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

net/publication/287232776

A review of hedonic pricing models in housing research

Article · February 2013

CITATION READS
1 3,315

1 author:

Anthony Owusu-Ansah
Ghana Institute of Management and Public Administration
23 PUBLICATIONS   68 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Housing market dynamics View project

All content following this page was uploaded by Anthony Owusu-Ansah on 05 September 2018.

The user has requested enhancement of the downloaded file.

In: Journal of International Real Estate and Construction Studies ISSN: 2153-6813
Volume 1, Number 1 © 2011 Nova Science Publishers, Inc.

A REVIEW OF HEDONIC PRICING MODELS

IN HOUSING RESEARCH

Anthony Owusu-Ansah
Department of Property, Business School, Aberdeen University
Edward Wright Building Annexe, Dunbar Street, Old Aberdeen
Aberdeen, AB24 3QY, United Kingdom

ABSTRACT
The aim of this study is to discuss hedonic pricing models that are often employed in
housing research. Parametric models (loglinear OLS, Box-Cox OLS, and weighted least
square models), nonparametric and kernel estimates (nearest neighbour and local
polynomial estimates models) and semi-parametric models (Robinson-Stock, Yatchew‟s
differencing and Clapp‟s local regression models) are discussed. The discussion shows
that parametric models are easy to estimate and interpret but impose some strong
assumptions, which can produce misleading results when the assumptions are violated.
Even though the nonparametric models avoid the strong assumptions of parametric
models, their computations are technical and complicated. The semi-parametric models
combine elements of the parametric and nonparametric models, thereby, reducing the
problems inherent in such models. Even though the simple parametric approach seems to
be the most widely employed model, the study provides useful insights about other
possible alternative models that can be applied in housing research. However, it is
important to note that based on the advantages and disadvantages of each approach, no
particular approach is necessarily superior to others. The overarching factors that
ultimately determine the choice of an approach are research aim and objectives.

Keywords: Hedonic pricing models, housing research, parametric models, nonparametric

models, semi-parametric models.

1. INTRODUCTION
Real estate contributes immensely to the socio-economic development of every country.
This contribution has been noted by Economic historians like North and Thomas (1973) and
Rosenberg and Birdzell (1986). According to Liu et al. (2002) in 2001, the real estate sector


Email: [email protected]
20 Anthony Owusu-Ansah

as a whole accounted for almost 30% of the gross domestic product (GDP) growth of China.
Wilhelmsson (2009) also observes that the total real estate value in Sweden in 2006 was
almost three times higher than the Gross Domestic Product (GDP) of the country. Thus, real
estate forms a major component of the total wealth of many countries.
Housing is a critical sub-sector and forms a major aspect of the real estate sector in the
world. Liu et al. (2002), for instance, notes that out of the various aspects of real estate
development in China, commercial housing investment is the largest and the most important.
It forms a major part of individual and household portfolios as well as growth rate of all
economies. Various studies including Green (1997), Coulson and Kim (2000), Chau and Zou
(2000), Wen (2001) Liu et al. (2002) and Leung (2003) have established the effects of
housing investment on a country‟s economic growth. These studies note that housing
investment may stimulate GDP growth more than other types of investment. According to
Case et al. (2004) the value of a house has a major impact on households‟ consumption and
savings opportunities. If households‟ consumption and savings opportunities are affected,
then the aggregate consumption and savings are also affected (Tsatsaronis and Zhu, 2004).
Hale (2008) opines that housing has helped in shaping the business cycles of countries like
Britain, Canada, New Zealand and the USA. Housing information is, thus, useful to, for
example, real estate developers, banks and policy makers (Schulz and Werwatz, 2004).
The hedonic regression model is one of the tools that have been used extensively in
analysing the housing market. It has been used by various researchers like: (a) Wilhelmsson
(2000), Janssen et al. (2001), Berry et al. (2003), Pagourtzi et al. (2003), Ong et al. (2003) and
Kim and Park (2005) to investigate house prices and housing characteristics relationship; (b)
Goodman (1978), Meese and Wallace (1997), Dunse and Jones (1998) and Wilhelmsson
(2009) to construct house price indices; and (c) Wyatt (1995 and 1997), Thorsnes and
McMillen (1998) and McCluskey and Borst (2007) to carry out automated valuation of
houses.
Despite the wide application of the hedonic pricing model, housing economists are still
not in agreement on the particular functional form to use. This is partly due to the fact that
economic theory provides little guidance and puts few restrictions on the functional form to
use (Cropper et al., 1988; Anglin and Gençay, 1996) and partly due to the flexibility, easiness
and accuracy of a particular functional form to the researcher as well as the suitability of the
functional form to the nature of the research. As a result, researchers have employed many
types of functional forms in hedonic price analysis. The aim of this study is to examine the
various hedonic pricing models that are often employed to analyse the housing market. To
achieve the stated aim, the objectives to be pursued are to identify and discuss the models that
are often employed, the advantages of employing a particular functional form over the other
as well as the drawbacks inherent in them and the various situations in which the functional
forms have been employed.
The rest of the paper is organised as follows. Section two defines and discusses the
meaning of hedonic modelling whilst sections three, four and five examine parametric,
nonparametric and semi-parametric hedonic pricing approaches respectively. Section six
concludes the paper.
 A Review of Hedonic Pricing Models in Housing Research 21

2. HEDONIC MODELLING
Rosen‟s (1974) study develops the theoretical framework for a hedonic pricing model and
provides the foundation for non linear hedonic pricing models (Sirmans et al., 2005) even
though studies had already applied the model in other areas such as in the automobile industry
(Court, 1939) and in the production of fertilizers (Griliches, 1958). Lancaster (1966) also in
his consumer theory provides a micro economic foundation for estimating the value of utility
generating attributes. Goodman (1998) gives a detailed history of the invention of the hedonic
price analysis.
Rosen (1974, p.34) defines hedonic prices as the implicit prices of attributes that are
revealed to economic agents from observed prices of differentiated products and the specific
amounts of characteristics associated with them. This definition suggests that housing as a
heterogeneous good could be viewed as a package of inherent characteristics relevant to
location, housing structure and neighbourhood or environmental amenities. The heterogeneity
of houses means that no two houses are the same. They may be different with respect to, for
example, location, size, neighbourhood amenities and quality. Each house has its own unique
attributes, which make them different (Sirmans, 2005). Since most of these housing
characteristics are not traded explicitly, their prices cannot be observed directly on the
market. The hedonic pricing model is, therefore, applied to estimate the marginal contribution
of each property and neighbourhood characteristic to the house price.
The hedonic pricing model has been employed enormously in the housing sector. Studies
that have employed the hedonic pricing model in the housing sector include Can (1992),
Ganesan et al. (1997), Sheppard (1999), Wilhelmsson (2000), Janssen et al. (2001), Berry et
al. (2003), Ong et al. (2003), Clapham et al. (2006) and Wilhelmsson (2009). These studies
have been used in the housing sector for making inferences about non-observable values of
different attributes like neighbourhood amenities such as access to hospitals and schools, air
quality and airport noise level to estimate households demand for the various housing
characteristics and to construct housing price indices.
A hedonic regression curve shows the relationship between a dependent variable (house
price in this case) and an independent or explanatory variable (a housing characteristic like
the number of bedrooms). For example if y represents the dependent variable to be estimated,
and x represents the independent variable, the regression equation can be written as:

yi = f(xi) + εi (1)

