Dept of Real Estate and Construction Management

Centre of Finance and Banking Master of Science Thesis no. 496

The determinants of the retail rent with

OLS and Spatial Regression models
-A case study of Shanghai, China

Author: Supervisor
Liang Jian Mats Wilhelmsson

Stockholm 2010
1 / 38
Master of Science Thesis
Title: The determinants of the retail rent with OLS and
Spatial Regression models-A case study of
Shanghai, China

Author: Liang Jian

Department: Department of real estate and construction
management. Centre of Finance and Banking
Master Thesis number: 496
Supervisor: Mats Wilhelmsson
Keywords: Retail rent, percentage rent, OLS model, spatial
autocorrelation, spatial regression model, RMSE
test.

Abstract:
This paper is an empirical study of retail rent determinants in Shanghai on city level.
Hedonic model and spatial regression model are used in the paper to estimate the
determinants. The significant explaining variables are the age, the area of retail space
in the building, the distance to the Jing An CBD center, the type of the retail property
and the district that the property locates in. A new classification of district in retail
research context is suggested in this paper, and it is proved to be more appropriate in
explaining the retail rent variation among different retail properties than the districts
set up by government.
The problem of spatial autocorrelation in the residuals of OLS models is tested by
Moran’ I statistics, then the spatial lag model and spatial error model are constructed
to estimate the determinants of retail rent in this paper. The root mean square error
(RMSE) test is performed to the OLS model and spatial regression models, and the
spatial lag model is finally selected as the best model in practice.

2 / 38
Acknowledgement

This Master of Science Thesis has been conducted at the Division of

Finance and Banking at the Royal Institute of Technology in Stockholm,

Sweden, during the winter of 2009 and spring of 2010.

I would like to express my gratitude to my thesis supervisor, Mats

Wilhelmsson, professor at the Centre of Finance and Banking, for his

knowledge, time and effort to assist and motivate me to understand the art

of econometrics and to write on a very complicated and interesting topic.

I am very grateful to all the teachers in the Program of Real Estate

Management. Thank them for bringing me into the wonderful world of

real estate, the courses they teach constitute the framework of my

knowledge, and it is the basis for this thesis and my future career.

Finally, I would like to thank my family and friends for their love and

support, especially for my ex-colleagues in CBRE branch in Shanghai,

the database they provided is the necessity for this thesis.

Stockholm, Sweden.2009

Liang Jian

3 / 38
Table of contents
Abstract: ............................................................................................................................... 2

Acknowledgement ............................................................................................................ 3

1. Introduction ...................................................................................................................... 5

2. Literature review – the scope of retail rent research.......................................................... 8

3. Methodology................................................................................................................... 10

3.1 The construction of OLS model .................................................................................. 11

3.2 Spatial autocorrelation test ....................................................................................... 13

3.3 Spatial Regression Model........................................................................................... 15

4. Data description .............................................................................................................. 16

5. Results analysis ............................................................................................................... 21

5.1 Interpretation of OLS model results ........................................................................... 22

5.2 Spatial autocorrelation test and spatial regression model .......................................... 26

5.3 The root mean square error test and model selection................................................ 27

6. Conclusion....................................................................................................................... 28

7. References ...................................................................................................................... 29

8. Appendix ......................................................................................................................... 31

Appendix 1- retail related figures about Shanghai ........................................................... 31

Appendix 2-weight matrix ............................................................................................... 32

Appendix 3-formulas for Z sore ....................................................................................... 33

Appendix 4-variables correlation ..................................................................................... 34

Appendix 5-estimate with OLS models............................................................................. 35

Appendix 6-VIF test results .............................................................................................. 37

Appendix 7-formulas for RMSE test ................................................................................. 37

4 / 38
 The determinants of the retail rent with
OLS and Spatial Regression models
A case study of Shanghai, China

1. Introduction
The research about real estate rent is of great importance. Not only the property
owners and tenants are concerned because the rent level directly influences the profit
they get, but also the investor would take the rent into consideration when they
calculate the return and make the decision to invest or not. Another area of importance
is that the rent is used to calculate the bubble indicator, cap rate for example. These
bubble indicators play an important role in investigating the stability of the finance
market and the economy growth.
One important lesson from the sub-prime crisis is the significance detrimental
power of the real estate bubble. Koetter and Poghosyan (2009) used panel data
collected from Germany to demonstrate the significant negative relation between the
stability of the financial market and the deviation of the house price from the
equilibrium value. In order to avoid the real estate bubbles, the transaction price of
property should be obtained by calculating the sum of present value of the future
profit — net operation income (NOI). The NOI is defined as the total revenue of the
property, and it equals to the rental income minus the total operation cost in most
cases. So the accurate estimation of the rent is the prerequisite of accurate prediction
of future NOI and keeping the stability of real estate market and economy growth, and
researches about rent and its determinants are a requisite to guide the real estate
investor and government to perform appropriate decision and administration.
Reviewing the previous research about rent estimation, most of the works

5 / 38
concentrate on the residential and office markets, the research about rents in retail
properties are limited. The retail rental market is more complicated than the other
types of property market because more factors are affecting the retail rent when
comparing with office and residential market, such as anchor tenant and the mix of
brands.
As the retail market in Asia develop at a rapid pace, more and more retail
tenants and landlords like to sign the retail space rental contract in a risk sharing way
according to the research by Yap (1996). The percentage rent of sales is used to pay as
the rent in most of the retail space rental business in lot of East Asian cities like Tokyo,
Hong Kong ,Singapore and Shanghai［Williams, Tagami（2002），Tay, Lau, Leung
（1999） and Yap (1996)］. The percentage rent is expected to vary among different
properties: if the retail property is attractive for customers in every aspect and attracts
a lot of customers every day, then the land lord can set a higher percentage rent for
tenants because all of the brands want to reside there to make more money, the
demand for space in the retail property is high. Nevertheless, very few literatures
about the determinants about percentage rent in retail are found. This paper will focus
on the percentage rent, and the retail rent means percentage rent in the rest part of
paper.
The knowledge in geography is now more and more widely applied to urban
economics, especially in real estate economic research, and the spatial econometrics
has now become an important subfield of econometrics. Nevertheless, as a classical
application of interaction of geography technology and econometrics theory, spatial
autocorrelation is seldom mentioned in the previous research pertaining to retail rent.
More research should be done to fill in the gap.
As one of the most developed cities in China, Shanghai has been exhibiting the
tremendous potential and opportunities in the retail market. According to the data
from the National Bureau of Statistics of China, the Gross Domestic Product (GDP)
and GDP per capita in Shanghai in year 2007 (1219 billion RMB and 66367 billion
RMB respectively) are forty five times and twenty seven times of that in year 1978.

6 / 38
The strong economy is the critical factor of retail business development according to
the researches by Tsolacos (1995) Hanna, et al. (2007) and Hendershott, et al. (2002).
And the change of economy structure also guarantees the growth of retail industry:
the proportion of service sector in GDP increases from 19% in 1978 to 53% in 2007.
Under this background, development of the retail business in Shanghai has been
keeping strong in the past 20 years. As the Figure.1 below shows, the total retail
turnover of Shanghai in year 2007 (384.8 billion RMB) is seventy one times of that in
year 1978.
Figure .1 The turnover change from 1978 to 2007.

