0.082 3.543 * 20,682.039 72,124.429
sdtb 43,626.152 0.059 2.597 * * 10,638.265 76,614.038
appt 40,587.221 0.221 6.156 * 27,640.821 53,533.620
interwar 1,145.387 0.004 0.175 211,675.462 13,966.236
postwar 26,304.529 2 0.026 21.078 2 17794.004 5,184.946
modern 17,080.592 0.099 3.699 * 8,013.350 26,147.835
newbuild 27,413.536 0.132 4.405 * 15,191.877 39,635.195
elec 222,639.642 2 0.036 21.634 249,853.769 4,574.485
gfch 21,138.437 2 0.007 20.242 210,390.651 8,113.778
solidfuel 25,693.632 2 0.003 20.132 290,366.005 78,978.740
Garage_present 22,310.359 0.119 4.460 * 12,487.056 32,133.661
R2 0.703
Adjusted R 2 0.694
F-statistic 73.910
n 645
Table VIII.
Global hedonic model Notes: Significant at: *0.001, * *0.05, and * * *0.10 levels; dependent variable: house price

Min. LQ Median UQ Max.

Intercept 2 162,069.02 82,925.97 127,309.70 154,721.40 348,042.20

Distance 2 67,201.5 28,384.2 2 1,739.92 3,056.162 17,958.36
Grammar_Dis 2164.744 239.2608 215.3355 7.34554 82.46365
GPs_Dist 276.2381 24.65704 30.3864 57.70345 326.3531
GreenSpace_Dis 296.8178 21.46091 29.29793 49.25342 153.769
Primary_Dis 292.6878 228.2324 12.39141 57.33146 265.7421
Dist_Amen 2 155,732 25,186.1 13,542.63 34,507.86 296,422.5
House_Dim 12,392.47 42539.37 52,304.76 66114.13 86,661.41
Garage_present 2 25,164.5 1,527.68 16,661.49 35,768.47 128,422.31
Type 2 9,288.760 1,135.73 5,206.3464 8,380.66883 17,115.2285
Age 2 15,313.282 22,133.39 2,395.7867 8,207.43034 28,473.427
Heating 2 32,982.836 29,179.01 280.60143 8,923.67705 29,331.3191
R2 0.889
Adjusted R 2 0.843
F-statistic 127.256 *
n 645
Table IX.
GWR parameter estimate Notes: Results for PRICE as a response variable and 11 predictor variable(s); spatial function:
summaries bi-squared; adaptive kernel: 12 per cent neighbours
 between £ 2 25,164 and £128,442. Moreover, the negative values for the heating type A determinant
and type of property coefficients indicate counter-intuition as they suggest that of house price
a particular lower classification of housing typology can sell for more than those with a
higher classification at some locations and that the type of heating contained within a
property reduces price in some areas. Of course this may be explained by properties
having heating types which are on paper better (such as full gas central heating) but
which is reality are outmoded, requiring complete replacement. This is an interesting 63
finding which requires further analysis particularly relating to outliers to shed further
light on the nature and causes of this issue.
Downloaded by FLINDERS UNIVERSITY OF SOUTH AUSTRALIA At 10:44 15 February 2016 (PT)

To distinguish between the model accuracies, the results of ANOVA tests on the
observations in Belfast are displayed in Table X. Comparison between the residual
sum of square (RSS) values indicates that the reduction in the residuals further
indicates that the GWR gives a better fit of data than the global OLS model, illustrating
the emphasis on spatial heterogeneity. In addition, the Akaike information criterion
(AIC) statistic also provides a means for comparison between the models, in this
instance the measure shows that the GWR model is the preferred model as it has a
smaller AIC value (Table XI).
To test for the presence of spatial auto-correlation the Morans I test was applied to
measure spatial dependence between values of random variables over geographic
locations. The Morans I test illustrates that there is limited spatial autocorrelation
present in the data (z-score ¼ 1.012703). In essence, this suggests that the error in the
model is “not” spatially concentrated.

