Spatial Analysis of House Price Determinants

1
SPATIAL ANALYSIS OF HOUSE PRICE DETERMINANTS: A
GREATER LONDON CASE STUDY
LAURENT J. L. SANTOS1
RUI JIANG1
December 2018
1
Department of Civil, Environmental & Geomatic Engineering,
University College London | UCL, UK
e-mail: laurent.santos.18@ucl.ac.uk and rui.jiang.18@alumni.ucl.ac.uk
How prices are determined by preferences and constrains is central to understand market trends and
patterns. This study aims to investigate the relationship between a range of explanatory variables and
house prices in greater London. More precisely, the research applies Exploratory Data Analysis and
Spatial Regression methods to investigate factors affecting house prices in greater London. To what
extent are housing prices determined by better health services and environmental areas? How does
distance to underground stations and schools affect property prices? To what extent are housing
values shaped by society income areas? To address these questions, a locally based spatial analysis is
presented, and local estimates of regression parameters area investigated. The strategy of inquiry
places emphasis on the structure of the housing market and the interrelationships of the variables that
influence, directly or indirectly the functioning of the market. The study demonstrates that different
variables have both positive and negative impacts in house prices across space.
1. INTRODUCTION
The focal point of this study is to investigate the relationship between a range of explanatory variables
and house prices in greater London. More precisely, the study stems from Exploratory Data Analysis
and Spatial Regression methods to investigate factors affecting house prices in greater London.
Over the last past decades, housing markets as an area of interest has long drawn attention given the
causes and consequences of macroeconomics events as the great financial crash, drivers (for instance,
labour, capital and material markets), but also for microeconomic theories (e.g. consumer decision
analysis) and urban planning studies (e.g. migration and mobility). Housing markets is a broad topic
of study with a wide spectrum of applications.
The mathematical modelling of spatial processes is useful in providing information on the
determinants of those processes through the estimate of variables. It also provides a framework to
predict spatial impacts of various actions and under different scenarios. To this end, a locally based
spatial analysis is presented, a technique derived for producing local estimates of regression

2
parameters which can be used to produce maps from which a geography of spatial relationships can
be examined.
Spatial analysis is a type of geographical inquiry to explain patterns of human behaviour and its spatial
expression in terms of mathematics and geometry. It also can be understood as a process that you can
model problems geographically, derive results by computer processing, and then explore and examine
results.
Spatial analysis provides a unique set of techniques and methods for analysing events that are located
in geographical space. By means of quantitative analysis, it seeks to predict the spatial patterns of
some phenomena, as for example the behaviour of property prices within a city.
2. LITERATURE REVIEW
2.1 Housing Markets
In academia or government, housing market is contented upon two broad frameworks: 1.
macroeconomics and 2. microeconomics research. The former aims to analyse the drivers that might
lead to economic outcomes. The latter is focused on consumer decision by modelling personal
preferences and constraints (e.g. income).
Macroeconomics emphasises the context that somehow leads to market behaviours. From one side,
the variables of interest are focused on the drivers that may affect pricing mechanisms. In this context,
variables as income, taxes and interest rates are associated to housing outcomes (Maclennan, 2012).
Such analysis also considers environmental statistics as labour, capital markets and the business of
constructing to predict market outcomes.
In microeconomics, consumer preferences and decision mechanisms are the core variables that must
be analysed. Housing markets are led by processes of search and matching (Maclennan, 2012). Prices
fluctuate in order to bring supply and demand into economic equilibrium. Price determination is the
outcome of rational choices made by individual agents (i.e. demand) and production (i.e. supply)
(Centre for Co-operation with European Economies in Transition, 1993). Figure 1.1 depicts the
traditional supply and demand curve model.
In this context, housing is regarded as a composite commodity with attributes that are valued by the
decision maker (the purchaser) in order to satisfy a range of preferences. This process is known as
consumer-utility function and assumes that buyers make informed decisions based on a matrix of
attributes, options and risks throughout a specific spatial search process.
Economics research has been good at establishing the existence of the complexity of housing as a
good. House price studies have been used to identify the economic significance of different,
distinctive attributes of housing (a good literature review can be found in Maclennan 2012). These
studies, that almost invariably have high levels of explanatory power, lead to some important
conclusions and confirm that housing prices are influenced by attributes or characteristics such as:
• Size, style, layout and internal amenity (variety).
• The location of the dwelling: households pay not just for size, type, quality but for the
characteristics of the location: + the costs of accessibility to the wider spread of locations
used by household (such as foot instance, employment), + shopping and leisure locations: the
quality and availability of neighbourhood amenity including neighbours, + access to local
retail and service facilities.
• The asset importance of their home and possibilities for (relative) gain and loss as well as