Retail turnover(100 million)

5000
4000
3000
2000
1000
0
1975 1980 1985 1990 1995 2000 2005 2010

The figures are downloaded from Statistics Bureaucracy of China. The table with more

details can be found in appendix.1 (http://www.stats-sh.gov.cn/2008shtj/index.asp)

The 2010 Shanghai World Expo has been bringing in unprecedented

opportunity for the development of the retail business in Shanghai. The fame that
brought in by the 2010 World Expo will attract more domestic and foreign investment
to the retail real estate market in Shanghai. Moreover, as the economy develops, more
top grade brands will choose to exploit the market in Shanghai because of the
improvement of the living standard and affordability of citizens in Shanghai. Within
this background, the researches and guidance about retail market in Shanghai are
required for the retail investors, tenants and related government departments.
This paper aims to estimate the determinants of retail rent in Shanghai. A cross
section data set is used in the paper, and regression models are estimated. The
description of the model and relative statistics test are discussed in chapter 3, and the
estimated result and analysis will be presented in chapter 5. In chapter 4, the data used
7 / 38
in the model will be described. The data set is provided by real estate service
company—CB Richard Ellis and the National Bureau of Statistics of China. In the
end, the article will be concluded in chapter 6.

2. Literature review – the scope of retail rent

research
The relationship between retail sales profit and retail rent was investigated by Chun,
et al. (2001). In their research, the cross section data about shopping centers of USA
in 1995 and time serials data of retail sales profit are used. The results of their
research suggest that the rent does not change correspondingly to the sales profit in
the retail property immediately; the impact from profitability of the retail property to
the rent of the retail property is lagged considerably.
On the city level, Sirmans, et al. (1992) demonstrated the significant impact of a
number of different characteristics and location of the retail property on the rent. In
their research, a panel of shopping centers in Baton Rouge and Louisiana, USA from
1989 to 1991 is used. In order to avoid the problem of heteroscedasticity, the weighted
least square estimation is used instead of ordinary least square estimation. The model
and variables used in their paper are specified as follows:
𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑗𝑗𝑗𝑗 = 𝑓𝑓(𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝑗𝑗 , 𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝑗𝑗 , 𝐿𝐿𝐿𝐿𝐿𝐿𝑗𝑗 , 𝑀𝑀𝑀𝑀𝑀𝑀𝑗𝑗 ) (2.1)

In this model, explanatory variables are separated into four types:(1) the variable
“Draw” means the customer drawing power, it includes “the total area, age of the
property and type of the anchor tenants”.(2) The “Design” means the design of the
retail building, it is measured by dummy variables that describing the configuration of
building .(3) The “LOC” means the location of the property, which is described as
dummy variables.(4) The “MKT” means the market condition ,it includes the external
amenities and the economy condition. All the variables mentioned above are
estimated to be significant in explaining the variation of the rent in retail according

8 / 38
the results of their paper.
Similar research is conducted by Tay，et al.(1996) and Hui, et al. (2006) to
investigate the retail properties rent in Hong Kong market. In the paper of Tay, et
al.(1996), the panel data set consisted of 405 retail tenants in 9 shopping center in
Hong Kong from 1991 to 1994 is used. The influence of variation among different
types of retail units on retail rent is estimated to be significant in their paper. In the
paper of Hui, et al. (2006), more variables are taken into their model such as the
numbers of floors and the vacant rate. It is worth of noticing that the explaining power
of vacancy rate in retail properties is estimated to be insignificant in Hong Kong retail
market .This contradicts the results of the research from Sirmans, et al. (1992) and
Hardin,et al.(2001) which focus on the retail market in US.
To investigate the explaination power of vacancy rate, instrumental variables and
two stage models are employed by Hardened al. (2002). In their paper, a cross
sectional dataset including 118 shopping centers in Atlanta, USA is used. The vacancy
rate is treated as endogenous variable and instrumented in the first stage model, and
then the estimated vacancy rate from the first model is used as along with other
explaining variables, to explain the rent variation among shopping centers. Significant
and negative impact of vacancy rate on rent is detected.
Jackson (2001) challenged the traditional regions classification when the retail
rental level is concerned. In their research, cluster analysis technique and panel data
from UK are used. It is suggested in his paper that the districts that set with
government administration purpose should be replaced by the classification which is
of economics meaning. For example, the classification of district should be based on
the annual rental growth rate. The influence from nearby retail properties on the retail
rental level of retail property is investigated by Hardin and Wolverton (2001) by using
cross section dataset with information about shopping centers in Atlanta, USA. The
results of their research show that the impact of the neighborhood retail property rent
is significant, and the impact is more significant for community centers than for
shopping mall.

9 / 38
 The results indicate the existence of spatial autocorrelation in retail properties
rent. To take care of the spatial autocorrelation, Carter, et al. (2000) give an example
of using weighted least squares (WLS) model to control for the spatial autocorrelation
in retail rent determinants research.
On aggregate level, the significant explanatory variables of retail rent are
estimated as Gross Domestics Production (GDP), real consumer expenditure, the rent
level in past years, the total retail sales volume and retail supply volume. In the
research of Tsolacos (1995), a time series dataset from 1977 to 1994 in UK is used. In
their paper, the autoregressive and moving average model is built to estimate the
determinants and make prediction of retail rent level in the next four quarters. Similar
research is conducted by D’Arcy, et al. (1997) in the research of comparing retail rent
determinants in Europe cities and Hanna, et al. (2007) in the investigation of retail
rental market in USA. Hendershott, et al. (2002) went further to apply error correction
model to estimate the adjustment speed of office and retail rent in UK. The variables
in their models are the same as the other aggregate level research mentioned above.
The influence of government retail planning policies on retail market is investigated
by Jackson and Watkins (2004) by using a time series data from UK. The effect of
government policies of retail market is confirmed in their research. The predictability
of prevailing rent prediction models are tested by Chaplin (2000), emphasizing the
application of “ex ante prediction” before choosing the best model to predict the
future rent in retail.
With this background, we investigate the determinants of rent for retail space in
Shanghai on the rent level in 2009.

3. Methodology
In this chapter, three OLS models which are based on the knowledge of the previous
research are first constructed. Efforts are put into the OLS models to control for the
possible spatial autocorrelation by including spatial related variables such as location
10 / 38
and distance to CBD. Then the test for spatial autocorrelation is introduced, and the
spatial regression model is also constructed in the last part of the chapter. All the
models and test are based on a cross sectional dataset.