The regression function f(.) is the average value of y given x. The error term εi represents
the observation errors, which is assumed to be normally distributed with a mean of 0 and a
constant variance σ2 for the purposes of hypothesis testing. As mentioned already, economic
theory provides little guidance and put few restrictions on the functional form f(.) of the
dependence of y on x (Cropper et al. 1988; Anglin and Gençay 1996). The functional form to
be employed for the hedonic pricing model is, therefore, often a struggled issue among
researchers. This has led researchers to employ many functional forms in estimating the
regression curve. The two main different ways in which the hedonic pricing can be modelled
are the parametric and the nonparametric approaches. There is also a hybrid approach – the
semi-parametric approach which combines the elements of the parametric and the
22 Anthony Owusu-Ansah

nonparametric approaches. The various approaches are discussed in detail in the next three
sections.

3. PARAMETRIC HEDONIC PRICING MODELS

The parametric hedonic pricing model is a type of hedonic model in which the regression
curve is assumed to have some pre specified functional form that is fully described by a finite
set of parameters, which are usually the coefficients of the independent variables (Härdle,
1990; Horowitz and Lee, 2002; Fox, 2004). For example, a multiple parametric hedonic
function can take the form of the following:

y = β0 + β1x1 + β2x2 + … + βvxv + u . (2)

where y is the dependent variable and x1 through to xv are the independent variables that
explain the value of x or on which y depends. The betas (β0 through to βv) are the unknown
parameters to be estimated from the equation and u is the error or the disturbance term, which
is assumed to be normally distributed with a mean of zero and a constant variance σ2. The
most commonly known parametric models are discussed below.

3.1. Log-linear Ordinary Least Squares (OLS) Model

The log linear functional form has been greatly employed in hedonic price analysis,
especially, in constructing price indices and analysing housing price determinants (Sirmans et
al., 2005). With this, the natural logarithm of the transaction price of the property is taken as
the dependent variable and the housing and locational characteristics are taken as the
independent variables. That is, instead of taking the absolute values of the transaction prices,
a natural logarithm of them is taken. A log-linear model can take the form:

lnYi = β0 + β1Xi + β2Zi+ui. (3)

where lnY represents the natural log of the transaction price of the property, X represents a
vector of housing characteristics, Z also represents a vector the two locational variables
(latitude and longitude), u is the error term which is assumed to normally distributed with a
mean of 0, and a constant variance σ2.
A lot of studies have used and emphasised the advantages of the semi-logarithmic form
over the other linear functional forms. Follain and Malpezzi (1980), particularly, find that the
semi-logarithmic form: (a) makes the interpretation of the regression coefficients easy – as
the percentage change in the price given a unit change in the housing attribute; (b) allows for
variations in the currency value of each housing characteristic; and (c) helps to minimise the
problem of heteroskedasticity. Wilhelmsson (2000 and 2009) Wilhelmsson (2009) applies
this method in constructing a price index in Sweden and uses it to assess the impact of traffic
noise level on the value of single family houses in Sweden respectively.
 A Review of Hedonic Pricing Models in Housing Research 23

3.2. Box-Cox OLS Model

Box and Cox (1964) have developed a statistical test for the functional form providing
the "best fit" based on likelihood ratio tests and most studies now employ this in choosing the
functional form. According to the formal Box-Cox model:

V (α) = b0 + ∑ bi Zi (β) + i∑ ∑ j ≠ i Ci j Zi (β) Zj (β) (4)

where:
V(α) = (Vα – 1)/α α≠0
V(α) = ln(V) α=0
Zi (β) = (Ziβ – 1)/β β≠0
Zi (β) = ln(Zi) β=0

The Box-Cox methodology has been used in studies by Goodman (1978), Linneman
(1980), Blomquist and Worley (1981), Halvorsen and Pollakowski (1981), Eberts and
Gronberg (1982) and Casssel and Mendelsohn (1985). In comparing the linear functional
forms with the Box-Cox transformation, Goodman (1978) rejects the linear functional forms.
He notes that the linear functional form is “overly restrictive” and so the Box-Cox
transformation provides a better estimate than the linear functional forms. Halvorsen and
Pollakowski (1981) also recommend the use of a Box-Cox transformation. Cassel and
Mendelsohn (1985) observe that even though the Box-Cox functional form is attractive when
the goodness of fit criterion is used, it should be noted that the fact that the best fit criterion is
used to choose a functional form does not necessarily mean that the estimates of the housing
characteristics are more accurate. Particularly, they opine that the accuracy of the Box-Cox
estimates is reduced when the number of coefficients to be estimated is extremely large. They
also argue that the coefficients of a non-linear transformation are “cumbersome” to use
properly and also the Box-Cox transformation may not be suitable when the dataset contains
negative numbers. This is because “any negative number raised to a non-integer real power is
imaginary”. Finally, Cassel and Mendelsohn (1985) have argued that the Box-Cox
methodology may be inappropriate for prediction.
The OLS models explained above usually call for five set of assumptions (called the
Gauss-Markov assumptions) to make the estimates valid. Among them are the zero
conditional mean assumption, the homoskedasticity assumption and the linearity in
parameters assumption (Wooldridge, 2009). If the five set of assumptions hold, then the OLS
is seen as the Best Linear Unbiased Estimator (BLUE).
One major criticism of the OLS models that has also received particular attention in the
literature is the problem of heteroskedasticity. The homoskedasticity assumption underlying
OLS models states that the variance of the unobservable error u, conditional on the
independent variables is constant. That is, Var(u/ x1, x2…, xk) = σ2. This is one of the
assumptions that make the OLS BLUE. However, it has been found in a lot of studies that the
homoskedasticity assumption mostly fails in multiple regression analysis (White, 1980;
Fleming and Nellis, 1984; Robinson, 1987; Fletcher et al., 2000; Stevenson, 2004). That is,
the variance of the error term changes across different segments of the population – a
situation called heteroskedasticity.
24 Anthony Owusu-Ansah

It should be noted that the homoskedasticity assumption is needed to justify the usual t
and F tests, as well as the confidence intervals for OLS estimation of the multiple linear
regression model. Therefore, in the presence of heteroskedasticity inferences cannot be made
since the standard errors and for that matter the F and t statistics are invalid. The estimates,
however, are not biased or inconsistent. Fletcher et al. (2000) also confirm this by observing
that when heteroskedasticity is present in economic models, OLS estimators are still unbiased
but not the best. According to Robinson (1987) the presence of heteroskedasticity in
economic models leads to “inefficient point estimators and hypothesis tests with suboptimal
asymptotic local power”. Heteroskedasticity can lead to both changes in the estimated
standard errors of the coefficients and changes to the value of the coefficients themselves
(Stevenson, 2004).