Spatial analysis of GWR findings

As previously identified, the GWR approach revealed significant spatial
non-stationarity. Visual representation of the findings capture the spatial expansion
of the GWR method of incorporating spatial heterogeneity and an improvement in
explanatory power and predictive accuracy above and beyond that of the
stationary-coefficient OLS model (Figure 4). In this regard, the GWR technique
appears to better represent the complex spatial patterns inherent in the Belfast data.
Inferentially, this also suggests that spatial heterogeneity may be due to discrete,
localized influences, as well as those operating in a broad, continuous manner over
space. The complex spatial trends depicted by the approach reflect the spatial variation
in the price of housing and reveal that marginal house prices may vary with locational
context and that the stationary coefficient model has obscured this important
relationship.
Visual analysis of the model results displays the significance of the local parameter
estimates, which denote strong evidence of varying local relationships. The spatial
distributions of individual attribute coefficients. For example, the House_Dim

Source of variation RSS DF MS F-test p AICc

þ 12 þ 09
OLS residuals 1.186 £ 10 12.00 1.874 £ 10 127.256 0.00 15,616.62
GWR residuals 2.54646 £ 10þ 41 487.56 2.08501 £ 10þ 29 4.65991 0.00 15,448.61 Table X.
GWR improvement 1.46089 £ 10þ 42 145.44 ANOVA comparison
between GWR and
Notes: RSS (residual sum of squares); DF (degree of freedom); MS (mean square) OLS model fit
 JFMPC
D. class Count Up. limit Y Morans’s I Est. Morans’s I Res. Morans’s I
17,1
1 14,260 912.333 0.284 0.332 20.041
2 27,664 1,824.667 0.184 0.209 20.004
3 38,176 2,737 0.089 0.102 0.002
4 41,054 3,649.333 0.064 0.074 ,0.001
64 5 45,402 4,561.666 0.015 0.015 20.002
6 47,616 5,474 20.059 20.07 ,0.001
7 45,536 6,386.333 20.105 20.117 ,0.001
Downloaded by FLINDERS UNIVERSITY OF SOUTH AUSTRALIA At 10:44 15 February 2016 (PT)

8 42,182 7,298.66 20.08 0.088 20.002

9 31,792 821 20.091 20.102 0.004
10 23,386 9,123.333 20.047 20.06 0.001
11 19,722 10,035.66 20.028 20.024 20.005
12 12,624 10,948 20.007 20.018 0.003
13 10,324 11,860.333 20.098 20.108 20.011
14 6,078 12,772.666 20.034 20.045 0.017
15 5,242 13,684.999 0.034 0.043 0.001
16 1,618 14,597.333 20.019 20.014 20.028
17 806 15,509.666 0.011 0.003 0.02
18 762 16,421.999 20.009 0.002 20.021
19 578 17,334.333 0.005 0.007 20.001
Morans I index 0.233726
Expected index 20.00155
Table XI. Z-score 1.012703
Morans I test Variance 0.053976

PRICE Estimated
550,000
550,000
500,000
500,000
450,000
450,000
400,000
400,000
350,000
350,000
300,000 300,000

250,000 250,000

200,000 200,000

Figure 4. 150,000 150,000

Visual output of GWR 100,000 100,000
analytical results 50,000 50,000

coefficient indicates that there is significant spatial distribution (Figure 5) across the
study area. The estimates are positive and significant throughout the Belfast area and
exhibit relatively smooth spatial trends. The highest marginal-price estimates are
found within east and south Belfast, where homes tend to be amongst the most sought
after. For example, the estimates tend to be high in the more valuable south Belfast
area where demand is high across the house type spectrum and lower towards the west
and north where demand is commensurately lower. The effects of location show a clear
differential in the marginal price of housing with regards to house dimensions
indicating that an additional square foot of living space and or additional bedroom
 GWR Coefficient HOUSE_DIM
85,000
GWR Coefficient GREENSPACE A determinant
140
80,000
120 of house price
75,000
100
70,000
80
65,000
60
60,000
40
55,000
20
50,000
45,000
0 65
–20
40,000
–40
35,000
Downloaded by FLINDERS UNIVERSITY OF SOUTH AUSTRALIA At 10:44 15 February 2016 (PT)