3
quality and maintenance obligations (fixity and durability).
All these theories have a very similar moral and socio-political grounding, namely they assume that
“people are rational, utility-maximising decision makers and that economic activity takes place in
freely competitive, equilibrium-seeking contexts or settings” (Brown, 1993, p. 186).
Home and neighbourhood are an ineluctably conjoined choice. This jointness not only adds to the
variety of attributes that have to be considered but also adds other distinctive aspects to housing
choice.
2.2 Geospatial Data Analysis
Geospatial analysis can be defined as the process of examining the locations, attributes, and
relationships of features within a spatial scale. We may better refer the techniques covered in this
book as spatial rather than geographic because they are applied to data arrayed in any space, not only
upon geographic space.
The methods of geospatial analysis provide a better understanding of spatial problems, offering a
special perspective to examine events, patterns and processes with spatial variables. Another
perspective is offered by De Smith (2007) who content that ultimately, geospatial analysis concerns
what happens where and makes use of geographic information that links features and phenomena on
the Earth’ s surface to their locations. Any analysis carried out without knowledge of location is
definitively not a spatial analysis.
Before diving in a more profound toolset of quantitative geography, it makes sense to introduce the
main elements of geospatial analysis.
• Attributes are the recorded characteristics or property of a place. The simplest type of
attribute, termed nominal, is one that serves to identify one property from another. For
example, the number of a property or house serves to identify that instance of a class of
entities, and to distinguish it from other members of the same class.
• Vector data is the conceptual view of discrete objects. In a geographic scale, the world is
represented using points, lines, and polygons. These data are created by digitizing the base
data. They store information in (xi,yi) coordinates. Vector models are useful for storing data
that has discrete boundaries, with country borders, land parcels, and streets.
• Density estimation is one of he most useful concepts for spatial analysis. It is a quantitative
measure to address a context, to what extent events or properties at some location are related
to the location’ s surroundings. The density expresses the number of discrete objects per unit
area.
• Spatial scale or simply scale refers to the order of magnitude of extent or size of a land area
or geographical distance studied or described. It also used in the sense of spatial resolution,
or the level of spatial detail in data.
2.3 Geospatial Data Analysis
The idea of using models in problem analysis is really not new. For complex problems, visual and
mathematical approaches might be used to study the relationships, to describe or represent an object
or a phenomenon. A model is a simplified representation of reality, a mathematical expression that

4
accurately represents the relevant characteristics of the object or problem being
studied. (Ragsdale 2004).
Models allows us to gain insight and understanding examining complex problems under investigation.
The modelling techniques in this book use mathematics to describe and explain a spatial pattern. But
before dealing with spatial modelling techniques, we have to examine three categories of modeling
techniques.
The first category is known as prescriptive models. In some situations, we face a problem involving
a very precise, well-defined functional relationship f() between the independent variable x1,x2,…xk
and the dependent variable Y. These types of models are called prescriptive models because their
solutions tell the decision maker what action to take.
A second category of problems is one in which the objective is to predict or estimate what value the
dependent variable Y will take on when the independent variables x1,x2,…xk take on specific values.
If the function f() relating the dependent and independent variables is known, we might simply enter
the specific values for x1,x2,…xk into the function f() and compute the value of Y. These types of
models are called predictive models. For example, a real estate appraiser might know that the value
of a house (Y) is influenced by several variables as proximity to the underground (x1) and parks (x2)
but the functional relationship f() that relates these variables to one another might be unknown. By
analysing the relationship between the dependent variables, the appraiser might be able to identify a
function f() that relates these variables in a reasonably accurate manner.
The third category of models are called descriptive models. In these situations, a problem has a very
precise, well-defined functional relationship f() between the independent x1,x2,…xk and the dependent
variable Y. However, there might be great uncertainty as to the exact values that will be assumed by
one or more of the independent variables. In these types of problems, the objective is to describe the
outcome or behaviour of a given operation or system.
In this research, various mathematical models are applied to study a specific spatial problem: how
housing prices are determined by health, transport, school and income within the greater London
authority boundary. Essentially, predictive techniques are applied to tackle a spatial analysis problem.
3. EXPERIMENT DESIGN AND METHODOLOGY
The process of spatial analysis generally follows a number of stages: problem formulation, data
gathering, exploratory analysis, modeling and testing and, ultimately, final reporting of the findings.
In this section, the methodologies applied in the study are introduced. The framework of the
experiments and the hypotheses that will be tested are first reviewed.
The following subsections are focused on geospatial statistical modeling approaches: exploratory
spatial data analysis, spatial regression and testing. The final section of this chapter is dedicated to
the description of the data applied in the experiment.
3.1 Analytical Methods

5
Spatial analysis consists of the analysis of numerical spatial data and the construction and testing of
mathematical model of spatial processes. The goal of all these activities is to add to our understanding
of spatial processes (Fotheringham, 2000).
The strategy of inquiry places emphasis on the structure of the housing market and the
interrelationships of the variables that influence, directly or indirectly the functioning of the market.
The strategy of inquiry that will be developed is based on a framework decomposed in three main
sections. Figure 3.1 outlines the line of inquiry.
FIGURE 3.1 Outline of Applied Methods
A first step is to examine how data vary in space and explore their general characteristics. Explanatory
data analysis aims to examine the relationships between variables to test causal relationships. The
autocorrelation indices of spatial heterogeneity in house prices across London have to be calculated
and measured.
The second part of the experiment is dedicated to spatial regression modelling. Spatial regression is
an extension of linear regression to deal with the properties of spatial data, such as autocorrelation
and heterogeneity. A regression model is built and accounts for the autocorrelation of the dependent
variable through an additional parameter or set of parameters. With the best regression fit, spatial
heterogeneity analysis can then be performed to identify non-uniformity phenomena in the dataset.
3.2 Quantitative Spatial Data Analysis
Exploratory data analysis (EDA) is a strategy to analyse datasets and explore their characteristics and
generate hypotheses. The approach has its roots in the work of American mathematician John Tukey
in the 1960s (see Tukey 1962). It emphasises graphical methods such as histograms, box-plots and
scatterplots.
Exploratory spatial data analysis (ESDA) is the extension of EDA to spatial data. A key aspect of
ESDA is the identification of patterns via data visualisation and statistics. It implies using techniques
to summarize and visualise the data in search of meaningful patterns and to identify problematic and
extraordinary data called outliners.
According to T. Cheng (2019), among the objectives of ESTDA, are included:
• maximising insight into a dataset
• uncovering underlying structure
• extracting important variables
• detecting outliers and anomalies
• testing underlying assumptions
3. Exploratory
spatial data
analysis
2. Spatial
regression
1. Analysis and
conclusions