3.3. Weighted Least Square (WLS) Models

The WLS models are another set of parametric models that researchers mostly rely on.
Like the OLS models, the WLS models are also specified in advanced. With the OLS models,
it is assumed that the error variance is identical for all the variables in the population and so
the OLS gives the same weight to each observation. Instead of given equal weights to all the
observations as in the case of the OLS, the WLS gives different weights to different
observations. Less weight is, therefore, given to observations with higher error variance
(Park, 1966; Harvey, 1976; Lardaro, 1992; Goodman and Thibodeau, 1995 and 1997;
Fletcher et al., 2000).
Where the heteroskedasticity in the OLS models are known up to a multiplicative
constant, then the Generalised Least Squares (GLS) proposed by Davidian and Carroll (1987),
which is a type of WLS have the potential of giving better estimates than the OLS
(Stevenson, 2004). With this model, the initial OLS model is re-estimated with WLS using
the reciprocals of the normalised predicted values from the absolute residual model.
Fletcher et al. (2000) find that in the presence of heteroskedasticity of unknown form, the
Estimated Generalised Least Square (EGLS) estimators are consistent and asymptotically
efficient than OLS estimators. The model uses the residuals from the original OLS model.
Fletcher et al. (2000) stress that EGLS gives a forecast closer to the actual ex post forecasting.
This means that the confidence intervals for the forecasts are narrower. However, they
indicate that the EGLS estimators are efficient in large samples. Thus, in case of small
samples, the procedure may not be the best. According to Stevenson (2004) the corrected
models have the desired effect of reducing the standard errors of the hedonic coefficients.
Unlike the iterative GLS that attribute the cause of heteroskedasticity to a particular variable
like age of a property and so has the ability to remove all heteroskedasticity from the hedonic
model, the EGLS does not assume that heteroskedasticity is caused by a single specified
variable and hence does not eliminate all evidence of heteroskedasticity.

3.4. Merits and Demerits of Parametric Models

Parametric models are the most widely employed models in hedonic regression analyses.
This can be attributed to the fact that the models are fully determined up to a parameter as the
 A Review of Hedonic Pricing Models in Housing Research 25

name suggests. If the assumptions made for these models are correct, then not only will the
estimates be accurate but the fitted models can also be easily interpreted. The simplicity in
estimating and interpreting the coefficients is the major advantage of the parametric models
and no wonder it still remains one of the popular approaches in hedonic modelling.
The approach has, however, come under serious criticisms in the literature. Härdle (1990)
indicates that the preselected parametric model might be too restricted or too low-dimensional
to fit unexpected features. This is confirmed by Anglin and Gençay (1996) who argue that
parametric model involves implicit restrictions. The parametric approach has also been
strongly criticised for imposing strong assumptions such as the linearity of the parameters of
the dependent variable (y) to the independent variables (x‟s) as discussed above. This can
make the model inconsistent and provide a regression relationship that is misleading if any of
the underlying principles or assumptions is violated. For example, when the homoskedasticity
assumption is violated, then the heteroskedasticity is present in the model.

4. NONPARAMETRIC HEDONIC PRICING MODELS

The previous section has discussed parametric models where the regression models are
specified in advanced. The nonparametric approach is flexible and does not pre-specify the
model in advance and also relaxes the strong assumptions of the parametric approach (Fix and
Hodges, 1951; Watson, 1964; Cover, 1968; Stone, 1977; Li, 1984; Robinson, 1988; Meese
and Wallace, 1991; Hastie and Loader, 1993). It, however, assumes a smooth population
regression function (Fox, 2004 and Härdle, 1990) and, therefore, provides an adaptable
method of exploring a general relationship between the dependent and the independent
variables. The nonparametric approach takes the exact form of equation (1) above:

yi = f(xi) + εi

Unlike the parametric regression approach, which specifies the regression (f) in advance,
the nonparametric traces the dependence of the explained variable (y) on the explanatory
variables (x’s) without pre specifying the regression function (f) that relates the explained to
the explanatory variables (Fox, 2004). The mean curve is, thus, estimated without a reference
to a specific form and this offers a flexible tool in analysing unknown regression relationships
(Härdle, 1990).
Like the parametric approach, a weighted sum of the y observations is usually used to
obtain the fitted values for the nonparametric regression approach. However, instead of using
equal weights as in the case of the OLS or weights proportional to the inverse of variance as
in the case of the WLS, with the nonparametric approach, the independent variables closest to
the focal independent variable (x0) are given more weight than those more remote from it.
There is, therefore, a decreasing function of the distances of the location of the independent
variables from the focal independent variable (x0). There are many types of nonparametric
regression techniques and these include the spline method (Reinsch, 1967), the nearest
neighbours method (Fix and Hodges, 1951; Cover and Hart, 1967; Stone, 1977; Li, 1984), the
kernel estimates (Watson, 1964; Nadaraya, 1964) and the locally weighted regression (LWR)
26 Anthony Owusu-Ansah

or local polynomial regression (LPR) approaches (Cleveland and Devlin, 1988; Cleveland et
al., 1988). The next subsection describes some of these techniques.

4.1. Kernel Regression Method

The Kernel estimation method is one of the famous nonparametric methods used in
housing research (Pace, 1993; Min, 2007; Parmeter et al., 2007). The method aims at
obtaining and using appropriate weights to yield the fitted values. The general idea is that in
estimating f(x0), it is appropriate to give greater weight to observations that are close to the
focal value x0. It is argued that the kernel smoothing technique is one of the simplest ways of
computing a weight sequence. According to Nadaraya (1964), Watson (1964) and Nadaraya
(1965) the kernel weight sequence {Wni(x) can be represented by describing the shape of
the weight function Wni(x) by a density function with a scale parameter that adjusts the size
and the form of the weights near x. This is defined by:

Wni(x) = (5)

where Khn(u) is a decreasing function of |u| and 0<hn<1 is called the bandwidth. By studying
the above equation carefully, it can be noticed that the weighting scenario is different for the
observations with observations at locations close to x having more weight than observations
remote from it. The kernel estimate, mh(x) is defined as a weighted average of the response
variables in a fixed neighbourhood around x, determined in shape by the kernel, K and the
bandwidth h. It is important to mention that the kernel is a continuous, bounded and
symmetric real function K that integrates or sums up to one. Since Nadaraya (1964) and
Watson (1964) propose the kernel weights of the form Whi(x), the result from the kernel
estimate in the equation below is often called Nadaraya-Watson estimator.

mh(x) = . (6)