–60 Figure 5.
30,000
25,000 –80 House_Dim and
20,000 –100 GreenSpace coefficient
15,000 –120 estimates
–140

adds more value in particular market segments, thereby exhibiting a more localized
pattern and sub-market differentiation. The findings are in line with local appraisal
knowledge of the housing market and endorsed by the GreenSpace coefficient (Figure 5)
which captures the effects accurately as the east of the city has a large recreational
park area, with the urban core having relatively little open GreenSpace, as expected.
The estimates for several variables reflect complex, localized spatial patterns. The
property type coefficient (Figure 6) appears to have the most beneficial effect on house
prices in the city centre, with a number of other pockets of positive and negative effects
evident, notably in the south-east of the market area, which demonstrates a
pronounced bunching of negative coefficients. Analysis of the age coefficient indicates
a large hinterland of a relatively benign age coefficient effect, populated with enclaves
of negative coefficients on house price estimates.
The preceding analysis infers that whilst the coefficients for type and age are both
positive and significant in the stationary global model, the GWR parameter estimates
indicate that, ceteris paribus, the type and age of property negatively affect price in
substantive sub-markets throughout the Belfast region, thus price estimates
differentiate across space and are dependent upon locational context.

GW R Coefficient TYPE GW R Coefficient AG E

16,000
14,000 25,000
12,000 20,000
10,000
8,000 15,000
6,000 10,000
4,000
2,000 5,000
0 0
–2,000
–5,000
–4,000 Figure 6.
–6,000 –10,000 Property type and age
–8,000 coefficients
–15,000
 JFMPC 6. Conclusions
17,1 This paper set out to undertake an empirical comparative analysis of the traditional
hedonic approach and GWR method specifically addressing the locational
characteristics of house price. The analysis reveals that spatial heterogeneity is
evident within the data. In this regard, it reveals that the traditional hedonic approach
is limited for modelling such data, as it fails to account for spatial non-stationarity.
66 Significantly, the GWR approach improves upon the results of the stationary OLS
model yielding a considerably higher R 2 and outperforming in terms of explanatory
power and predictive accuracy. The GWR technique provides strong evidence that the
Downloaded by FLINDERS UNIVERSITY OF SOUTH AUSTRALIA At 10:44 15 February 2016 (PT)

marginal prices of key housing attributes are not constant throughout the Belfast
market area, but vary with the spatial and locational context which thereby reflects
localised demand and supply dynamics and specific consumer behaviour.
Significantly, the findings highlight the complex spatial structure of housing
markets and the need to explicitly address spatial heterogeneity in housing market
models. A failure to do so may result in a loss of explanatory power, lead to erroneous
conclusions, and obscure important housing market dynamics. GWR in particular
provides a means to visualize the spatial structure of housing markets. Importantly,
the application of GWR provides more informative insight and accuracy in delineating
specific sub-market classifications and analysis. Whilst broad locational “dummies”
can be used to capture location, this does not fully capture the micro locational effects
that can be important in property valuation terms. Improving the market area
“dummies” would involve considerable expert valuer input, or indeed further spatially
based analysis to identify such areas which can then be utilised in a global model.
In terms of the relevance of this research to policy and practice, it is clear that
housing market analysis underpinning automated valuation models for supporting
lending decisions and crucially, supporting property tax revaluation exercises, has
much to gain from adopting spatial approaches such as GWR. Whilst further analysis
using longer time periods, larger datasets, richer attribute data and alternative spatial
approaches is necessary to better specify models for such activity, it is apparent that
spatial analysis is capable of improving performance in these areas. Whilst it is
possible that the results here identify a peculiarity of the Belfast market, it is more
plausible to assume that they are more broadly applicable to housing markets
elsewhere in the UK and beyond. It is also important to note that the results do not
invalidate the use of the standard hedonic approach per se, but rather suggest that in
terms of housing markets at least, the approach should be expanded to better
incorporate the inherently spatial nature – a modern echo of the “location, location,
location” truism at the heart of traditional appraisal approaches. The capacity of
spatial analysis to improve performance in this way depends upon the availability and
quality of spatially referenced data, supporting the drive to improve the quality of
public and private data, in terms of commonality of addressing and spatial referencing.
Moving forward, it is intended that this research will be further developed to
examine the comparative performance of different spatial techniques and the stability
of the performance improvement of spatial techniques suggested herein, over different
data sets, and examining both rental and capital value. By comparing performance of
the leading techniques in different settings, it is hoped that a deeper understanding of
both the underlying market forces and the merits of the techniques themselves will
emerge. Better market analysis using optimised techniques has the capacity to lead to
 more robust lending decisions and more efficient, fairer and more equitable taxation A determinant
systems, thereby strengthening key aspects of both the economy and civil society. of house price
Against this backdrop the emerging findings provided here and the results of future
and similar work can be seen to be a valuable resource in terms of improving practice
and informing public policy.