6
• developing parsimonious models (simplest model that can achieve a certain performance
level).
The most widely known way to explain a pattern in a dataset is a technique known as correlation. In
spatial analysis, spatial autocorrelation aims to examine the magnitude of a variable at one location
correlates with nearby locations (Cliff, 1981). The magnitude of the effects can be measured using a
number of statistics which describes spatial variation in terms of a function that shows how spatial
autocorrelation decreases with increasing distance. This core concepts have been captured by Tobler
(1970) in what has become known as his First Law of Geography: “All thinks are related, but nearby
things are more related than distant things” (p.235).
Testing for spatial autocorrelation and incorporating it into a model requires a formal specification of
the spatial structure. This is done through a spatial weight matrix and spatial lags. A spatial weight
matrix is a nxn square matrix, designated as W , where each element wij describes the strength of the
spatial relationship between each of the n spatial observations. The weight matrix embodies the
analyst’ s knowledge and assumptions about the study area and data within the model.
A distance weight matrix considers the distance between observations when defining the spatial
structure. It works well with point data since precise distances can be captured. The weight matrix
captures the strength of the relationships, so two points that are close in space should have a larger
weight than those further apart, typically going to zero at some cutoff point (Singleton, 2017).
The spatial weight matrix formalizes the neighbourhood around each observation. A spatial lag is
created by multiplying the weight matrix by a variable measured for each observation.
3.3 Spatial Regression
In the study, spatial regression models are used to explore the effect of several explanatory variables
on house price data in London by taking spatial autocorrelation and heterogeneity into account.
Spatial regression is an extension of linear regression to deal with the properties of spatial data, such
as autocorrelation and heterogeneity.
Through the regression approaches, including linear regression and spatial regression methods, a
series of explanatory variables are applied used in the analyses:
• retail environment
• health services accessibility
• physical environment
• distance to undergrounds
• distance to schools
• annual household income.
These explanatory variables are utilised to explain and predict the dependent variable, the mean value
of the house price in London. Four variables have been log transformed in all models. Replacing the
data with the log aims to remove the skew and normalise the data, making the pattern more visible.
When dealing with spatial data, conventional multiple linear regression is often not enough to account
for the presence of spatial regression. After the residual autocorrelation has been tested, the spatial
regression models are used to fit the model.

7
Three widely used spatial regression models (Anselin, 2010) will be presented:
• spatial lag model
• spatial error model
• spatial Durbin model
These techniques are going to be applied by considering a spatial lag term, a spatial error term and a
lagged term of the independent variables.
3.3.1 Multiple Linear Regression
As one of the most widely used and simplest statistical methods, multiple linear regression is
conducted to summarize and explain the relationships between dependent variable and independent
variables based on their statistical relationships.
Scatterplots between each independent variable and the dependent variable are created to verify
associations between the variables which may indicate significant relationships between house prices
and other factors.
Also, prior to linear regression, multicollinearity must be tested to ensure the reliability of the model.
Variance inflation factor (VIF) and Kappa statistic are used to diagnose multicollinearity. Both
statistics aimed to explore the correlation among independent variables. VIF, the reciprocal of the
tolerance, quantifies the severity of multicollinearity by measuring the variance in a model in the
presence of multiple terms and one single term1
while Kappa, the condition number, explains the
multicollinearity by calculating the eigenvalues of the variables (Belsley, 2004).
In general, a value of VIF less than 5 or a Kappa statistic less than 1000 indicates that it is safe to
ignore multicollinearity. When a VIF goes above 5 and Kappa goes above 1000, it can be assumed
that the regression coefficients are modelled incorrectly and requiring correction. In this case, the
independent variable with a high VIF/Kappa should be removed from the regression model, or Partial
Least Square Regression would be used instead.
3.3.2 Spatial Autocorrelation in Regression Errors
When working with spatial data, it is important to test for spatial autocorrelation in the residuals of
linear models. There are a number of tests that can be used. We examine some of them in the following
sections.
3.3.2.1 Spatial Autocorrelation in Ordinary Least Squares (OLS) Regression Errors
Although the association between covariates must be examined and excluded by conducting
multicollinearity tests, spatial correlations among the variable itself may still exist. When dealing
1
For more details, see James, 2013.

8
with spatial variables, it is essential to test the regression errors after the model been initially fitted in
the linear model due to the independency of the explanatory variables in terms of their spatial
correlations.
Therefore, to avoid spatial model misspecification, it is critically necessary to calculate the spatial
autocorrelation in the residuals of multiple linear regression model by employing spatial test models.
Based on the assumption that the residual errors are normally distributed, different approaches of
diagnostics are adopted to quantify spatial dependence in the presence of covariates (Darmofal, 2015).
Generally, there are two types of tests, unfocused diagnostics and focused diagnostics. Unfocused
diagnostics include Moran’s I and Kelejian-Robinson (KR) diagnostics, which simply measure the
presence of spatial dependence. On the contrary, focused diagnostics apply the effect of spatial lag or
spatial error dependence on the present dependence as well as LM test for spatial lag dependence and
LM test for spatial error dependence.
3.3.2.2 Moran’s I
Moran’s I stands as the most common diagnostic approach for regression residuals.
Unlike the usual Moran’s I test which measures autocorrelation in spatial data, Moran’s I for residual
spatial autocorrelation assumes that all the variables are the residuals from a linear regression.
3.3.2.3 Lagrange Multiplier Test for Spatial Lag Dependence and Error Dependence
Lagrange Multiplier (LM) test is usually used for situation where several sources of misspecification
are considered. Specifically, the test is applied for situations in which spatial dependence in the form
of an omitted spatially lagged variable or spatial autocorrelation are presented.
3.3.2.4 Spatial Lag Model
Spatial lag model assumes that direct interactions between observations at near locations exist (Ward,
2008). The model is purely autoregressive and posits that autocorrelated dependent variable causes
the presence of spatial autocorrelated residuals. The model can be expressed as the following
equation:
where ρ represents the spatial autocorrelation parameter and W represents the row-sum standardized
spatial weight matrix.
3.3.2.5 Spatial Error Model
The spatial error model assumes that autocorrelation exists in the errors as a result of unobserved
variables. The model can be expressed as the following equation:

9
where γ is the spatial autocorrelation parameter, W is the row-sum standardized spatial weight matrix
and δ is a vector of the error terms.
3.3.2.6 Spatial Durbin Model
Based on spatial lag model, the spatial Durbin model utilises both the explanatory variables and
predictor variable as its autoregressive terms.
where θ is a vector of parameters.
3.4 Geographically Weighted Regression
The global regression model (GWR) estimates a constant parameter for the relationship between each
independent variable and the dependable variable across the study area.
GWR is a technique that allows local, rather than global, parameters to be estimated. It recognizes
that spatial variations might exits and should be measured. That is, observed data near to point i have
more influence than those located farther than i (Fotheringham, 2000).
In this model, the equation measures the relationship around each point i. Weighted least square
estimation provides the basis to understanding how GWR operates. It attempts to model heterogeneity
using geographically varying regression coefficients. This enables maps of the coefficients to be
produced, providing a better understanding of the relationship between the dependent and
independent variables across space.
The general form of the GWR model is:
Where: βi0 is the intercept term at location i, i=1,2,...,Ni=1,2,...,N; N is the number of spatial
locations; βip is the value of the pth
parameter at location i, p=1,2,...,m; m is the number of
independent variables; ϵi is a random error.
3.5 Data Collection and Description
Economic studies content the existence of a broad range of attributes for housing. In his work,
Maclennan (2012) recognizes this and suggests that they have high levels of “explanatory power” (p.

10
7), confirming that prices are influenced by a spectrum of characteristics and factors dictated by
housing characteristics, location and access to local retail and service facilities. Quite in line with this,
Malpezzi (2002) asserts that the value of houses correlate with different attributes.
That said, it can be suggested that house pricing is a direct application of the economic significance
of different factors, ranging from internal attributes as well as locational characteristics (e.g. services
facilities, public transport, neighborhood, etc.). Maclennan (2012) points to the fact that space effects
interact with variety in shaping the dynamics of housing markets: “households pay not just for size,
type, quality but for the characteristics of the location” (p. 7).
A fundamental feature of housing is the availability of underground station nearby. Indeed, many
studies suggest a positive relationship between house prices and proximity to rail station. In their
work, Cervero et al. (2004) found a 6.4% to 45% increase in the house prices near to rail station. This
claim is also supported by Debrezion et al. (2007) who found that for every 250m closer to the station
is worth 2.4% of the house price.
By the same extent, living close to school is an asset for couples with children. Their life course is
particularly salient when living nearby school facilities in the sense that it turns to be an important
attribute for housing search. In this context, Clark and Davies Withers (2007) argue that there is higher
demand for schools and facilities when young families move into a community.
The access to hospitals, has also been pointed out as an important component for house prices. Clark
(2012) argues that housing prices is seen as a function of demands for associated services. The quality
and availability of health services might lead to a positive relationship to house prices.
Additionally, Zhang and Yi (2018) argue that house prices increase significantly as a result of a
willingness to pay for a pleasing environment, that is, zones with low traffic congestion,
environmental pollution and availability of leisure areas. In this line of reasoning, environmental
health might also present a positive effect in housing markets.
A further component affecting housing choices is the association with neighbourhood attributes. In
this sense, Maclennan suggests that as neighbourhood attributes are often associated with “social
situational good” (2012, p. 9). Butler and Hamnett (2012) explain that residential patterning presents
strong differences in the geography of income and social class. Income and housing prices might also
be positively correlated.
Taking into account this brief review, Table 3.1 details the description and sources of the data applied
to the study.

11
TABLE 3.1 Data Description
Category Data base Description Type Source
Geographical
scale
Maintainer Data range
Pricing House price data
Price Paid Data includes information
on all property sales in England and
Wales that are sold for full market
value and are lodged with us for
registration.
Continuou
s data
https://www.gov.uk/government/statistical
-data-sets/price-paid-data-downloads
Point data
HM Land
Registry
Mar-17
Education Education List of London schools.
Location
data
https://data.london.gov.uk/dataset/london-
schools-atlas
Point data
Consumer Data
Research Centre
2017
Environmental
health
Environmental
domain
London environmental domains
Discrete
data
https://data.cdrc.ac.uk/dataset/access-to-
healthy-assets-and-hazards-
ahah/resource/dd146935-8181-4cef-95cf-
68e11b47c1ae
Lower layer super
output areas
(LSOA)
Consumer Data
Research Centre
2017
Health Services
Health Services
domain
London health services.
Location
data
https://data.cdrc.ac.uk/dataset/access-to-
healthy-assets-and-hazards-
ahah/resource/dd146935-8181-4cef-95cf-
68e11b47c1ae
Lower layer super
output areas
(LSOA)
Consumer Data
Research Centre
2017
Transport
Underground
stations
List of London stations with their
Ordnance Survey coordinates.
Include all tube stations and mainline
stations.
Location
data
https://github.com/oobrien/vis Point data OpenStreetMap
November,
2018
Income
CDRC Individual
Income Estimates
(PAYE)
Individual income from Pay As You
Earn (PAYE) and benefits.
Discrete
data
https://data.cdrc.ac.uk/dataset/cdrc-2016-
individual-income-estimates-greater-
london-area-geodata-pack-london-
e12000007
Lower layer super
output areas
(LSOA)
Consumer Data
Research Centre
2016