3.1 The construction of OLS model

First, Ordinary Least Square (OLS) model is built to estimate the determinants of
retail percentage rent on city level. All the explaining variables that are chosen are
discussed in Sirmans and Guidry (1993) and Sirmans, Gatzlaff and Diskin (1996)
except the vacancy rate, the configuration of the retail building and the externalities
amenities. The main reason that vacancy rate is not chosen in paper is due to lack of
information about vacancy rate. Another reason is that the vacancy rate would have an
endogenous relationship with rent (Hardin, Wolverton and Carr, 2001). The design
of the retail building are highly correlated with the age of the building (newer building
has more attractive configuration) and the externalities amenities is highly correlated
with the location of the building(the districts with higher rent level are closer to the
CBD and have better external amenities). The problem of collinearality would occur
if we include them into the OLS model.
Before the model is estimated, some transformations are performed to the variable in
the model. First, the dependent variable (percentage rent) and the explaining variables
TA (total area) and RA (retail area) are in nature logarithm, and the open year of the
retail property is taken placed by the age of the retail property. The model is
constructed as follow:

Y = α0 +𝛽𝛽1 × L𝑇𝑇A + 𝛽𝛽2 × L𝑅𝑅A + 𝛽𝛽3 × Age + β4 × NoL + ∑ni=1(β1i × DGi ) + ∑nj=1(β2j × Tj) + ∑nk=1 β3k DTCk +ε (3.1)

The dependent variable Y , independent variables LTA and LRA equal to the nature
logarithm of retail rent, total area and retail area respectively, that is
Y=ln(Rp),LTA=ln(TA) and LRA=ln(RA). The vector age equals to 2010 minus open
year. 𝐷𝐷𝐷𝐷𝑖𝑖 is dummy variable for the district that the property is located in, it is worth
of noticing that the district 𝐷𝐷𝐷𝐷𝑖𝑖 is constructed by the government. 𝑇𝑇𝑗𝑗 is dummy

variable for different type of retail and DTCK means the distance from the property to
11 / 38
the center of the city (CBD or sub center).
According to Jackson (2001), the districts constructed by the government are
usually inadequate to describe the location attributes for properties in real estate
research. In order to get a better estimation, new classification of districts should be
proposed to take the place of the districts dummy variables in model (3.2), new model
is constructed as follow:

Y = α0 +𝛽𝛽1 × L𝑇𝑇A + 𝛽𝛽2 × L𝑅𝑅A + 𝛽𝛽3 × Age + 𝛽𝛽4 × 𝑁𝑁𝑁𝑁L + ∑ni=1(βsi × Di) + ∑nj=1(β2j × Tj) + ∑nk=1 β3k DTCk +𝜀𝜀 (3.2)

In the model (3.2) above, the new districts dummy variable Di is constructed
according to the different prosperity level of the retail industry among different
regions in Shanghai, more description about the districts classification is presented in
the chapter 4.
In order to investigate the different impact of retail property characteristics on the rent
value among different districts and different retail Type, the variables TA, RA, Age
and NoL are integrated with the dummy variables Di and 𝑇𝑇𝑗𝑗 to generate new
explaining variables in model (3.3) below.
Y = α0 +𝛽𝛽1 × L𝑇𝑇A + 𝛽𝛽2 × L𝑅𝑅A + 𝛽𝛽3 × Age + 𝛽𝛽4 × 𝑁𝑁𝑁𝑁L + ∑ni=1(βsi × Di ) + ∑nj=1(β2j × Tj ) + ∑nk=1 β3k DTCk +

∑ni=1(βsti × Di × LTA) + ∑nj=1(β2tj × Tj × LTA) + ∑ni=1(βsri × Di × LRA) + ∑nj=1(β2rj × Tj × LRA) + ∑ni=1(βsai ×

Di × Age) + ∑nj=1(β2aj × Tj × Age) + ∑ni=1(βsni × Di × NoL) + ∑nj=1(β2nj × Tj × NoL) + ∑ni=1 ∑nj=1(βstij × Di ×

Tj)+𝜀𝜀 (3.3)

The coefficients of integrated variables can be interpreted as the marginal effect of the
retail property characters on the retail rent for different districts and different type. For
example, the coefficient βsti represents the percentage change of retail rent given one

percentage of total area increase in district i. Similarly, the β2tj means that the

percentage change of retail rent given one percentage of total area increase for retail

property type j. Moreover, the coefficient βstij is estimated to be the percentage

change of retail rent if the retail property belongs to retail type j and district location i.

12 / 38
3.2 Spatial autocorrelation test

The test for spatial autocorrelation should be performed on the error term in the OLS
models. The problem of spatial autocorrelation can be detected by testing the statistics
of Moran’s I［Moran (1948, 1950a, 1950b)］. The problem of spatial autocorrelation in
the economics research especially in real estate economics is described by
Anselin(1999) as : Due to the discrepancy among different regions, for one specific
individual ,the value for the variables with concern correlates with that of the nearby
individuals more than the individuals which are far away from it. So when the OLS
model is regressed using the data collected among all the regions, the error terms of
the OLS model may be system positive in some region while system negative in other
region. Hence the variance of the error term will not be constant among different
regions. In such situation, the OLS model would not be best model capturing the
relation between the dependent variables and independent variables. Moreover, the
hypothesis test performed to the estimators in the OLS model is not valid to because
of the biasness of the error term.
In this case, the spatial autoregression (SAR) model should be constructed to
take the place of the OLS model to estimate. To perform the test and construct the
SAR model, a spatial weight matrix should be constructed to measure the spatial
correlation among different observations. There is no strict theoretical framework in
the construction of the weight matrix, and the weight matrix can be built according to
the need of specific practice application .However, one principle need to be noticed is
that the matrix should be exogenous to the regression model ［Manski (1993)］.
In this paper, the elements of the weight matrix are specified as inverse-distance
matrix (weight matrix in appendix2) according to the spatial distance between the
observations [Anselin (1980, Ch. 8)]. The spatial matrix indicates the spatial
relationship among all the observations, and the value of the elements in the matrix is
equal to:

𝑤𝑤𝑖𝑖𝑖𝑖 = 1�𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑 𝑏𝑏𝑏𝑏𝑏𝑏𝑏𝑏𝑏𝑏𝑏𝑏𝑏𝑏 𝑖𝑖 𝑎𝑎𝑎𝑎𝑎𝑎 𝑗𝑗 (3.4)

13 / 38
The distance between observations i and j is obtained by computing:
1
𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑 𝑏𝑏𝑏𝑏𝑏𝑏𝑏𝑏𝑏𝑏𝑏𝑏𝑏𝑏 𝑖𝑖 𝑎𝑎𝑎𝑎𝑎𝑎 𝑗𝑗 = ((𝑙𝑙𝑙𝑙𝑙𝑙𝑖𝑖 − 𝑙𝑙𝑙𝑙𝑙𝑙𝑗𝑗 )2 + (𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑗𝑗 − 𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑖𝑖 )2 )2 (3.4)

Where 𝑙𝑙𝑙𝑙𝑙𝑙𝑗𝑗 and 𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑗𝑗 represent the latitude and longitude of observation respectively.

So according to the way that the matrix is constructed, the higher value of element
𝑤𝑤𝑖𝑖𝑖𝑖 indicates the shorter distance between i and j, and vice versa.