Notes 67
1. District council represents a geo-political local government area in Northern Ireland.
Downloaded by FLINDERS UNIVERSITY OF SOUTH AUSTRALIA At 10:44 15 February 2016 (PT)

2. The University of Ulster House Price Index shows that house prices in Northern Ireland
peaked in Q3 2007 when the average house price stood at £250,586, the subsequent
correction in house prices resulted in six consecutive quarters of negative growth
consequently by the end of Q1 2009 the average house price in Northern Ireland had fallen
37.4 per cent to £156,857.
3. NISRA Population Projections – 2010.
4. In this period there were a total of 3,383 transactions throughout Northern Ireland.
5. This is a “pseudo”-R 2, calculated as the squared correlation coefficient between the observed
and predicted values for all 646 regressions.
6. The GWR parameter estimates vary at each of the 646 observation points.

References
Adair, A.S. (1992), “The determination of significant variables in the valuation of residential
property”, unpublished PhD thesis, University of Reading, Reading, PA.
Adair, A.S. and McGreal, W.S. (1988), “The application of multiple regression analysis in
property valuation”, Journal of Valuation, Vol. 6 No. 1, pp. 57-67.
Bitter, C., Mulligan, G.F. and Dall’erba, S. (2007), “Incorporating spatial variation in housing
attribute prices: a comparison of geographically weighted regression and the spatial
expansion method”, Journal of Geographical Systems, Vol. 9 No. 1, pp. 7-27.
Borst, R.A. and McCluskey, W.J. (2008), “Using geographically weighted regression to detect
housing submarkets: modelling large-scale spatial variations in value”, Journal of Property
Tax Assessment & Administration, Vol. 5 No. 1, pp. 21-51.
Bourassa, S.C., Hoesli, M. and Peng, V.S. (2003), “Do housing submarkets really matter?”, Journal
of Housing Economics, Vol. 12 No. 1, pp. 12-28.
Bover, O. and Velilla, P. (2002), “Hedonic house prices without characteristics: the case of new
multi unit housing”, Working Paper No. 117, European Central Bank, Frankfurt am Main.
Bover, O., Muellbauer, J. and Murphy, A. (1989), “House prices, wages and the UK labour
market”, Oxford Bulletin of Economics and Statistics, Vol. 51, pp. 97-136.
Brunsdon, C., Fotheringham, A.S. and Charlton, M.E. (1996), “Geographically weighted
regression: a method for exploring spatial non-stationarity”, Geographical Analysis, Vol. 28,
pp. 281-98.
Brunsdon, C., Fotheringham, A.S. and Charlton, M.E. (1999), “Some notes on parametric
significance tests for geographically weighted regression”, Journal of Regional Sciences,
Vol. 39 No. 3, pp. 497-524.
Brunsdon, C., Fotheringham, A.S. and Charlton, M.E. (2002), “Geographically weighted summary
statistics: a framework for localized exploratory data analysis”, Computers, Environment
and Urban Systems, Vol. 26, pp. 501-24.
 JFMPC Can, A. (1990), “The measurement of neighbourhood dynamics in urban house prices”, Economic
Geography, Vol. 66, pp. 254-72.
17,1
Casetti, E. (1972), “Generating models by the expansion method: applications to geographical
research”, Geographic Analyst, Vol. 4, pp. 81-91.
Casetti, E. (1997), “The expansion method, mathematical modelling, and spatial econometrics”,
International Regional Science Review, Vol. 20, pp. 9-32.
68 Choy, L.H.T., Mak, S.W.K. and Ho, W.K.O. (2007), “Modelling Hong Kong real estate prices”,
Journal of Housing and Built Environment, Vol. 22, pp. 359-68.
Downloaded by FLINDERS UNIVERSITY OF SOUTH AUSTRALIA At 10:44 15 February 2016 (PT)