12
4. EXPLORATORY SPATIAL DATA ANALYSIS
To explore the characteristics of the variables, exploratory spatial data analysis (ESDA) was
applied prior to the spatial regression models.
4.1 Non-spatial Characteristics
Table 4.1 summarizes the statistics of variables. From the table, a wide range of values of variables
can be observed. The minimum value of house price is 190722, while the maximum value is more
than four times greater than that. The wide range can also be observed in the distance to
undergrounds, with LSOA area containing undergrounds having a value of 0, while for some
LSOA areas, the distance to the nearest underground is as far as 18847.6 meters. Compared with
undergrounds, schools seem to be more evenly distributed. The maximum distance from a LSOA
area to the nearest school is 753.88 meters, which is about 1% of the maximum distance from
LSOA to the underground. The indexes of retail outlets, health service and physical environments
vary from 0 to 100. The mean annual household income ranges from 25020 to 88090.
TABLE 4.1 Summary of Variables
House price Retail Health Environment Underground School Income
Min. 190722 4.116 0.0049 4.898 0 0 25020
1st Qu 375081 26.962 1.3618 28.692 327.2 0 34130
Median 483722 40.418 3.4774 44.501 970.5 14.52 38440
Mean 671684 42.915 5.5587 47.221 2370.9 62.7 39565
3rd Qu 671684 56.956 7.5203 63.93 3329.7 98.77 43545
Max. 8061563 99.255 46.1394 100 18847.6 753.88 88090
4.2 Spatial Characteristics
Figure 4.1 uncovers wide range values of different variables. Different spatial patterns can be
observed between the central of London and the peripheral areas. In general, variables appear to
be spatially correlated since adjacent LSOA areas tend to have similar values than distant LSOA
areas.
Figure 4.1 also demonstrates that houses with higher property prices cluster in the central area of
London. Together with figure 4.2, it seems that houses with a higher price appear to have a better

13
accessibility to retail outlets and a shorter distance to the nearest underground. However, those
houses tend to be surrounded by worse physical environments, and they appear to have less
accessibility to health services. Also, the distance to schools has less spatial patterns and does not
present strong correlation with house prices. As expected, annual household income appears to
have a positive correlation with the house prices.
FIGURE 4.1 Map of House Prices

14
FIGURE 4.2 Maps of (from top to bottom, from left to right): 1. Accessibility to Retail Outlets and 2.
health service, 3. Quality of Physical Environment; 4. Annual Household Income; 5. Distance to School
and 6. To Underground stations.
4.3 Global Spatial Autocorrelation
As Tobler (1970) mentioned, everything tends to be related to everything else, but things close to
each other tend to be more related.

15
In the study, Moran’s I have been employed to uncover the presence of global spatial autocorrelation
by measuring a single value of autocorrelation level across the study area for each variable.
TABLE 4.2 Outputs of Monte-Carlo Simulation
Monte-Carlo
simulation
Moran’s I
observed
rank
p-value
House Price 0.681 1000 0.001
Retail 0.705 1000 0.001
Health 0.642 1000 0.001
Environment 0.943 1000 0.001
Underground 0.987 1000 0.001
School 0.150 1000 0.001
Income 0.733 1000 0.001
Table 4.2 summarizes the values of calculated Moran’s I, observed rank and p-value for each
variable. The positive values of Moran’s I suggest that all variables are positively correlated to
nearby variables: distance to underground and physical environment present a strong correlation
whereas distance to school has a weaker correlation. The observed ranks are all 100 and p-values
are all less than 0.001, which suggests that the probability of the observed Moran’s I being due to
chance is less than 0.1%. Therefore, it can be concluded that all variables are significantly
positively autocorrelated.
4.4 Local autocorrelation
Due to the presence of spatial heterogeneity, the level of autocorrelation could vary. In this case, a
single value is not enough to explain the variance in the variables. Therefore, local autocorrelation
methods were employed in the study as well.
4.4.1 Moran Scatterplot
The Moran scatterplot on figure 4.3 identifies clusters of high and low values in house price data.
Indicated by the labels on the plot, there are some clusters of extremely high price houses in city
of Westminster and Camden town, while most data are within the ‘Low-Low’ quadrant. The
missing district information is due to the combination of two datasets, with one dataset has less
data.

16
FIGURE 4.3. Moran Scatterplot of House Prices

17
4.4.2 Getis and Ord’s Gi and Gi *
The figures below are the outputs of Getis and Ord’s Gi and Gi*. The figure 4.4 indicates the
distribution of raw Gi* values across the study area. Central London has a higher value of Gi*
compared with peripheral area, which suggests that the house price in central London is more
spatially autocorrelated.
Figure 4.5 displays the distribution of hotspots and coldspots. As expected, the central area of
London stands as a cluster of high property prices while some parts of peripheral areas are
identified as a cluster of relatively low property price, especially east part. No significant clusters,
presented by red colour, are found in most of the areas.
FIGURE 4.4. The Output of Getis and Ord’s Gi* FIGURE 4.5. Hotspots and Coldspots
4.4.3 Local Moran’s I
Areas with significant clusters have been uncovered by the local Moran approach. Consistent with
the results calculated by Getis and Ord’s Gi and Gi *, there is a small area in the central London
with high-high clusters. No low-low clusters have been found in this study area by local Moran’s
I (see figures 4.6 and 4.7).