In order to simplify the interpretation and estimation process, the spatial weight
matrix is standardized by dividing the each cell of the matrix wi,j with the sum of
respective row cells:
∗ 𝑤𝑤𝑖𝑖,𝑗𝑗
𝑤𝑤𝑖𝑖,𝑗𝑗 = �∑105 𝑤𝑤 (3.6)
𝑗𝑗 =1 𝑖𝑖𝑖𝑖

The standard weighted matrix can be generated by the expression above, and the sum
of each row equals to 1.
Now that the standard weight matrix which indicates the spatial structure of the
observation has been built, the statistics of Moran’I can be calculated by the following
expression［Moran (1948, 1950a, 1950b)］:
′
I=�𝑁𝑁�𝑆𝑆 � �𝑒𝑒 𝑊𝑊𝑒𝑒�𝑒𝑒𝑒𝑒 ′ � (3.7)
0

𝑆𝑆0 = ∑𝑖𝑖 ∑𝑗𝑗 𝑊𝑊𝑖𝑖𝑖𝑖 (3.8)

N is the number of the observations, W is the standardized weight matrix, i and j
indicate the number of the element in the weight matrix, 𝑒𝑒 is the residual of the
regression model and 𝑒𝑒 ′ is the transpose of 𝑒𝑒.
The range of the Moran’I statistics is (-1, 1).The more it is close to 1, the more
similar spatial structure is observed, the more it is close -1, and more dissimilar spatial
structure is observed. When the statistics of Moran’I equals to 0, the problem of
spatial autocorrelation is convinced to be not bothering the estimation process. With
the purpose of more precise test, the Z scores can be used and hypothesis test(t-test)
can be utilized [Cliff and Ord (1972, 1981)]. If the residuals of the regression model are
distributed normally, the distribution of Moran’s I would be normal, then the Z scores
can be used and a hypothesis test(t-test) can be performed to estimate whether the
Moran’I is significant different from the value which indicates no spatial
14 / 38
autocorrelation(as expression 8.1 in appendix3 shows). This formula derivation can be
found in appendix.
The null hypothesis of test is Moran’s I=E (I), which means that no spatial
autocorrelation is detected. If the Z value is higher than 1.96, the hypothesis would be
rejected, then spatial autocorrelation is significant at 95% confidence level.

3.3 Spatial Regression Model

Once the spatial autocorrelation level is tested to be significant, spatial regression

model [Cliff (1981) and Anselin (1988)] should be built to estimate as follow:
𝑌𝑌 = 𝛾𝛾𝑊𝑊1 𝑌𝑌 + 𝛽𝛽𝛽𝛽 + 𝑢𝑢 ; 𝑢𝑢 = 𝜌𝜌𝑊𝑊2 𝑢𝑢 + 𝜖𝜖 (3.9)

1
where the Y is the independent variable(Y= ln(1−𝑅𝑅𝑅𝑅 − 1)) expressed as a N× 1 vector,

W1and W2 are 𝑁𝑁 × 𝑁𝑁 weight matrixes, W×Y is a N× 1 spatial lag vector,which

indicates the influence from the neighbor observations, and 𝛾𝛾 is the coefficient of
the spatial lag. 𝑋𝑋 is a N× 𝑀𝑀 matrix with M explanatory variables (including the total
area and the retail section area of the property, the open year, the number of layers for
retail section, dummy variables for districts and dummy variables for different type of
retail) in row and N observations in column, and 𝛽𝛽 is an 1× 𝑀𝑀 vector indicating the
coefficients of the M explanatory variables. u is the residual of the regression model
with spatial correlation and 𝜌𝜌 is the coefficient of the spatial correlation of the
residual u, 𝜖𝜖 is the independent error term without any spatial correlation.
In this paper, spatial autoregression model and spatial error model, which derive
from the general spatial regression model are used. In the expression of (3.12),
when𝜌𝜌 = 0, and 𝛾𝛾 and β ≠ 0 the spatial regression model is generated as follow:
𝑌𝑌 = 𝛾𝛾𝑊𝑊1 𝑌𝑌 + 𝛽𝛽𝛽𝛽 + 𝜖𝜖 (3.10)
In retail rent context it means the rent level in one shopping center is influenced by
the rent level of the shopping center nearby. When 𝛾𝛾 = 0, 𝜌𝜌 and β ≠ 0 the spatial
error model is generated as follow:
𝑌𝑌 = 𝛽𝛽𝛽𝛽 + 𝜌𝜌𝑊𝑊2 𝑢𝑢 + 𝜖𝜖 or ( 𝐼𝐼 𝑛𝑛 − 𝜌𝜌𝜌𝜌 )𝑦𝑦 = ( 𝐼𝐼 𝑛𝑛 − 𝜌𝜌𝜌𝜌 )𝑋𝑋 𝛽𝛽 + 𝜖𝜖 (3.11)
15 / 38
In this model, the influence from both the dependent and independent variables of the
shopping center nearby are estimated.
What should be noticed when regressing this model is that the vector of W1 Y
is not exogenous of the model, when OLS model is applied to this case, the
estimations would be biased and inconstant. To cope with the problem, the techniques
of generalized spatial two-stage least squares (GS2SLS) with instrumental variables
which developed by Kelejian and Prucha (2004) along with coauthors Arraiz et al.
(2008) or Maximum Likelihood Estimation (MLE) proposed by Cliff (1981), should
be used in estimating the parameters. In this paper, the MLE technique is used
because it is widely used in the application of spatial econometrics field, and all our
independent variables as exogenous.

4. Data description
The data used in the regression analysis is provided by real estate service company
CB Richard Ellis branch in Shanghai. It contains information about 109 typical retail
properties in Shanghai on the level in August 2009. The variables in the regression
model include the average rent in each property; the districts (stipulated by
government and set up with economics research purpose) that retail properties locate
in; the total area of each building; the retail area of each building; the type of the retail
property belongs to (department store, shopping center, trade market or outlets); the
open year; the number of layer in the retail sector of the building and the distance to
some specific center. Table 1.gives the descriptive statistics description of the data set:

Table 1. Descriptive statistic

Variable Obs Description Mean Std. Dev. Min Max

Y 89 Y=ln(percentage rent) -1.53544 0.308357 -2.52573 -0.96758

16 / 38
persentage~t 89 Rent in term of percentage 0.224943 0.063173 0.08 0.38

Distance to CBD in Jing An district

dtocbd 103 center 0.054916 0.053271 0 0.397612

Distance to the shopping center with

Dtobund6 103 highest rent 0.059055 0.060191 0 0.436812

District for bunds and East Nanjing

D1 103 Road 0.135922 0.344382 0 1

D2 103 District for Jing An CBD 0.067961 0.25291 0 1

D3 103 Sub CBD district in Lu Wan 0.242718 0.430823 0 1

D4 103 New CBD district in PU Dong 0.106796 0.310364 0 1

Districts that are adjacent with D1

D5 103 and D2 0.300971 0.460923 0 1

Districts that aren’t adjacent with D1

D6 103 and D2 0.145631 0.354461 0 1

DG1 103 Bao Shan district 0.029126 0.168983 0 1

DG2 103 Chang Ning disrict 0.097087 0.297525 0 1

DG3 103 Hong Kou district 0.087379 0.28377 0 1

DG4 103 Huang Pu district 0.184466 0.38976 0 1

DG5 103 Jing An district 0.07767 0.26896 0 1

DG6 103 Lu Wan district 0.106796 0.310364 0 1

DG7 103 Ming Hang district 0.097087 0.297525 0 1

DG8 103 Pu Dong district 0.116505 0.322398 0 1

DG9 103 Pu Tuo district 0.048544 0.215963 0 1

DG10 103 Qing Pu district 0.019418 0.138662 0 1

DG11 103 Song Jiang district 0.029126 0.168983 0 1

DG12 103 Xu Hui district 0.048544 0.215963 0 1

DG13 103 Yang Pu district 0.048544 0.215963 0 1

DG14 103 Zha Bei district 0.009709 0.098533 0 1

age 100 The age of the retail individuals 8.38 4.888205 1 26

Type1 98 Shopping center 0.591837 0.494021 0 1

Type2 98 Department store 0.346939 0.478443 0 1

Type3 98 Trade market 0.040816 0.198882 0 1

Type4 98 Out lets 0.020408 0.142119 0 1

nooflayer 98 The floors number for retail sector 6.081633 2.296 1 13

openyear 100 The year that the retail open 2002.62 4.888205 1985 2010

retailgfas~m 99 The dimension for retail sector 51271.34 83260.66 3800 700000

totalareas~2 98 The total dimension for the building 71504.22 81280.89 4541 430000

lnTotalArea 98 The logarithm of total dimension 10.71538 0.968388 8.420902 12.97154

lnRetailArea 99 The logarithm of retail dimension 10.27051 1.007179 8.242756 13.45884