The bandwidth h determines the rate at which the weights decrease as the distance from
x0 increases. Since the rate at which the weights decreases relative to the locations of the x’s
controls the smoothness of the resulting estimate of f(.), the choice of the bandwidth becomes
very necessary to the efficient performance of the nonparametric fit (Hardle, 1990). In a
scenario where the bandwidth is small (h close to zero), the most closest point to the focal x0,
will have majority of the weight with only the other closest observations to this point
receiving the remainder of the weight. The resulting fit would, therefore, connect the dots
formed by the observed data points and will make it under smoothed with possibly high
variance.
On the contrary, if the bandwidth is very large (equal or close to equal to the entire range
in x values), then instead of concentrating the weights on a few data points, the weight is now
evenly distributed across all the observations. This, thus, makes the fit over smoothed and
 A Review of Hedonic Pricing Models in Housing Research 27

possibly with a high bias because it essentially fits the value of y at each data point. Stock
(1991) also indicates that the choice of kernel function can significantly affect the estimates
more than the effect that the bandwidth choice will have. Where the bandwidth is a standard
deviation of a normal distribution centred at the focal x0, it is called the Gaussian or Normal
kernel and when the bandwidth is a half-width of a window centred at the focal x0, it is called
the tricube kernel (Fox, 2004).
As pointed out already, the kernel and for that matter their weights should be
symmetrical. However, when the x approaches a boundary of the data, whether to the left or
to the right, the kernel weights can no longer be symmetric. For example, if the focal point x0
is located at or very close to the right boundary, only points to the left of x0 are capable of
receiving kernel weights. This is because there are simply no (few if any) points to the right
of x0 to receive any weight. This is considered as the main problem of the kernel method.

4.2. Nearest Neighbour Method

The Nearest Neighbour (NN) method was first used by Fix and Hodges (1951) and since
then has been extensively employed in non parametric density and regression estimation and
discrimination (Cover and Hart, 1967: Cover, 1968; Stone, 1977). Pace (1996) has applied
and compared the nearest neighbour method to the OLS method in the housing sector. The
kernel regression estimation as discussed above can be seen as a way in which the weighted
averages of the response variables in a fixed neighbourhood around x can be computed with
the width of the neighbourhood being governed by the bandwidth h. The k-nearest neighbour
(K-NN) estimator can also be viewed as a weighted average of the response variables in a
neighbourhood x. The important difference is that while the neighbourhood width is fixed
with the kernel regression, the neighbourhood width is not fixed but variable in the case of the
nearest neighbour estimate. That is, the values of Y used in computing the average are those
which belong to the k observed values of x that are nearest to the point x at which one would
like to estimate m(x).
Härdle et al. (2004) note that if the modeller is confronted with sparsely data and want to
estimate m(.) at a point of x, it may be the case that the k nearest neighbours are rather far
away from x and farther from each other as well. Therefore, the neighbourhood around x for
which the average of the corresponding y values is computed will be wide. Like the kernel
estimation where h is the bandwidth, k is the smoothing parameter of the estimator with the
nearest neighbour method. When k is increased, the estimate becomes smoother and vice
versa.

The k-nearest neighbour can be viewed as a kernel estimator with uniform kernel,
K(u)= and variable bandwidth R=R(k). The bandwidth or smoothing parameter
R(k) is the distance between x and its furthest k-nearest neighbour. In this case, the nearest
neighbour estimator can mathematically be written as:

k(x) = (7)
28 Anthony Owusu-Ansah

Robinson (1987) indicates that the use of nonparametric nearest neighbour (NN)
regression estimators of σ2 in place of kernel estimators avoids the technical problems
imposed by the random denominator of the kernel estimators.

4.3. Local Polynomial Regression (LPR) or Locally Weighted Regression

(LWR)

Cleveland and Devlin (1988) and Cleveland et al. (1988) have provided a detailed
description of the local polynomial regression. The local polynomial regression is a technique
for estimating a regression surface in a moving average manner. As discussed earlier, the
Nadaraya-Watson kernel regression corresponds to a local constant least squares fit. The local
polynomial regression however helps “to fit a polynomial in a neighbourhood of x”. The
neighbourhood is realised by including kernel weights into the minimization problem below.

(8)

Härdle et al. (2004) indicate that from the above equation denotes a vector of
coefficients ( and represents the kernel weights. The underlying
principle is that instead of fitting only local constant least square fit, with the LPR, the
smooth function can be well approximated by high degree polynomial in the neighbourhood
of any point x. The LPR varies with x and hence can be described as a really local regression
at the point x (Härdle et al., 2004).
The local polynomial estimator of the regression function m is p,h(x)= 0(x). By
running the local polynomial regression with varying x, the curve p,h(.) can be obtained.
Härdle et al. (2004) show that in equation (1) above, when p=0, reduces to 0, which means
that the local constant estimator is the same as the Nadaraya-Watson estimator. That is:
0,h(x)= h(x)= (9)

As discussed in the kernel regression section above, the Nadaraya-Watson estimate has a
problem in that one-sided neighbourhoods at the boundaries are typically observed. This is
because in local constant modelling, almost the same points are used to estimate the curve
near the boundary. The LPR overcomes this by fitting a higher degree polynomial. Again, the
LPR fit can improve the function estimation in regions with sparse observations (Härdle et al.,
2004), a problem commonly encountered with the k-NN estimate. The locally weighted
regression approach has been used to estimate house price index by researchers like Meese
and Wallace (1991) and has been combined with a parametric model by Clapp (2004) to
estimate local house price indices.

4.4. Merits and Demerits of Non Parametric Regression

The nonparametric regression approach provides an adaptable method of exploring a

general relationship between the dependent and the independent variables. It has the ability to
provide predictions of observations yet to be made without reference to a fixed parametric
 A Review of Hedonic Pricing Models in Housing Research 29

model. The approach is also flexible and works well even with small dataset and even when
the dataset has missing values. Also, the approach provides an objective estimation of the
regression curve as compared to the parametric regression where the underlying assumptions
are often unreliable.
One major criticism of the nonparametric regression analysis is the problem of “curse of
dimensionality”. Since nonparametric regression estimates are based on local averaging, in
higher dimensions, the observations become sparsely distributed even for large sample sizes
and so the estimators perform unsatisfactorily. Again, unlike parametric regression estimate,
there are no parameters to describe in nonparametric regression. It is, thus, very difficult to
make quantitative statements about the actual difference between two or more populations.
Nonparametric regression has also not received the needed attention because of the technical
and cumbersome nature of its estimation. The estimates are very quantitatively intensive and
very technical and cumbersome. There is difficulty in assigning weights and also to choose
the appropriate kernel function and smoothing parameter (bandwidth) in nonparametric
regression. All these make the approach difficult to test complex and sophisticated statistical
models. Finally, Ullah (1988) finds that even though the nonparametric models increase the
robustness of the estimates, it tends to be less precise.