18
FIGURE 4.6 Local Moran’s I FIGURE 4.7 Map of Significant Clusters
5. RESULTS
5.1 Multiple Linear Regression
To acquire an initial overview of the data, a scatter plot of matrices was created. It helps to visualise
the relationships among each variable. In figure 5.1, the diagonal, bivariate scatter plots with the linear
regression fits (red lines) reveals the statistical relationships between independent and dependent
variables. On the diagonal, histograms are presented, which indicates most of the variable data are
normal distributed after log transformation. Pearson correlation coefficients are presented in the area
above the diagonal. Overall positive associations were found between house price and retail,
environment and income variables, which negative associations were found between house price and
health service, distance to the underground and school.

19
FIGURE 5.1 A Scatter Plot of Matrices
TABLE 5.1 Summary of Multicollinearity Analysis
Variables retail log(health+1) environment log(school+1) log(underground+1) income
VIF 2.96 2.80 1.72 1.01 1.30 1.04
Mean VIF 1.80
Kappa 16.23
Then, multicollinearity analysis was employed on the independent variables (see table 5.1). The
relative low values of VIF and Kappa statistic indicate that there is no severe multicollinearity among
the explanatory variables.
After that, the multiple linear regression was conducted. From the result below, it turns that all
explanatory variables are statistically significant because their p-values are less than the common
alpha value 0.05. The value of multiple R-squared (0.7116) indicates that the multiple linear
regression explains 71.16% of the variance in the data. The statistically significant p-value in F-test
also explains a significant amount of the variance in house price data.

20
FIGURE 5.2 Result of Multiple Linear Regression
5.2 Spatial Autocorrelation in OLS Regression Errors
Figure 5.3 displays the spatial distribution of the residuals of the linear regression model. Spatial
patterns can be observed as positive residuals (in red) tend to cluster together in the north and south-
east where the house prices are overestimated by the fitted model, while negative residuals (in blue)
are clustered together in the east and west London where the house prices are underestimated.
FIGURE 5.3 Residuals of Multiple Linear Regression Model
Residuals:
Min 1Q Median 3Q Max
-1.33031 -0.16122 -0.01249 0.13891 1.51392
Coefficients:
Estimate Std. Error t value Pr(>|t|)
(Intercept) 1.120e+01 3.834e-02 292.279 < 2e-16 ***
retail 7.731e-04 3.176e-04 2.434 0.014954 *
log(health+1) -1.615e-02 7.569e-03 -2.134 0.032913 *
environment 3.740e-03 2.138e-04 17.491 < 2e-16 ***
log(school+1) -5.809e-03 1.673e-03 -3.472 0.000522 ***
log(underground+1) -2.624e-02 2.042e-03 -12.849 < 2e-16 ***
income 4.988e-05 5.377e-07 92.765 < 2e-16 ***
---
Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1
Residual standard error: 0.2563 on 4536 degrees of freedom
Multiple R-squared: 0.7116, Adjusted R-squared: 0.7112
F-statistic: 1865 on 6 and 4536 DF, p-value: < 2.2e-16

21
5.2.1 Moran’s I
To test the spatial autocorrelation in the residuals, the Moran test has been applied. It returns a value
of 0.432 (expected value -0.001), with a p-value of 0.000, which shows that the residuals are
significantly autocorrelated.
5.2.2 LM test for Spatial lag and spatial error dependence
The lag test returns a chi-squared statistic of 104.12 with a p-value of 0.000, while the error test
returns a chi-squared statistic of 1219.8 with a p-value of 0.000. The high values of chi-square statistic
along with the small p-values indicate that both regression models can be used as the regression
models.
Indicated by the higher value of chi-squared statistic in spatial error test, spatial error model might be
used as a better interpretation of the spatial autocorrelation.
TABLE 5.2 Lagrange Multiplier (LM) test for Spatial Dependence.
Test Value DF p-Value
LM lag 104.12 1 0.0000
LM error 1219.8 1 0.0000
5.3 Spatial Regression Models
5.3.1 Global Spatial Regression Models
TABLE 5.3 Summary of Spatial Regression Models
ρ (rho) / Lambda Log-likelihood AIC LM-test
Spatial Lag Model 0.34791*** 253.8128 -489.63 474.71***
Spatial Error Model 0.70226*** 603.4894 -1189
Spatial Durbin Model 0.38217*** 396.0785 -762.16
Note: *** represents the significance level is at 0.001.

22
Table 5.3 displays the summarized results of three spatial regression models. Four spatial parameters
including rho, log-likelihood, AIC and LM-test were selected to compare the performance of different
models.
In all models, the lag factor (rho/Lambda) is always statistically significant with a p-value less than
0.001, which indicates a high level of spatial autocorrelation presents in the house price data.
Therefore, the result of multiple linear regression would be misleading. The LM test for spatial lag
model reveals that significant spatial autocorrelation remains after the model has been fitted.
The log-likelihood and Akaike Information Criterion (AIC) were used to compare the suitability of
each model. The higher value of log-likelihood and the lower negative value of AIC, the better the
model. In the study, a larger log-likelihood and a larger negative AIC reveals that spatial error model
has the best fitness compared with other two models. However, although this model performs best
among three models, spatial autocorrelation still exists and cannot be fully explained by the model.
TABLE 5.4 Significant coefficients in spatial Durbin model
Variable Non-lagged Lagged Significance
Retail No Yes positive
log(Health+1) Yes Yes positive/negative
Environment No No
log(school+1) Yes Yes negative/negative
log(underground+1) No No
Income Yes Yes positive/negative
Figure 5.4 shows the residuals of three spatial models, spatial lag model, spatial error model and
spatial Durbin model respectively. In general, the house price in north London has been overestimated
while east and west London have been underestimated.