In the previous research about retail rent estimation, the rent in real term is
taken as independent variable in most of the papers ［ Sirmans, et al. (1992).
17 / 38
etc］.Nevertheless, more and more retail contracts in Shanghai are signed in a more
risk-sharing way according to the results of survey and data base provided by CB
Richard Ellis branch in Shanghai. The principle of the making the percentage rent is
that the amount of the rent is obtained by multiplying the turnover of tenants per
month with a certain percentage specify in the rental contract beforehand. For
example, if the percentage rent for a tenant is 20%, then the tenant have to pay 20
RMB to the landlord as rent for every 100 RMB goods he sells to the customers. All
the other operation expenses are included in the 20 RMB. This percentage rent
mention above is represented in the table as percentage~t.
Some transformations are performed to the dependent variable percentage rent
as mentioned above in the methodology chapter, that is Y=ln(percentage~t) . The
reason of the transformation is that the marginal effect of the explaining variables on
the percentage change of percentage rent can be estimated by doing so.
The dummy variables 𝐷𝐷𝐷𝐷𝑖𝑖 indicate the government specifying district that the
retail property locates in. For example, if the variable 𝐷𝐷𝐷𝐷1 equals to 1, we know that
the property is in Bao Shan district. However, the Shanghai retail market pattern is not
divided as the districts planed by government, so the dummy variables D1 to D6 are
generated to specify the different location of the properties according to the prosperity
level of retail business in different region ［Wang, et al. (2006), Ye,Xiong (2002) and
Jackson (2001)］. The rent level varies among districts because of the diversity of the
business propensity level in different district. The rent level is higher in the region
where the retail business is highly developed. According to the survey investigation
by Wang, et al. (2006) and Ye,Xiong (2002), the retail market in Shanghai presents a
two polar pattern: the turnover per square meter of shop is highest in the regions of
Jing An CBD district and Bund area on the west side of the Huang Pu river, and the
turnover level decreases as the distance to these two polar increase. The retail rent
among the retail properties are expected to follow this kind of pattern, and the dummy
variables 𝐷𝐷𝑖𝑖 indicating the location districts are set up according the this pattern. For
example, if the variable 𝐷𝐷1 equals to 1, it means that the retail property is in the one

18 / 38
the two most “expensive” regions that is the Bund area.

Figure.2 The districts of Shanghai

The picture above describes the districts planed by government (𝐷𝐷𝐷𝐷𝑖𝑖 ) and the
newly constructed districts𝐷𝐷𝑖𝑖 . In this figure, the districts planed by government are
indicated by the characters “MinHang”,“LuWan” ,etc. While the newly constructed
districts D1 to D4 are indicated by the circle 1 to 4, and the district D5 are the shaded
area with numbers 5, and the rest of the region in the picture is district D6.
The rent level also varies among different types of retail: The rent for shopping
centers and department stores are expected to be higher than that of trade markets and
outlets. Shopping centers and department stores are usually located in the downtown
in big East Asian city like Singapore, Hong Kong and Shanghai, and the price of the
space in the city center is highest. The top grade bands tenants always choose to settle
in department store and shopping centers, so the higher retail rent level is found in
department stores and shopping centers. While trade markets and factory outlets often
choose to locate in the place far away from the city center and the cost for land is
usually low there, there for the rent level is expected to be lower in trade markets and
19 / 38
factory outlets.
The variable age indicates that the latest thorough renovation and reopen time
of the retail property. Some papers ［Sirmans, et al. (1992). etc］argue that the age of
the retail property is reversely correlated to the rent level because new retail property
is usually better than the old ones in architecture design and function. However, the
older retail properties may also have an advantage of customer awareness and loyalty,
and new open retail property may lower their rent to attract anchor tenants to lease.
Therefore the model will be estimated in the latter chapters to test the two hypotheses.
The variables total A and retail A represent the total area of the whole building
and the total area of the retail sector of the building respectively. Some properties
have not only retail section in the building, but also office and residential. It is very
common to see some buildings have a supermarket, even a shopping center in the
lower floors, and the higher stories are rented out or sold out for office or housing in
the downtown of many cities. Bigger retail properties are expected to attract bigger
customers’ flows because more function and content can be provided in bigger retail
property to cater for the customers with diverse needs, so the rent is expected to be
higher in giant shopping centre.
The variables Total area and Retail area are transformed to their natural
logarithm because the estimated parameters can be interpreted as elasticties. The
marginal impact of percentage change in area on rent level can be estimated by nature
logarithm.
In order to get better control of the spatial correlation, the variables dtocbd and
dtobunds6 are generated as independent variables to indicate the distance to the CBD
Jiang An district and shopping centre Bund6. The two regions mentioned above have
the highest rent level in Shanghai, and the two independent variables are expected to
be negatively coorelated to the rent level.

20 / 38
5. Results analysis
Before the models are estimated in stata, some specifications need to be stated. First,
no outliers are eliminated from the dataset although higher fitness of the model may
be achieved by doing so, because the number of the effective observations is limited
(85 observations). Second, as the table of correlation of the variables in appendix 3
shows, the correlation between some independent variables is high, for example the
correlation between the variables total area and retail area is 0.581, and the correlation
between retail area and the dummy variable type4 for factory outlets is 0.58. The
problem of multicollinearity would occur if the independent variables with high
correlation are put into the same OLS model to regress against the dependent variable.
Variance inflation factors (VIF) should be derived to test for the problem of
multicollinearity.
The results from the OLS regression models (3.1)-(3.3) specified in chapter 3
are shown in table 2 and appendix 5.
Table.2 model1 and model2 in OLS models

model1 model2(Robust)
coefficient t-value coefficient t-value
lnTotalArea -0.02 -0.29 -0.03 -1.09
lnRetailArea -0.02 -0.42 0.00 0.05
age -0.01 -2.38 -0.01 -2.56
nooflayer 0.01 0.65 0.00 0.42
DG1 0.28 1.10
DG2 0.25 1.11
DG3 0.18 0.84
DG4 0.25 1.19
DG5 0.30 1.40
DG6 0.15 0.77
DG7 0.35 1.60
DG8 0.15 0.67
DG9 0.06 0.28
DG11 0.11 0.46
DG12 0.12 0.55
DG13 -0.07 -0.30
DG14 -0.03 -0.16
D2 0.02 0.42

21 / 38
D3 -0.22 -6.44
D4 -0.22 -5.78
D5 -0.46 -11.08
D6 -0.65 -14.28
type1 -0.48 -1.29 0.22 2.85
type2 -0.53 -1.44 0.21 2.74
type3 -1.20 -3.06 -0.37 -3.04
dtobund6 -2.91 -3.28 0.28 0.62
dtocbd -0.05 -0.04 -0.30 -0.52
_cons -0.56 -1.25 -1.03 -5.44
Number of obs 85.00 85.00
Root MSE 0.18 0.09
F-value 7.91 206.01
R-squared 0.74 0.92
F-value for B-P
test before robust 1.46 5.27
VTF(average) 9.1 8.43
Z for Moran’I
(distance inverse) 0.02 0.02
Z for
Moran’I(square
distance inverse) 1.65 1.55