Cliff, A.D. and Ord, J.K. (1981), Spatial Processes, Pion, London.
Crespo, R., Fotheringham, A.S. and Charlton, M.E. (2007), “Application of geographically
weighted regression to a 19-year set of house price data in London to calibrate local
hedonic price models”, Proceedings of the 9th International Conference on
Geocomputation, September 2007, Maynooth, Ireland.
Dubin, R.A. and Sung, C.H. (1987), “Spatial variation in the price of housing: rent gradients in
non-monocentric cities”, Urban Studies, Vol. 24, pp. 193-204.
Eckert, J.K. (1990), Property Appraisal and Assessment Administration, International Association
of Assessing Officers, Chicago, IL.
Evans, A.W. (1973), The Economics of Residential Location, Macmillan, London.
Farber, S. and Yates, M. (2007), “A comparison of localised regression models in a hedonic house
price context”, Canadian Journal of Regional Science, Vol. 29 No. 3, pp. 405-20.
Follain, J.R. and Malpezzi, S. (1980), Dissecting Housing Value and Rent, The Urban Institute,
Washington, DC.
Foody, G.M. (2004), “Thematic map comparison: evaluating the statistical significance of
differences in classification accuracy”, Photogrammetric Engineering and Remote Sensing,
Vol. 70, pp. 627-33.
Fotheringham, A.S. and Brunsdon, C. (1999), “Local forms of spatial analysis”, Geographical
Analysis, Vol. 31, pp. 340-58.
Fotheringham, A.S., Brunsdon, C. and Charlton, M.E. (2002), Geographically Weighted
Regression: The Analysis of Spatially Varying Relationships, Wiley, Hoboken, NJ.
Fotheringham, A.S., Charlton, M.E. and Brunsdon, C. (1998), “Geographically weighted
regression: a natural evolution of the expansion method for spatial data analysis”,
Environment and Planning A, Vol. 30, pp. 1905-27.
Galster, G. (1996), “William Grigsby and the analysis of housing submarkets and filtering”,
Urban Studies, Vol. 33 No. 10, pp. 1797-806.
Goodman, A.C. and Thibodeau, T.G. (1998), “Housing market segmentation”, Journal of Housing
Economics, Vol. 7, pp. 121-43.
Greaves, M. (1984), “The determinants of residential values: the hierarchical and statistical
approaches”, Journal of Valuation, Vol. 3, pp. 5-23.
Grigsby, W. (1963), Housing Markets and Public Policy, University of Pennsylvania Press,
Philadelphia, PA.
Huang, B., Wu, B. and Barry, M. (2010), “Geographically and temporally weighted regression for
modelling spatio-temporal variation in house prices”, International Journal of Geographical
Information Science, Vol. 24 No. 3, pp. 383-401.
Hutcheson, G.D. and Sofroniou, N. (1999), The Multivariate Social Scientist: Introductory
Statistics Using Generalized Linear Models, Sage, Thousand Oaks, CA.
 Kain, J.F. and Quigley, J.M. (1975), Housing Markets and Racial Discrimination, National Bureau A determinant
of Economic Research, New York, NY.
of house price
Kaiser, H.F. (1974), “An index of factorial simplicity”, Psychometrika, Vol. 39, pp. 31-6.
Lancaster, K.J. (1966), “A new approach to consumer theory”, Journal of Political Economy,
Vol. 40, pp. 2207-21.
McMillen, D.P. and Redfearn, C. (2010), “Estimation and hypothesis testing for nonparametric
hedonic house price functions”, Journal of Regional Science, Vol. 50, pp. 712-33. 69
Malpezzi, S. (1998), “Welfare analysis of rent control with side payments: a natural experiment in
Downloaded by FLINDERS UNIVERSITY OF SOUTH AUSTRALIA At 10:44 15 February 2016 (PT)