23
FIGURE 5.4 Residuals from Spatial Lag Model, Spatial Error Model and Spatial Durbin Model

24
Figure 5.5 presents the fitted house price values calculated by spatial Durbin model and the actual
house price values. By comparing two figures, it can be concluded that spatial Durbin model predicts
the house price well. However, since the model is only a simplified version of the real world, residuals
remain at some extent.
FIGURE 5.5 Fitted values from Spatial Error Model and Observed House Prices in London
5.3.2 Geographically Weighted Regression
In the study, a Gaussian kernel and a Bisquare kernel have been applied when fit the GWR model.
Fixed bandwidths are chosen by leave-one-out cross-validation method for Gaussian and Bisquare
kernel respectively.
Summarized in table 5.5, the values of optimal bandwidth are estimated to be 1243.81m and
3822.40m for the Gaussian kernel and the Bisquare kernel, which in turn result in different effective
numbers of parameters (776.57 and 561.72). Also, a larger negative AIC (-2703.02) and a larger
positive R2 (0.88) in Gaussian kernel indicates that Gaussian kernel performs better compared with
Bisquare kernel.
The result of Moran test of GWR model quantifies the residuals for autocorrelation. A Moran’s I
value of 0.02 and a significant p-value (p<0.000) demonstrate significant spatial autocorrelation
remains in the house price after fitted by a GWR model.

25
TABLE 5.5 Comparison between Gaussian kernel and Bisquare Kernel
Gaussian Bisquare
GWR
Bandwidth 1243.81 3822.40
Effective number of parameters 776.57 561.72
AIC -2703.02 -2388.83
Quasi-global R2 0.88 0.87
Moran test
Residuals 37864.00 33159.00
df 41483.00 37002.00
p-value < 2.2e-16 < 2.2e-16
I 0.02 0.06
TABLE 5.6 Summary of GWR Coefficient Estimates at Data Points by Gauss Kernel.
Min. 1st Qu. Median 3rd Qu. Max. Global
X.Intercept. -5.55e+00 1.13e+01 1.16e+01 1.20e+01 2.10e+01 11.20
retail -2.40e-02 -2.51e-03 -6.15e-04 1.07e-03 2.51e-02 0.00
log.health -2.04e-01 -3.17e-02 2.67e-03 4.77e-02 5.42e-01 -0.02
environment -2.37e-02 -1.74e-03 4.54e-04 2.49e-03 2.67e-02 0.00
log.school. -8.24e-02 -8.50e-03 -1.49e-03 3.92e-03 4.67e-02 -0.01
log.underground. -9.71e-01 -3.03e-02 -6.90e-03 6.10e-03 1.74e+00 -0.03
income 7.30e-06 3.69e-05 4.18e-05 4.71e-05 1.83e-04 0.00
Table 5.6 indicates the distribution of the explanatory variables. The effects of different explanatory
variables on the house price are also visualized in the distribution map in Figure 5.5.

26
FIGURE 5.6 Coefficient Maps
In addition, maps of the coefficients have been produced. As the figures shown, different dependent
variables have different influences on the dependent variable across space. Expect for household
income, both positive and negative impacts are found between different variables and house price
data across space. Figure 4.6a indicates that the accessibility to retail outlets has a positive effect on
the house price in east and west London, while it has a negative impact on the house price in centre
and north the London. Similarly, map of the accessibility to health service (figure 5.6b) reveals a
similar spatial pattern on the house price. In terms of the physical environment (figure 5.6c), it has a
greater positive influence in the centre London while has a relatively negative influence in the
peripheral area. Distances to both schools (figure 5.6d) and undergrounds (figure 5.6e) appear to have
a greater influence on the house price in the peripheral area, while a less influence in the central area.
On the contrary, the coefficient map of mean household income (figure 5.6f) tells a different story.
An overall positive impact on house price has been visualized by the map, with larger positive values
in the central area of London and along the Times river.
DISCUSSIONS AND CONCLUSION
6.1 Spatial autocorrelation
The result of global spatial autocorrelation reveals that geospatial clustering exists in all the
variables. The highest level of autocorrelation was observed for distance to underground (I =