5. SEMI-PARAMETRIC HEDONIC PRICING MODELS

The semi-parametric approach incorporates some parametric information into a non
parametric regression so as to benefit from the advantages of both models and reduce the
problems inherent in the previous approaches. The semi-parametric regression is seen as a
useful alternative to the pure non parametric regression. Consider a semi-parametric model of
the form:

yi = ziβ + f(xi) + ui

where yi represents the dependent variable (log of house price in this case), ziβ represents the
parametric part of the equation, f(xi) represents the nonparametric part and ui is the usual error
term. Robinson (1988) is seen as the originator of this idea. Stock (1989, 1991) subsequently
uses this semi-parametric approach to estimate the effect of removing hazardous waste on
house prices. This Robinson-Stock model is further illustrated and applied by Anglin and
Gençay (1996) to estimate a hedonic price function using data from the UK. The Robinson-
Stock model together with two other varieties proposed by Yatchew (1997 and 1998) and
Clapp (2002, 2003 and 2004) are discussed below:

5.1. Robinson-Stock Model

A semi-parametric model is of the form:

yi = ziβ + f(xi) + ui (i = 1, 2,…, n). (10)

30 Anthony Owusu-Ansah

where yi is the ith observation on a dependent variable, zi and β are 1 × p vectors, xi is a 1 × k

vector and the error term ui has a mean of zero and conditional variance σ2(ui | zi, xi). It is,
therefore, the functional form which is unknown to the researcher and so the ziβ and f(xi) as
indicated above form the parametric and the nonparametric parts of the equation respectively.
According to Robinson (1988), the above model can be rewritten as:

pi – E( pi | xi) = (zi – E(zi | xi)) β + ui. (11)

Robinson (1988) and Stock (1989) show that β can be estimated in a two-step procedure.
Firstly, nonparametric technique is used to estimate the unknown conditional means E(pi | xi)
and E(zi | xi) and secondly, the estimates are substituted in place of the unknown functions in
the equation. Parametric OLS model is then used to estimate β. Robinson (1988) shows that
the resulting estimate of β is asymptotically equivalent to those where the true mean functions
are known and used in the estimation. Various nonparametric techniques explained above can
be applied to estimate the unknown conditional means E( pi | xi) and E(zi | xi). Robinson
(1988), for instance, suggests the Nearest Neighbour (NN) technique. Anglin and Gençay
(1996), however, applies the nonparametric kernel estimator to estimate the conditional
means.
In applying the Robinson-Stock semi-parametric model, Anglin and Gençay (1996) enter
all the dummy variables such as driveway, recreation room, finished basement, gas heating,
central air, garage and neighbourhood dummy variables in the parametric or linear part of the
model and enter the discrete and continuous variables into the nonparametric part or the
unknown f(.). They note that the curvature of the function will not be affected if the dummy
variables are included in the nonparametric part but would only cause scale effects.

5.2. Yatchew’s Differencing Model

Another type of the semi-parametric approach is the differencing method proposed by

Yatchew (1997 and 1998):

yi = zi β + f(xi) + ui (12)

Like the other semi-parametric models, the differencing model also assumes the function
f to be smoothed. In addition to this, the model assumes that the function f is single-valued
with a bounded first derivative (Lokshin, 2006). Yatchew (1997 and1998) indicates that it is
very imperative that the data is rearranged in an ascending order such that x1 ≤ x2…≤ xT,
where T is the number of observations in the sample. After the data is sorted, the above
equation is then first differenced to obtain:

yi(n) – yi(n-1) = {zi(n) – zi(n-1)}β + [f{xi(n) – xi(n-1)}] + ui(n) – ui(n-1) n = 2,…,T... (13)

It should be noted that because the derivative of f is bounded, when the sample size
increases, [f{xi(n) – xi(n-1)}] approaches zero and so the nonparametric effect is removed.
Yatchew observes that so far as z and x are not correlated, the OLS model can be used to
estimate β using the differenced data. In this case, the vector of estimated parameters βdiff has
 A Review of Hedonic Pricing Models in Housing Research 31

the approximate sampling distribution. Efficiency of the β estimate can be improved

substantially by using higher order differencing (Yatchew, 1997). Yatchew (1998) applies the
differencing method to estimate the hedonic price of housing attributes. The parametric
variables he uses include the lot size, the presence of various amenities and the living area of
the house. Since the location effect has no natural parametric specification, he puts it in the
nonparametric part.

5.3. Clapp’s Local Regression Model

One difficulty in hedonic models is how to accurately model house prices over space.
According to Clapp et al. (2002) and Clapp (2003 and 2004), house price vary dramatically
over space and most at times difficult to model. Researchers most often omit a lot of
unmeasured spatial variables from the hedonic equation by attempting to include some
neighbourhood and accessibility variables (Dubin, 199; Gillen et al., 2001). Gillen et al.
(2001) indicate that most of the omitted spatial variables can be modelled with spatial
autocorrelation. Clapp's LRM combines this spatial autocorrelation approach with
nonparametric technique to model the omitted spatial variables (Clapp et al., 2002).
According to Clapp (2004) the model like other data mining methods requires the assumption
that the underlying space-time surface is smooth and also, the model does not assume
information on distances to the CBD or other points of interest such as airport. This makes the
model more efficient since consistent identification of these points may not be available.
The LRM has two parts – an OLS model to hold constant for the housing characteristic
and a nonparametric smoother (LPR), which calculates location value as a function of
longitude and latitude. In using the LRM, the coefficients on housing characteristics and
measureable spatial variables are estimated using a linear function such as the log linear
model. A nonlinear function of latitude and longitude is used to estimate the unmeasured
spatial variables and specifically, the Local Polynomial Regression (LPR). Both the
parametric and nonparametric parts of the mean function are jointly estimated. Consider a
semi-parametric model:

yi = ziβ + f(xi) + ui. (14)

where yi is the log of house price, zi is a vector of housing characteristics and xi is a two
dimensional vector of latitude and longitude. According to Clapp (2003, 2004), in using the
LRM, firstly, the log of sale price is regressed on a vector of housing characteristics zi and a
vector of physical location variables of latitude and longitude xi. This is represented by the
equation below:

yi = ziβ1 + xiβ2 + εi. (15)

Secondly, the residuals, η, are calculated from the β estimates of the above equation. That
is:

η = y – xβ (16)
32 Anthony Owusu-Ansah

The calculated residuals are then used to estimate the nonparametric part:

f(X) = Ŋ = smooth (η|x) (17)

where x has two column vectors of equally spaced latitude and longitude coordinates and Ŋ
can be interpreted as a matrix of location values. The function that produces the smooth is the
Local Polynomial Regression (LPR). The LPR produces a local average by down-weighting
observations that are more distant from the fixed point. Assuming a weighted average of Ŋ in
the local neighbourhood around (x0) needs to be estimated, then:

Ŋ(x0) = (18)

This is the Nadaraya-Watson estimator discussed under the nonparametric regression.

The weighting function, Kh(.) is an inverse function of distance and h is the bandwidth. Lesser
weight is, thus, given to more distant points. Clapp (2003) applies the kernel weights to
estimate this equation. The product kernel gives;

Kh(Xi – x0) = Khxi(Xi1 – x01) Khx2(Xi2 – x02). (19)

where the subscripts 1 and 2 on the X variables indicate latitude and longitude.