Cairo, Egypt”, Regional Science and Urban Economics, Vol. 28 No. 6, pp. 773-96.
Malpezzi, S. (2003), “Hedonic pricing models: a selective and applied review”, Housing Economics
and Public Policy: Essays in Honour of Duncan Maclennan.
Michaels, R. and Smith, V. (1990), “Market segmentation and valuing amenities with hedonic
models: the case of hazardous waste sites”, Urban Economy, Vol. 28, pp. 223-42.
Miles, D. (1994), Housing, Financial Markets and the Wider Economy, Wiley, New York, NY.
Moore, J.W. and Myers, J. (2010), “Using geographic-attribute weighted regression for CAMA
modelling”, Journal of Property Tax Assessment & Administration, Vol. 7 No. 1, pp. 5-29.
Muth, R. (1960), “The demand for non-farm housing”, in Harberger, A.C. (Ed.), The Demand for
Durable Goods, University of Chicago Press, Chicago, IL.
O’Sullivan, A. and Gibb, K. (2003), Housing Economics and Public Policy, Blackwell/RICS, Oxford.
Orford, S. (1999), Valuing the Built Environment: GIS and House Price Analysis, Ashgate,
Aldershot.
Pavlov, A. (2000), “Space varying regression coefficients: a semi-parametric approach applied to
real estate markets”, Real Estate Economics, Vol. 28, pp. 249-83.
Pendleton, E.C. (1965), “Statistical inference in appraisal and assessment procedures”,
The Appraisal Journal, Vol. 37, pp. 501-12.
Ridker, M., Ronald, G. and Henning, J.A. (1967), “The determinants of residential property values
with special reference to air pollution”, Review of Economics and Statistics, Vol. 49, May,
pp. 246-57.
Rosen, S. (1974), “Hedonic prices and implicit markets: product differentiation in pure
competition”, The Journal of Political Economy, Vol. 82, pp. 34-55.
Rothenberg, J., Galster, G., Butler, R. and Pitkin, J. (1991), The Maze of Urban Housing Markets,
University of Chicago Press, Chicago, IL.
Sirmans, G.S., Macpherson, D.A. and Zietz, E.N. (2005), “The composition of hedonic pricing
models”, Journal of Real Estate Literature, Vol. 13 No. 1, pp. 3-43.
So, H.M., Tse, R.Y.C. and Ganesan, S. (1997), “Estimation the influence of transport on house
prices: evidence from Hong Kong”, Journal of Property Valuation & Investment, Vol. 15
No. 1, pp. 40-7.
Stegman, M.A. (1969), “Accessibility models and residential location”, Journal of America
Institute of Planners, Vol. 35, pp. 22-9.
Stevens, M. (2002), Applied Multivariate Statistics for the Social Sciences, 4th ed., Lawrence
Erlbaum, Hillside, NJ.
Straszheim, M. (1975), An Econometric Analysis of the Urban Housing Market, National Bureau
of Economic Research, New York, NY.
Thériaut, M., Des, R.F., Villeneuve, P. and Kestens, Y. (2003), “Modelling interactions of location
with specific value of housing attributes”, Property Management, Vol. 21, pp. 25-48.
 JFMPC Tobler, W. (1970), “A computer movie simulating urban growth in the Detroit region”, Economic
Geography, Vol. 46 No. 2, pp. 234-40.
17,1
UU HPI (2008), University of Ulster House Price Index, Quarter 4.
Xu, T. (2008), “Heterogeneity in housing attribute prices: a study of the interaction behaviour
between property specifics, location coordinates and buyers’ characteristics”, International
Journal of Housing Markets and Analysis, Vol. 1 No. 2, pp. 166-81.
70 Yu, D.L. (2006), “Spatially varying development mechanisms in the greater Beijing area:
a geographically weighted regression investigation”, The Annals of Regional Science,
Downloaded by FLINDERS UNIVERSITY OF SOUTH AUSTRALIA At 10:44 15 February 2016 (PT)

Vol. 40 No. 1, pp. 173-90.