27
0.987) and physical environment (I = 0.943), followed by income level and accessibility to retail
outlets.
The high correlation in underground distance can be explained by the high density of undergrounds
in central London, while low density in the peripheral area driven by the requirement of residences.
Surprisingly, relatively weak autocorrelation has been found in distance to school. The reason
might be explained from the distribution map, where schools are relatively evenly distributed
across London.
Hotspots in central London have been identified through local spatial autocorrelation methods,
while no coldspots have been observed. It is aligned with the traditional economic theory of bid-
rent (Alonso, 1964) which refers to the relationship between house price and real estate demand
and the distance from the central business district.
6.2 Multiple linear regression
Visualized by the scatter plot, strong associations exist between independent variables and
dependent variables. As expected, household income is positively correlated with house price and
distances to the underground and school are negatively correlated with house prices. The
accessibly to the retail outlet and health service has opposite impacts on house prices. It is
understandable that the higher accessibility to the retail outlet, the higher the house price is.
However, a negative association has been found between health service accessibility and house
price, which could be explained by the cost of health service infrastructure. In central London,
fewer but larger health service institutes would be found, while more but smaller size institutes
might be found in peripheral areas. In addition, a R-squared value of 0.7116 indicates that the
linear model alone can explain most of the variability.
6.3 Spatial Regression Models
Global regression models have been employed to explain the overall spatial autocorrelation across
the area. Among three models, spatial error model performs best in the study by comparing AIC
and log-likelihood. However, spatial autocorrelation remains between fitted values and observed
values.
To further explain the spatial heterogeneity across the study area, a local spatial regression model,
GWR model, has been applied. A Quasi-global R2 of 0.88, compared with the aspatial multiple
linear regression (R2=0.71), suggests that the GWR model performs well by taking spatial
autocorrelation into account. The chosen explanatory variables can therefore explain most of the
variability in house prices. However, from the coefficient map, it could be concluded that although
most variability being explained, the presence of heterogeneity in variables affect its predictivity
level of dependent variable and should be taken into consideration.
6.4 Limitations and Future Works
The pros and cons in the ‘local’ and the ‘global’ in spatial regression must be understood in the
context of searching for broad generalisations versus local distinctions.

28
On the other hand, a major concern that must be raised when applying techniques to spatial data
analysis refers to the issue of the geographical scale applied in the study. Spatially aggregated data
as LSOA could draw major issues that should be noted.
The modifiable areal unit problem (MAUP) refers to questions whether the correlation coefficient
is an appropriate statistic for use with aggregate spatial data. That is, some relationships can be
relatively stable to data aggregation while others can be rather sensitive. Fotheringham et al (2000)
outline two components of the MAUP problem:
1. The scale effect: different results can be obtained from the same analysis at different levels
of spatial resolution.
2. The zoning effect: different results can be obtained through the regrouping of zones.
These issues suggest the importance of applying different aggregation to demonstrate the
sensitivity of the results for different scales. This would provide a way to increase the robustness
of the results and improve the confidence of the conclusions.
Finally, other statistical techniques other than regression could provide good opportunities for
future works as for instance principal components analysis and discriminant analysis. The
awareness that aspatial techniques could be transplanted to spatial problem would makes room for
new spatial analysis possibilities.

29
REFERENCES
Alonso, W. (1964). Location and land use: Toward a general theory of land rents. Harvard University Press,
Cambridge, MA.
Anselin, L. (2010). Spatial econometrics. Dordrecht: Kluwer Academic.
Belsley, D., Kuh, E., & Welsch, R. (2004). Regression diagnostics. Hoboken, N.J.: Wiley.
Brown, S., 1993. Retail location theory: evolution and evaluation. The International Review of Retail,
Distribution and Consumer Research, 3(2), pp.185–229.
Butler, T., & Hamnett, C. (2012). Social geographic interpretations of housing spaces. Sage handbook of
housing studies, 147-162.
Centre for Co-operation with European Economies in Transition (1993). Glossary of industrial organisation
economics and competition law. Organization for Economic.
Cervero, R., Murphy, S., Ferrell, Christopher, et al., 2004. Transit-oriented Development in the United States:
Experiences, Challenges, and Prospects (TCRP 102). Transportation Research Board, Washington, DC.
Clark, W. A.V. and Davies Withers, S. (2007). Family migration and mobility sequences in the United States:
Spatial mobility in the context of the life course. Demographic Research (Max Planck Institute for
Demographic Research)17(20): 591–622.
Clark, W. A. (2012). Residential mobility and the housing market. The Sage handbook of housing studies, 66-
83.
Darmofal, D. (2015). Spatial analysis for the social sciences (pp. 68-95). Cambridge: Cambridge University
Press.
Debrezion, G., Pels, E., Rietveld, P., 2007. The impact of railway stations on residential and commercial
property value: a meta-analysis. J. Real Estate Financ. Econ. 35, 161–180.
Fotheringham, A. S., Brunsdon, C., & Charlton, M. (2000). Quantitative geography: perspectives on spatial
data analysis. Sage.
Haworth, James (2018). Spatial Analysis and GeoComputation: A tutorial guide. Unpublished manuscript.
James, G., Witten, D., Hastie, T., & Tibshirani, R. (2013). An Introduction to Statistical Learning. New York,
NY: Springer.
Krugman, P., & Wells, R. (2010). Microeconomics. New York: Worth Publishers.
Maclennan, D. (2012). Understanding housing markets: Real progress or stalled agendas. The SAGE
Handbook of Housing Studies, 5-26.
Malpezzi S (2003)‘Hedonic Pricing Models: A selective and applied review’. In T. In TO'Sullivan , S.and
K.Gibb. Housing Economics and Public Policy: Essays in Honor of Duncan Maclennan. Blackwell Science.
Oxford, pp. 67–85.
Tobler, W. (1970). A Computer Movie Simulating Urban Growth in the Detroit Region. Economic
Geography, 46, 234. doi: 10.2307/143141
Tukey, J. (1977). Exploratory data analysis. Reading, Mass.: Addison-Wesley Pub. Co.
Ward, M., & Gleditsch, K. (2008). Spatial regression models. London: SAGE.
Zhang, L., & Yi, Y. (2018). What contributes to the rising house prices in Beijing? A decomposition
approach. Journal of Housing Economics, 41, 72-84.

Spatial Analysis of House Price Determinants

More Related Content

Similar to Spatial Analysis of House Price Determinants (20)

Recently uploaded (20)

Spatial Analysis of House Price Determinants