The LPR takes the form:

Ŋi = γ0 + (Xi-x0) γ1 + (Xi-x0)2 γ2 + … + (Xi-x0)p γp + ξi. (20)

where x0=(x01, x02) represents a specific point on space and Xi=(Xi1, Xi2) represents the
location variables. The subscripts 1 and 2 on the W and w variables indicate latitude and
longitude, γj (j=1,…,p) are column vectors with number of elements equal to the column of
Wi, γ0 is a scaler quantity and Ŋi represents the unobserved location values to be estimated.
Clapp (2003) observes that w is a rectangular grid of equally spaced latitude and longitude
points that span the data and the level of Ŋ is estimated conditional on each knot. Both the
parametric and nonparametric parts of the semi-parametric equation must be estimated
simultaneously and both must be consistent estimators of the underlying functions.
It is seen from this model that while the log linear OLS model capture measurable effects,
the LPR capture the effects on price of immeasurable spatial variables. There is a common
usage of the LPR and models of the form of the LRM in the field of medicine (Wand and
Jones, 1995; Fan and Gijbels, 1996; Hastie et al., 2001) but Clapp (2003) is one of the few
researchers to apply this to the housing sector.
According to Clapp (2004), the semi-parametric approach of LRM offers greater
flexibility and so allows the identification of space-time asymmetries, which other models are
not able to identify. He finds the model to increase the explanatory power and improves out-
of-sample prediction when compares to hedonic OLS model. The LRM efficiently uses scarce
neighbourhood data to estimate the evolution of house prices at each point on the grid. The
model does not require knowledge of the neighbourhood boundaries. Clapp (2003) indicates
that the model‟s inability to forecast house prices is the main drawback of the LRM.
 A Review of Hedonic Pricing Models in Housing Research 33

5.4. Merits and Demerits of Semi-Parametric Regression

The semi-parametric approach as indicated above has the advantages that both the
parametric and the nonparametric approaches possess. They keep the easily interpretability of
the parametric approach and retain some of the flexibility of the nonparametric approach.
These two features are well combined to produce accurate and robust estimate for the
regression model.
In testing the semi-parametric and parametric models, Anglin and Gençay (1996, p.634)
find that the semi parametric model outperforms their benchmark parametric model even
though the benchmark parametric model “passes most of the specification tests that
researchers use to reassure themselves that the parametric specification is reasonable”. Again,
the problem of the “curse of dimensionality” is reduced drastically if not eliminated since
most of the regressors enter into the parametric part. However, as with the parametric
approach, if the underlying assumptions are violated, the model can produce inconsistent
estimates. Again, the computational problems associated with the nonparametric regression
are still associated with the semi-parametric model.

CONCLUSION
Hedonic pricing models have been extensively used in housing research and the simple
parametric models seem to be the well known and the mostly employed methods. Three broad
approaches which are parametric, nonparametric and semi-parametric approaches have been
discussed and the advantages and problems in using each of the approaches have been
identified. The parametric models considered in this paper include the log-linear OLS, Box-
Cox OLS and Weighted Least Square models. The parametric approach assumes the
regression curve has a pre-specified functional form that is fully described by a finite set of
parameters. It has been shown that the WLS reduces the problem of heteroskedasticity, which
is a major problem in parametric models. The approach has the advantage of accurately
estimating and interpreting the model. Its major demerit is that when the assumptions
underlying the model are violated, the model may give a misleading picture of the regression
relationship.
The nonparametric approach avoids specifying the regression function in advance and so
prevents the strong assumptions underlying the parametric approach. Some of the
nonparametric models discussed include the kernel method, the nearest neighbour estimate
and the local polynomial regression. The approach is, however, criticised for its “curse of
dimensionality” problem that makes the estimates inaccurate if the number of regressors is
large. Again, it is very technical and complex and in computing the estimates and this makes
it uncomfortable for most modellers in using the method.
The semi-parametric approach combines elements of the parametric and the
nonparametric approaches. This approach has the advantage of reducing the problems
associated with the parametric and the nonparametric approaches but at the same time
maintain the advantages they possess. The hybrid approach has the advantage of making the
estimate easy to interpret. It also reduces the “curse of dimensionality” problem because it
reduces the number of regressors which enter into nonparametric part. The technical
34 Anthony Owusu-Ansah

computational problem and some of the underlying assumptions of the parametric approach
are, however, still present though they are reduced.
The final and obvious question to ask from the above discussion is: “does the choice of
an approach depend on the advantages and the disadvantages of the approach?” The answer is
not necessarily “yes”. No particular approach is superior to others. The nature and objectives
of the particular housing analysis to be carried out would to a large extent determine the
approach or model to be employed. This is emphasised by Mertens (2003) and Creswell
(2003 and 2007) who note that the type of methodology to be employed in any research
depends on the research problem, questions or aim and objectives and not on the merits or
otherwise of a particular research approach.

AUTOBIOGRAPHICAL NOTE
Anthony Owusu-Ansah is a Doctoral Researcher and a Tutor at the Department of
Property in Aberdeen University, Scotland. He holds an MSc degree in Real Estate
Management from The Royal Institute of Technology (KTH) in Stockholm, Sweden and a
BSc (Hons) degree in Land Economy (First Class) from Kwame Nkrumah University of
Science and Technology (KNUST), Kumasi in Ghana. He has worked as an Assistant
Lecturer at KTH and KNUST in Sweden and Ghana respectively. Anthony has published in a
number of reputed international journals and conferences.

REFERENCES
Anglin, P. and Gencay, R. (1996) Semiparametric Estimation of a Hedonic Price Function,
Journal of Applied Econometrics, 11(6), pp.633-648.
Berry, J., McGreal, S., Stevenson, S., Young J. and Webb J.R. (2003) Estimation of
apartment submarkets in Dublin, Ireland, Journal of Real Estate Research, 25(2), pp.159-
70.
Blomquist, G. and Worley, L. (1981) Hedonic prices, demands for urban housing amenities
and benefit estimates, Journal of Urban Economics, 9(2), pp. 212-221.
Box, G. E. P. and Cox, D. R. (1964) An Analysis of Transformations, Journal of the Royal
Statistical Society, 26, pp.211-43.
Can, A. (1992) Specification and estimation of hedonic housing price models, Regional
Science and Urban Economics, 22, pp.453-74.
Case, B. J., Clapp, R. and Rodriguez, M. (2004) Modeling spatial and temporal house price
patterns: a comparison of four models, Journal of Real Estate Finance and Economics,
29(2), pp.167-91.
Cassel, E. and Mendelsohn, R. (1985) The choice of functional forms for hedonic price
equations: Comment, Journal of Urban Economics, 18(2), pp.135-142.
Chau, K. W. and Zou, G. (2000) The interaction between economic growth and residential
investment, Working Paper, University of Hong Kong.
Clapp, J. M. (2004) A semiparametric method for estimating local house price indices,
Journal of Real Estate Economics, 32(1), pp.127-160.
 A Review of Hedonic Pricing Models in Housing Research 35