5.1 Interpretation of OLS model results

The F-values for Breusch-Pagan test are 1.46, 5.27 and 1.03 for the three models
respectively, indicating that the problem of Heteroskedasticity is bothering the
estimation in model (3.2).The standard error of the estimates in model (3.2) is biased
in this case, so the t statistics and F in the model cannot be trusted. To fix this problem,
the heteroskedasticity robust procedure should be used in model (3.2), and the results
interpretation of model (3.2) is based on the robust model. The results of F test for all
the models are high (7.91, 206.01 and 28.86 respectively) rejecting the null hypothesis
that the explanatory variables are jointly not significant to capture the variation of the
percentage rent.
In the first model, the adjusted R-square is 0.74; it means that about 74 percent
of the variation of the dependent variable is explained by the model. The average VIF
22 / 38
for the model is 9.1(more detail about VIF test results can be found in appendix-6),
indicating that the multicollinearity is not a problem. The marginal effects of one
percentage increase in total area and retail area are estimated to be -0.015and -0.023
respectively, but both of the variables are not significant in explaining the dependent
variable. The marginal effect of age is estimated to be negative; the rent for the retail
property is estimated to decrease 1.1 percent if it is one year older. The explaining
power of number of layer is estimated to be positive and insignificant.
The dummy variables 𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑖𝑖 are all significant in explaining the variation of
percentage rent. According to the first model, it is quite surprising that the factory
outlets are estimated to be the most expensive retail type if all the other condition is
the same. The rent for shopping centers, department stores and trade market is 48.4,
53.2 and 119.6 percentages lower than that for outlets. The marginal impacts of
variables distance to Jing An CBD and distance to Bund area are both negative
(-0.051and -2.912 respectively), neither of them is significant.
It is worth noting that the explanatory power of all the dummy variables 𝐷𝐷𝐷𝐷𝑖𝑖
is insignificant in explaining the variation of the independent variable. Although the
joint explaining power of these variables is tested to be significant (the F-value of the
test is 3.78, we can reject the null hypothesis that all the 𝐷𝐷𝐷𝐷𝑖𝑖 can be equal to 0), none
of these variables has t-test value above 1.96 in the regression result. Better
classification with retail research purpose should be defined, other than the district
planed by government.
In the second model where the new districts dummy variables 𝐷𝐷𝑖𝑖 are used,
around 92% of the variation of the independent variables can be explained by the
model. The average VIF for the second model is 8.43(more detail about VIF test
results can be found in appendix-6), indicating that the multicollinearity is not
bothering the estimation. And the F value for second model increase from 7.91 (for
the first model) to 206.01, the null hypothesis that the coefficients for all the
independent variables equal to 0 is strongly rejected.
In the second model, the marginal effect of age is estimated to be negative

23 / 38
(-0.01) and significant (-2.56), the general rent level of retail property is 1% lower if it
is one year older. The number of layer, total area of the property and the retail area of
the property are still not significant in the second model. According to the estimation
of the second model, one percentage increase in the total area and retail area can
induce the general percentage rent increase 3.16% and decrease 0.11% respectively.
The different types of retail are significant explaining the independent variable
in the second model. If all the other conditions are the same, the rent level in shopping
center and department store are estimated to be the higher than factory outlets by 22%
and 21% respectively. The general level of percentage rent for trade market is lower
than the outlets by 37%.
The explaining power of the distance to CBD in Jing An and distance to Bun
area are both insignificant in the second model. The coefficients for the two variables
are -0.299 and 0.283 respectively.
It is very different from the first model in that the dummy variables 𝐷𝐷𝑖𝑖 in the
second model are all significant in explaining the rent level individually and jointly.
The t-value of the all the variables 𝐷𝐷𝑖𝑖 is higher than 1.96. According to the critical
value for the F test (53.01), we can strongly reject the null hypothesis that all the
coefficients of 𝐷𝐷𝑖𝑖 equal to 0. The region with the highest general rent level is
indicated by 𝐷𝐷2 (the district of Jing An CBD). The general rent level in 𝐷𝐷2 is higher
than that in 𝐷𝐷1 by 2%. The average rent level in 𝐷𝐷3 (the sub CBD district in Lu Wan
and Hong Kou), 𝐷𝐷4 (the new CBD region in Pu Dong district) and 𝐷𝐷5 (the districts
that adjacent with 𝐷𝐷1 and𝐷𝐷2 ) is lower than that in 𝐷𝐷2 by 21.55%, 21.66% and 46.31%
respectively. The lowest rent level is found in 𝐷𝐷6 (The districts that are not adjacent
with 𝐷𝐷1 and 𝐷𝐷2 ), it lower than that in 𝐷𝐷1 by 65.38%.
The dummy variables 𝐷𝐷𝑖𝑖 are more appropriate than 𝐷𝐷𝐷𝐷𝑖𝑖 to describe the
spatial characters in retail research context as the explanatory power of the model is
improved substantially (from 74% to 92%) by replacing 𝐷𝐷𝐷𝐷𝑖𝑖 with𝐷𝐷𝑖𝑖 . Moreover, the
dummy variables 𝐷𝐷𝑖𝑖 are all significant in the second model.
Although the estimators of the variables in the second model are very different

24 / 38
form than in the first model, the estimation of the second model is proved more
reliable because more variation of the rent is captured in the second model.
In the third model, integration variables are taken into regression. Although all
the t-values of the independent variables are below 1.96, the explaining power of the
third model is the highest of the three models (91.17% Adjusted R-square), and the F
value is also high (27.64) to reject the null hypothesis that the joint explaining power
of all the independent variables is insignificant. The VIF for the model is 18429(more
detail about VIF test results can be found in appendix-6), indicating that the some
independent variables in the model are severely linearly related with each other
［Rosiers, and Riault, (2001)］. In such case, the variance of the estimators of the
variables would be very high, because the effects of that bias can be multiplied by
orders of magnitude, then the estimation of the coefficients of the variables in model
become unreliable［O'Brien, et al( 2007)］. Nevertheless, the model can be used for
prediction purpose in cross section data set, and if the sample size increases, the
estimation can be more precise.
For the integrated variables, the estimated coefficients for variables from
AgeD1 to AgeD6 are 0.0113, 0.0043, 0.0014, 0.0019 and -0.0021 respectively,
indicating the marginal effects of age on the percentage rate in different districts. For
example, the general rent level in district 1 would increase by 1.13% if the property is
one year older, with the assumption that all the other condition be the same for all the
properties in the district. Similarly, the marginal effect of age on rent level among four
different retail types is estimated to be -0.0078,-0.0053, 0.0935 and -1.5541
respectively. And the marginal effect of total area on the retail rent level among
different district is negative (-0.4245,-0.3395,-0.4216,-0.3467,-4123 and -3663
respectively) while the effect among different retail type is positive (0.3902, 0.3742
and 0.8293 respectively).