Yu, J., de Jong, R. and Lee, L.F. (2007), “Quasi-maximum likelihood estimators for spatial
dynamic panel data with fixed effects when both n and T are large: a nonstationary case”,
working paper, The Ohio State University, Columbus, OH.

Further reading
Anselin, L. (1988), Spatial Econometrics: Methods and Models, Kluwer, Dordrecht.
Anselin, L. (1998), “GIS research infrastructure for spatial analysis of real estate markets”,
Journal of Housing Research, Vol. 9, pp. 113-33.
Arimah, B.C. (1992), “Hedonic prices and the demand for housing attributes in a third world city:
the case of Ibadan, Nigeria”, Urban Studies, Vol. 29 No. 5, pp. 639-51.
Ball, M.J. (1973), “Recent empirical work on the determinants of relative house prices”, Urban
Studies, Vol. 10, pp. 213-33.
Borukhov, E., Ginsberg, Y. and Werczberger, E. (1978), “Housing prices and housing preferences
in Israel”, Urban Studies, Vol. 15, pp. 187-200.
Brunsdon, C., Fotheringham, S. and Charlton, M.E. (1998), “Geographically weighted regression –
modelling spatial non-stationarity”, Journal of the Royal Statistical Society: Series D, Vol. 47
No. 3, pp. 431-49.
Brunsdon, C., McClatchey, J. and Unwin, D. (2001), “Spatial variations in the average rainfall
altitude relationship in Great Britain: an approach using geographically weighted
regression”, International Journal of Climatology, Vol. 21 No. 4, pp. 455-66.
Can, A. (1992), “Specification and estimation of hedonic housing price models”, Regional Science
and Urban Economics, Vol. 22, pp. 453-74.
Fan, G.-Z., Ong, S.E. and Koh, H.C. (2006), “Determinants of house price: a decision tree
approach”, Urban Studies, Vol. 43 No. 12, pp. 2301-15.
Fotheringham, A.S. and Pitts, T.C. (1995), “Directional variation in distance decay”, Environment
and Planning A, Vol. 27 No. 5, pp. 715-29.
Fotheringham, A.S., Charlton, M.E. and Brunsdon, C. (1996), “The geography of parameter space:
an investigation of spatial non-stationarity”, International Journal of Geographical
Information Science, Vol. 10, pp. 605-27.
Fotheringham, A.S., Brunsdon, C. and Charlton, M.E. (2000), Quantitative Geography:
Perspectives on Spatial Data Analysis, Sage, Thousand Oaks, CA.
Fotheringham, A.S., Charlton, M.E. and Brunsdon, C. (2001), “Spatial variations in school
performance: a local analysis using geographically weighted regression”, Geographical
& Environmental Modelling, Vol. 5 No. 1, pp. 43-66.
Gao, X., Asami, Y. and Chung, C.J.F. (2006), “An empirical evaluation of spatial regression
models”, Computers and Geosciences, Vol. 32 No. 8, pp. 1040-51.
 Goodman, A.C. (1978), “Hedonic prices, price indices, and housing markets”, Journal of Urban A determinant
Economics, Vol. 5 No. 4, pp. 471-84.
Haas, G.C. (1922), “A statistical analysis of farm sales in blue earth county, Minnesota, as
of house price
a basis for farm land appraisal”, Master’s thesis, The University of Minnesota,
Minneapolis, MN.
Hamid, A. (2008), “Combining geographic information systems and regression models to
generate locational value residual surfaces in the assessment of residential”, Pacific Rim 71
Property Research Journal, Vol. 13 No. 1, pp. 35-62.
Huang, B., Zhang, L. and Wu, B. (2009), “Spatio-temporal analysis of rural-urban
Downloaded by FLINDERS UNIVERSITY OF SOUTH AUSTRALIA At 10:44 15 February 2016 (PT)

land conversion”, International Journal of Geographical Information Science, Vol. 23,