Clapp, J. M. (2003) A semiparametric method for valuing residential location. Application to

automated valuation, Journal of Real Estate Finance and Economics, 27(3), pp.303-320.
Clapp, J.M., Kim H-J. and Gelfand A. E. (2002) Spatial 9rediction of house prices using LPR
and Bayesian smoothing, Journal of Real Estate Economics, 30(4), pp.505-532.
Clapham, E., Englund, P., Quigley, J. and Redfearn, C. (2006) Revisiting the past and settling
the score: index revision for house price derivatives, Real Estate Economics, 34(2),
pp.275-302.
Cleveland, W. S., Devlin, S. J., and Grosse, E. (1988) Regression by local fitting, Journal of
Econometrics, 37, pp. 87-14.
Cleveland, W. S. and Devlin, S. J. (1988) Locally Weighted Regression: An Approach to
Regression Analysis by Local Fitting, Journal of the American Statistical Association,
83, pp.596-610.
Coulson, N. E. and Kim, M. S. (2000) Residential investment, non-residential investment and
GDP, Real Estate Economics, 28(2), pp.233-247.
Court, A. (1939) Hedonic price indexes with automotive examples, in: American Statistical
Association (Ed.) The dynamics of automobile demand, New York: General Motors
Corporation, pp.99-117.
Cover, T. M. (1968) Estimation by the nearest neighbour rule, IEEE Transactions on
Information Theory, IT-14, pp.50-55.
Cover, T. M. and Hart, P. E. (1967) Nearest Neighbour Pattern Classification. IEEE
Transactions on Information Theory, IT-13(1), pp.21-27.
Creswell, J. W. (2003) Research design: Qualitative, quantitative and mixed methods
approaches, Thousand Oaks: SagePublications.
Creswell, J. W. (2007) Qualitative Inquiry and Research Design: Choosing Among Five
Approaches, Thousand Oaks CA: Sage Publications.
Cropper, M. L., Deck, B.L. and McConnell, M. (1988) On the choice of functional form for
hedonic price functions, The Review of Economics and Statistics, 70(40), pp.668-75.
Davidian, M. and Carroll, R. (1987) Variance function estimation, Journal of American
Statistical Association, 82, pp.1079-1091.
Dubin, R. A. (1992) Spatial autocorrelation and neighborhood quality, Regional Science and
Urban Economics, 22, pp.433-452.
Dunse, N. A. and Jones, C. A. (1998) A hedonic price model of office rents, Journal of
Property Valuation and Investment, 16, pp.297-313.
Eberts, R. and Gronberg, J. T. (1982) Wage Gradients, Rent Gradients and the Price Elasticity
of Demand for Housing: An Empirical Investigation, Journal of Urban Economics, 12(2),
pp.168-176.
Fan, J. and Gijbels, I. (1996) Local Polynomial Modelling and Its Applications, New York:
Chapman and Hall.
Fix, E. and Hodges, L. H. (1951) Discriminatory analysis-nonparametric discrimination:
consistency properties, Project 21-49-004 Report No. 4, USAF School of Aviation
Medicine, Randolph Field, Texas, pp.261-279.
Fleming, M. C. and Nellis, J. G. (1984) The Halifax House Price Index: Technical Details,
Halifax Building Society, Halifax.
Fletcher, M. Gallimore, P. and Mangan, J. (2000) Heteroscedasticity in hedonic house price
models, Journal of Property Research, 12(2), pp.93-108.
36 Anthony Owusu-Ansah

Follain, J. R. and Malpezzi, S. (1980) Dissecting housing values and rent, The Urban
Institute, Washington DC.
Fox, J. (2004) Nonparametric Regression [online], available from http://socserv.mcmaster.ca
[accessed 12th December, 2009].
Ganesan, S., Tse, R. Y. C. and So, H. M. (1997) Estimating the influence of transport on
house prices: evidence from Hong Kong, Journal of Property Valuation and Investment,
15(1), pp.40-47.
Gillen, K., Thibodeau, T. and Wachter, S. (2001) Anisotropic autocorrelation in house prices,
Journal of real estate Finance and Economics, 23(1), pp.5-30.
Goodman, A. C. (1998) Andrew Court and the invention of hedonic price analysis, Journal of
Urban Economics, 44(1), pp.291-389.
Goodman, A. C. and Thibodeau, T. G. (1997) Dwelling-age-related heteroskedasticity in
hedonic house price equations: an extension, Journal of Housing Research, 8, pp.299–
317.
Goodman, A. C. and Thibodeau, T. G. (1995) Age-related heteroskedasticity in hedonic
house price equations, Journal of Housing Research, 6, pp.25–42.
Goodman, A. C. (1978) Hedonic prices, price indices and housing markets, Journal of Urban
Economics, 5, pp.471–484.
Green, R. K. (1997) Follow the Leader: How changes in residential and non-residential
investment predict changes in GDP, Real Estate Economics, 25(2), pp.253-270.
Griliches, Z. (1958) The Demand for Fertilizer: An Econometric Reinterpretation of a
Technical Change, Journal of Farm Economics, 40, pp.591-606.
Hale, D. (2008) The Global Driver: How Housing is Driving the World Economy, Gale:
Cengage Learning.
Halvorsen, R., Pollakowski, H. O. (1981) Choice of functional form for hedonic price
equations, Journal of Urban Economics, 10, pp.37-49.
Härdle, W. (1990) Applied Nonparametric Regression, Econometric Society, Monograph
No. 19, New York: Cambridge University Press.
Härdle, W., Müller M., Sperlich S. and Werwatz A. (2004) Nonparametric and
Semiparametric Models, Heidelberg, Berlin: Springer-Verlag.
Harvey, A. C. (1976) Estimating regression models with multiplicative heteroskedasticity,
Econometrica, 44, pp.461–465.
Hastie, T. and Loader, C. (1993) Local regression: automatic kernel carpentry, Statistics
Science, 8(2), pp.120-143.
Hastie, T., Tibshirani, R. and Friedman, J. (2001) Elements of Statistical Learning: Data
Mining, Inference and Prediction, New York: Springer.
Horowitz, L. J. and Lee, S. (2002) Semiparametric methods in applied econometrics: do the
models fit the data? Statistical Modelling 2, pp.3–22.
Janssen, C. B. and Soderberg, J. Z. (2001) Robust estimation of hedonic models of price and
income for investment property, Journal of Property Investment and Finance 19(4),
pp.342-360.
Kim, K. and Park, J. (2005) Segmentation of the housing market and its determinants: Seoul
and its neighbouring new towns in Korea, Australian Geographer, 36(2), pp. 221-32.
Lancaster, K. J. (1966) A new approach to consumer theory, Journal of Political Economy,
74, pp.132-57.
Lardaro, L. (1992) Applied Econometrics (2nd Ed.), New York: Macmillan.
 A Review of Hedonic Pricing Models in Housing Research 37