25 / 38
5.2 Spatial autocorrelation test and spatial regression
model

The spatial autocorrelation test with distance inversed matrix shows the statistics of Z
score of Moran’s I are very proximate to 0 for the all the three models. However,
when the square inverse distance matrix is used in the test, the statistics of the Z
scores increase to 1.62, 1.55 and 1.52 respectively, the effect of spatial autocorrelation
is close to significant level. In this case, spatial lag model and spatial error model are
constructed to capture the spatial dependency.
The spatial lag model and spatial error model are using the same variables as
the second model to regress. The reason why the first and the third models are not
chosen is that the first model performs poorly in capturing the variation of the rent
and the problem of multicollinearity is bothering the third model.
Table.3 Estimate the determinants of retail rent with spatial regression model.
Spatial lag model Spatial error model OLS model
age -0.00555(-2.72) -0.00567(-2.78) -0.00634(-2.85)
nooflayer 0.002917(0.65) 0.002635(0.59) 0.002308(0.46)
type1 0.18143(1.74) 0.174282(1.64) 0.221329(1.94)
type2 0.179522(1.74) 0.17153(1.63) 0.212405(1.87)
type3 -0.41715(-3.91) -0.42283(-3.9) -0.37221(-3.21)
dtobund6 0.154368(0.36) 0.136733(-.31) 0.282713(0.59)
dtocbd -0.37078(-0.73) -0.36495(-0.71) -0.29905(-0.53)
lnTotalArea -0.02262(-0.89) -0.02341(-0.92) -0.03152(-1.13)
lnRetailArea -0.00413(-0.18) -0.00378(-0.16) 0.001063(0.04)
D2 0.007965(0.15) 0.005727(0.11) 0.016236(0.28)
D3 -0.20622(-5.48) -0.21101(-5.55) -0.21554(-5.16)
D4 -0.23054(-5.52) -0.23591(-5.46) -0.21621(-4.71)
D5 -0.46313(-11.46) -0.47071(-11.41) -0.46314(-10.23)
D6 -0.66169(-14.16) -0.66806(-13.95) -0.65384(-12.55)
_cons -1.0013(-6.44) -0.98249(-6.2) -1.03375(-5.98)

The t-value is within the bracket.

As the table above shows, the difference of estimation between spatial regression
model and the OLS model is very small. Most of the coefficients in the models have
the same sign and significant level, except that the variable of retail area in the OLS
26 / 38
model has a positive coefficient (0.001063) while the sign in spatial regression model
is negative (-0.00413 in spatial lag model and -0.00378 in spatial error model).
Although the sign of retail area is different in the three models, the explanatory power
is all insignificant (-0.18,-0.16 and 0.04 respectively). All the other estimators in the
three models are similar.

5.3 The root mean square error test and model selection

To compare the three competing models and select the most appropriate one to
estimate in practice, the root mean square error (RMSE) test is perform to the three
models［e.g. McCracken,(2000)］. According to the principle of the RMSE test, 75
observations are randomly selected from the total 85 observations to estimate the new
parameters, while the other 10 observations are left for prediction. Then the new
estimated model is used to predict the new residuals in the sample of the 10
observations. The new model with the smallest average deviation is selected. The
derivation procedure is shown in the table in appendix 7. The test results are shown in
the table below:
Table.3 the results of RMSE test.
OLS model Spatial error model Spatial lag model
RMSE 0.0606 0.0592 0.0589

RMSE/mean rent 26.95% 26.34% 26.18%

According to the test result, although the differences of the RMSE between the three
models are very small, the spatial lag model with the smallest deviation of residuals is
selected as the best model. As the table.3 shows, the difference between the estimators
of spatial lag model and OLS model are very small, most of the variables have the
same sign, very similar coefficients and significant levels in the two models, except
the variable of retail area.

27 / 38
6. Conclusion
The empirical study investigates the determinants of retail rent in Shanghai on city
level by taking 85 observations in into regression models. The significant explanatory
variables in the models are the age of the property, the area of retail building and the
distance from the property to the Jing An CBD. The types of the retail property and
the districts that the property located in are also explaining the variation of the retail
rent level among the individuals. Higher retail space price is found in the CBD and
centre of the city, as well as in shopping centre and department store compared to the
trade markets and outlets.
A new classification of district in retail research context is suggested in this paper,
and it is proved to be more appropriate in explaining the retail rent variation among
different retail properties than the districts set by the government.
The degree of spatial autocorrelation is estimated to be close to significant level
when using OLS model. The spatial lag model and spatial error model are constructed
to estimate the determinants of retail rent, and their results are compared with that
from OLS model. In order to find out the most suitable model to estimate in practice,
the root mean square error (RMSE) test is performed, and the spatial lag model is
selected as the best model to estimate.
Although the dataset and regression models give a satisfied result, the estimation
process can still be improved by taking more observation into models. In the third
OLS regression model, the estimators of the marginal effect of independent variables
in different regions and property types are unreliable because of the existence of
multicollinearity. More observations are needed to reduce the problem and give a
better estimation. Besides that, more data about other characteristics of the retail
building (such as the sales revenue of the shopping center) should also be used in the
regression model to test for their explaining power in retail rent.

28 / 38
7. References
1. Akerlof. (1997), “Social distance and social decisions”, Journal of the Econometric Society,

65, 1005–1027.

2. Anselin. (1988) Spatial Econometrics: Methods and Models, Dordecht: Kluwer.

3. Arraiz, I, Drukker, Kelejian, and Prucha. (2008). A Spatial Cliff-Ord-type Model with

Heteroskedastic Innovations: Small and Large Sample Results. Technical report, Department of

Economics, University of Maryland. http://www.econ.umd.edu/ prucha/Papers/WP Hetero GM IV

MC.pdf

4. O’Roarty, McGreal and Adair. (1996), “The impact of retailers’ store selection criteria on the

estimation of retail rents”, Journal of Property Valuation & Investment, Vol. 15 No. 2, 1997, pp.

119-130.

5. Jackson. (2000), “A Model of Spatial Patterns across Local Retail Property Markets in Great

Britain”, Urban Studies, Vol. 38, No. 9, 1445–1471, 2001

6.Sirmans and Krisandra, A. (1992), “The Determinants of Shopping Centre Rent”, Journal of

Real Estate Research, Vol.8 No.1:pp.107-115

7. Carter, Haloupek. (2000), “Spatial Autocorrelation in a Retail Context”, Journal of

International Real Estate Review, 2000, Vol. 3 No. 1: pp. 34 – 48

8. Brounen, Colliander and Eichholtz.(2005),“Corporate real estate and stock performance in the

international retail sector”, Journal of Corporate Real Estate ,Vol. 7 No. 4, 2005, pp. 287–299.

9 D'Arcy, McGough. (1997),“An empirical investigation of retail rents in five European cities”,

Journal of Property Valuation &Investment, Vol. 15 No. 4, 1997,pp. 308-322. © MCB University

Press, 0960-2712

10. Hui , Yiu and Yau.(2007),“Retail properties in Hong Kong: a rental analysis”, Journal of

Property Investment &Finance ,Vol. 25 No. 2, 2007 pp. 136-146

11. Rosiers, Thériault and Menetrier (2004), “spatial versus non-spatial determinants of shopping

center rents: modeling location and neighborhood-related factors”, Journal of Real Estate

Research, Vol. 21, Number3, 2001, pp.1-29.

12.Chun, Eppli, Shilling.(2001), “A Simulation Analysis of the relationship between Retail Sales

29 / 38
and Shopping Centre Rents”, Journal of Real Estate Research, Vol. 21,Number3,2001,pp.163-186.