pp. 379-98.
Leung, Y., Mei, C.L. and Zhang, W.X. (2000), “Statistical tests for spatial non-stationarity based
on the geographically weighted regression model”, Environment and Planning A, Vol. 32,
pp. 9-32.
Lochl, M. and Axhausen, K.W. (2010), “Modelling hedonic residential rents for land use and
transport simulation while considering spatial effects”, The Journal of Transport and
Land-Use, Vol. 3 No. 2, pp. 39-63.
Long, F., Pàez, A. and Farber, S. (2007), “Spatial effects in hedonic price estimation:
a case study in the city of Toronto”, Working Paper 020, June 2007, Centre for Spatial
Analysis.
Maclennan, D. (1982), Housing Economics Harlow, Longmans, Essex.
Matthews, S. (2007), Geographically weighted regression publication listing, R25 Advanced Spatial
Analysis Training Program, Pennsylvania State University Population Research Centre,
available at: www.pop.psu.edu/gia-core/litsearches/SAM_GWR_list.pdf
McCluskey, W.J., Deddis, W.G., Lamont, I. and Borst, R.A. (2000), “The application of surface
generated interpolation models for the prediction of residential property values”, Journal
of Property Investment & Finance, Vol. 18 No. 2, pp. 162-76.
McMillen, D.P. and McDonald, J.F. (1997), “A nonparametric analysis of employment density in a
polycentric city”, Journal of Regional Science, Vol. 37, pp. 591-612.
Pàez, A., Uchida, T. and Miyamoto, K. (2002), “A general framework for estimation and
inference of geographically weighted regression models: location-specific kernel
bandwidths and a test for local heterogeneity”, Environmental and Planning A, Vol. 34,
pp. 733-54.
Páez, A., Long, F. and Farber, S. (2007), “Moving window approaches for hedonic price
estimation: an empirical comparison of modelling techniques”, Urban Studies, Vol. 45
No. 8, pp. 1565-81.
Sirmans, G.S., Sirmans, C.F. and Benjamin, J.D. (1989), “Determining apartment rent: the value of
amenities, services and external factors”, Journal of Real Estate Research, Vol. 4 No. 2,
pp. 33-54.
Song, S. (1994), “Determination of bargaining outcome in single-family housing transactions: an
empirical examination”, Urban Studies, Vol. 32, pp. 605-14.
Song, S. (1998), “Home buyers’ characteristics and selling prices”, Applied Economics Letters,
Vol. 5, pp. 11-14.
Tu, Y., Hua, S. and Shi-Ming, Y. (2007), “Spatial autocorrelations and urban housing market
segmentation”, Journal of Real Estate Finance & Economics, Vol. 34, pp. 385-406.
Wang, P. (2006), “Exploring spatial effects on housing price: the case study of the City of
Calgary”, Master dissertation, University of Calgary, Alberta.
 JFMPC Appendix. Principal component analysis
17,1
4

3
72

Eigenvalue
2
Downloaded by FLINDERS UNIVERSITY OF SOUTH AUSTRALIA At 10:44 15 February 2016 (PT)

0
Figure A1. 1 2 3 4 5 6
Scree plot of eigenvalues
Component Number

1.0 nfa
beds
receptions

0.5
Component 2

Bus_Distan
CBD_Dist
0.0 Train_Dist

–0.5

Figure A2. –1.0

Component plot in –1.0 –0.5 0.0 0.5 1.0
rotated space
Component 1

Corresponding author
P.T. Davis can be contacted at: [email protected]

To purchase reprints of this article please e-mail: [email protected]

Or visit our web site for further details: www.emeraldinsight.com/reprints
 This article has been cited by:

1. Mohd Faris Dziauddin, Mustika Misran. 2016. Does Accessibility to the Central Business District (CBD)
Have an Impact on High-Rise Condominium Price Gradient in Kuala Lumpur, Malaysia?. SHS Web of
Conferences 23, 02004. [CrossRef]
Downloaded by FLINDERS UNIVERSITY OF SOUTH AUSTRALIA At 10:44 15 February 2016 (PT)