Leung, C. (2003) Economic growth and increasing housing price, Pacific Economic Review,
8(2), pp.183-190.
Li, K-C. (1984) Consistency for cross-validated nearest neighbour estimates in nonparametric
regression, The Annals of Statistics, 12(1), pp.230-240.
Linneman, P. (1980) Some empirical results on the nature of the hedonic price function for
the urban housing market, Journal of Urban Economics, 8, pp.47-68.
Liu, H., Yun, W. P. and Zheng, S. (2002) The interaction between housing investment and
economic growth in China, International Real Estate Review, 5(1), pp.40-60.
Lokshin, M. (2006) Difference-based semiparametric estimation of partial linear regression
models, The Stata Journal 6(3), pp.377–383.
McCluskey, W. J. and Borst, R. A. (2007) Specifying the Effect of Location in Multivariate
Valuation Models for Residential Properties, Property Management, 25(4), pp.312-343.
Meese, R. A. and Wallace, N. E. (1997) The Construction of Residential Housing Price
Indices: A Comparison of Repeat-Sales, Hedonic-Regression and Hybrid Approaches,
The Journal of Real Estate Finance and Economics, 14(1-2), pp.51-73.
Meese, R. and Wallace, N. (1991) Nonparametric Estimation of Dynamic Hedonic Price
Models and the Construction of Residential Housing Price Indices, Real Estate
Economics, 19(3), pp.308-332.
Mertens, D. M. (2003) Mixed methods and the politics of human research: The
transformative-emancipatory perspective, in: Tashakkori, A. and Teddlie, C. (Eds.)
Handbook of mixed methods in the social and behavioural sciences, Thousand Oaks CA:
Sage Publications, pp.100-120.
Min I. (2007) A nonparametric test of the conditional normality of housing demand, Applied
Economics Letters, 14(2), pp.105-109.
Nadaraya, E. A. (1964) On Estimating Regression, Theory of Probability and its
Applications, 9(1), pp.141–142.
Nadaraya, E.A. (1965) On non-parametric estimates of density functions and regression
curves, Theory of Probability Application, 10, pp.186-190.
North, D. C. and Thomas, R. P. (1973) The Rise of the Western World: A New Economic
History, New York: Cambridge University Press.
Ong, S. E., Ho, K. H. D. and Lim, C. H. (2003) A constant-quality price index for resale
public housing flats in Singapore, Urban Studies, 40(13), pp.2705-729.
Pace R. K. (1993) Nonparametric Methods with application to hedonic models, Journal of
Real Estate Finance and Economics, 7, pp.185-204.
Pagourtzi, E., Assimakopoulos, V., Hatzichristos, T. and French, N. (2003) Real Estate
Appraisal: a review of valuation methods, Journal of Property Investment and Finance,
21(4), pp.383-401.
Parmeter C. F., Henderson D. J. and Kumbhakar S. C. (2007) Nonparametric estimation of a
hedonic price function, Journal of applied Econometrics, 22, pp.695-699.
Park, R. E. (1966) Estimation with Heteroscedastic Error Terms, Econometrica, 34, pp.888.
Reinsch, C. (1967) Smoothing by spline functions, Numer. Math. 10, pp.177-83.
Robinson, P. M. (1988a) Root-N-Consistent Semiparametric Regression, Econometrica, 56,
pp.931-954.
Robinson, P. M. (1987) Asymptotically Efficient Estimation in the Presence of
Heteroskedasticity of Unknown Form, Econometrica, 55, pp.875-891.
 38 Anthony Owusu-Ansah

Rosen, S. (1974) Hedonic prices and implicit markets: Product differentiation in pure
competition, Journal of Political Economy, 82(1), pp.34-55.
Rosenberg, N. and Birdzell, L. E. (1986), How the West Grew Rich: The Economic
Transformation of the Industrial World, New York: Basic Books.
Schulz, R. and Werwatz, A. (2004) A state space model for Berlin house prices: estimation
and economic interpretation, Journal of Real Estate Finance and Economics, 28(1),
pp.37-57.
Sheppard, S. (1999) Hedonic analysis of housing markets, in: Cheshire, P. C. and Mills, E. S.
(Eds.), Handbook of Regional and Urban Economics, Amsterdam: Elsevier.
Sirmans, G. S., Macpherson, D. A. and Zietz, E. N. (2005) The composition of hedonic
pricing models, Journal of Real Estate Literature, 13(1), pp.3-43.
Stevenson, S. (2004) New empirical evidence on heteroscedasticity in hedonic housing
models, Journal of Housing Economics, 13(2), pp.136-153.
Stock, J. H. (1991) Confidence intervals for the largest autoregressive root in US economic
time series, Journal of Monetary Economics 28, pp.435-460.
Stock, J. H. (1989) Nonparametric policy analysis, Journal of American Statistical
Association, 84, pp.567-575.
Stone, C. J. (1977) Consistent nonparametric regression (with discussion), Annal Statistics, 5,
pp.595-645.
Thorsnes, P. and McMillen, D. P. (1998) Land Value and Parcel Size: A Semiparametric
Analysis, The Journal of Real Estate Finance and Economics, 17(3), pp.233-44.
Tsatsaronis, K. and Zhu, H. (2004) What drives housing price dynamics: cross-country
evidence? BIS Quarterly Review.
Ullah, A. (1988) Non-parametric estimation of econometric functional, Canadian Journal of
Economics, 21(3), pp.625-658.
Wand, M. P. and Jones, M. C. (1995) Kernel smoothing, New York: Clapman and Hall.
Watson, G. S. (1964) Smooth regression analysis, Sankhya Series A (26), pp.359-372.
Wen, Y. (2001) Residential investment and economic growth, Annals of Economics and
Finance, 2(2), pp.437-444.
White, H. (1980) A Heteroskedasticity-Consistent Covariance Matrix Estimator and a Direct
Test for Heteroskedasticity, Econometrica, 48(4), pp.817-38.
Wilhelmsson, M. (2000) The Impact of Traffic Noise on the Values of Single-family Houses,
Journal of Environmental Planning and Management, 43(6), pp.799–815.
Wilhelmsson, M. (2009) Construction and updating of property price index series: The case
of segmented markets in Stockholm, Property Management, 27(2), pp.119-137.
Wooldridge, J. M. (2009) Introductory Econometrics – A Modern Approach. Mason, OH:
Thomson.
Wyatt, P. J. (1995) Using a Geographical Information System for Property Valuation, Journal
of Property Valuation and Investment, 14(1), pp.67-79.
Wyatt, P. J. (1997) The Development of a GIS-Based Property Information System for Real
Estate Valuation, International Journal of Geographical Information Systems, 11(5),
pp.435-450.
Yatchew A. (1998) Nonparametric regression techniques in economics, Journal of Economic
Literature, 36, pp.669–721.
Yatchew A. (1997) An elementary estimator of the partial linear model, Economic Letters 57,
pp.135–143.

View publication stats