13. Lakshmanan, Hansen (1965), “A retail market potential model”, Journal of the American

Planning Association, 31:2,134 — 143

14. Hanna, Lake and Caples (2007), “Factors impacting price for retail space in houston”, Journal

of Economics and Economic Education Research, Volume 8, Number 2, 2007

15. Hendershott and MacGregor (2005), “Investor Rationality: Evidence from U.K. Property

Capitalization Rates”, Journal of Real Estate Economics, 2005 V33 2: pp. 299–322

16. Hendershott, MacGregor and White (2005), “Explaining Real Commercial Rents Using an

Error Correction Model with Panel Data” ， Journal of Real Estate Finance and

Economics,24:1/2,59-87

17. Sivitanidou.(1994), “Urban Spatial Variations in Office-Commercial Rent: The Role of Spatial

Amenities and Commercial Zoning ”,Journal of Urban Economics 38.23-49(1995).

18. Chaplin. (2000), “Predicting real estate rents: walking backwards into the future”, Journal of

Property Investment & Finance, Vol. 18 No. 3, 2000, pp. 352-370.

19. Wang and Zhang (2006), “Opportunities and Challenges of Shopping Centre Development in

China: A Case Study of Shanghai”, Journal of Shopping Center Research, Volume 13, Number 1,

2006

20. Tsolacos (1995), “An Econometric Model of Retail Rent in the United Kingdom”, Journal of

Real Estate Research, Vol. 5,Number 5,1995,pp.519-529

21. Warnes and Ripley (1987). Problems with likelihood estimation of covariance functions

of spatial Gaussian processes. Biometrika 74, 640–642.

22. Hardin and Wolverton (2002), “A n Empirical Analysis of Communities Center Rent”,

Journal of Real Estate Research, Vol.23, N o s. 1 / 2 – 2 0 0 2

30 / 38
8. Appendix

Appendix 1- retail related figures about Shanghai

Turnover Population GDP GDP per service

(100 million) (10 thousand) (100 million) capital sector/GDP
1978 54.1 1104 272.81 2497 19%
1979 68.28 1137 286.43 2568 19%
1980 80.43 1152 311.89 2737 21%
1981 88.73 1168 324.76 2813 22%
1982 89.8 1186 337.07 2877 22%
1983 100.68 1201 351.81 2963 24%
1984 123.72 1217 390.85 3259 25%
1985 173.39 1233 466.75 3855 26%
1986 196.84 1249 490.83 4008 28%
1987 225.25 1265 545.46 4396 29%
1988 295.83 1288 648.3 5162 29%
1989 331.38 1311 696.54 5487 29%
1990 333.86 1334 781.66 6107 31%
1991 382.06 1350 893.77 6954 35%
1992 464.82 1365 1114.32 8650 36%
1993 675.92 1381 1519.23 10729 38%
1994 834.76 1398 1990.86 13807 40%
1995 1050.96 1414 2499.43 17022 41%
1996 1258 1451 2957.55 19779 44%
1997 1435.38 1489 3438.79 22583 46%
1998 1593.27 1527 3801.09 24513 49%
1999 1722.33 1567 4188.73 26527 51%
2000 1865.28 1608.6 4771.17 29671 52%
2001 2016.37 1614 5210.12 32201 52%
2002 2203.89 1625 5741.03 35329 53%
2003 2404.45 1711 6694.23 39128 51%
2004 2656.91 1742.15 8072.83 46338 51%
2005 2972.97 1778.42 9164.1 51529 50%
2006 3360.41 1815.34 10366.37 57695 51%
2007 3847.79 1858.08 12188.85 66367 53%
The figures are downloaded from Statistics Bureaucracy of China
(http://www.stats-sh.gov.cn/2008shtj/index.asp)

31 / 38
Appendix 2-weight matrix

The weight matrix for spatial autocorrelation test:

1 2 3 4 5 6 7 8 …… 96 97 98 99 100 101 102 103
1 0.00 1.77 242.32 423.85 569.41 509.25 529.46 424.00 …… 507.15 505.78 114.59 114.33 114.04 113.05 278.13 189.84
2 1.77 0.00 266.18 461.88 603.26 548.83 569.49 461.77 …… 553.91 552.52 144.01 143.95 143.55 142.31 314.03 218.12
3 242.32 266.18 0.00 33.46 69.21 53.93 59.92 32.86 …… 76.88 76.63 115.61 123.75 120.84 116.62 26.62 16.04
4 423.85 461.88 33.46 0.00 28.99 4.74 6.97 0.02 …… 11.04 11.04 174.79 183.81 180.78 176.71 23.35 48.40
5 569.41 603.26 69.21 28.99 0.00 13.93 12.24 27.79 …… 55.94 56.32 325.75 338.75 334.30 328.09 98.93 126.89
6 509.25 548.83 53.93 4.74 13.93 0.00 0.22 4.50 …… 14.27 14.44 237.09 247.55 244.05 239.32 49.03 82.23
7 529.46 569.49 59.92 6.97 12.24 0.22 0.00 6.68 …… 15.86 16.07 251.51 262.26 258.66 253.81 55.65 90.81
8 424.00 461.77 32.86 0.02 27.79 4.50 6.68 0.00 …… 11.79 11.79 176.47 185.57 182.51 178.39 24.07 48.77
9 493.88 533.08 49.52 3.33 15.44 0.13 0.67 3.12 …… 13.43 13.58 226.39 236.63 233.20 228.56 44.28 75.91
10 429.69 467.17 33.11 0.17 24.77 3.63 5.61 0.09 …… 12.96 12.98 182.97 192.28 189.15 184.90 26.67 51.38
.. ….. ….. …… …… …… …… …… …… …… …… …… …… …… …… …… …… …..
94 551.09 600.68 97.13 18.94 64.80 19.97 21.00 19.85 …… 1.18 1.22 213.62 221.77 219.19 216.02 46.22 95.34
95 538.49 587.39 91.75 16.83 63.00 18.62 19.84 17.71 …… 0.66 0.68 207.01 215.11 212.54 209.36 42.64 90.02
96 507.15 553.91 76.88 11.04 55.94 14.27 15.86 11.79 …… 0.00 0.00 192.96 201.10 198.48 195.20 34.43 76.94
97 505.78 552.52 76.63 11.04 56.32 14.44 16.07 11.79 …… 0.00 0.00 191.93 200.03 197.43 194.16 34.05 76.43
98 114.59 144.01 115.61 174.79 325.75 237.09 251.51 176.47 …… 192.96 191.93 0.00 0.14 0.06 0.01 70.77 47.84
99 114.33 143.95 123.75 183.81 338.75 247.55 262.26 185.57 …… 201.10 200.03 0.14 0.00 0.02 0.11 76.43 53.07
100 114.04 143.55 120.84 180.78 334.30 244.05 258.66 182.51 …… 198.48 197.43 0.06 0.02 0.00 0.04 74.53 51.22
101 113.05 142.31 116.62 176.71 328.09 239.32 253.81 178.39 …… 195.20 194.16 0.01 0.11 0.04 0.00 72.03 48.61
102 278.13 314.03 26.62 23.35 98.93 49.03 55.65 24.07 …… 34.43 34.05 70.77 76.43 74.53 72.03 0.00 9.49
103 189.84 218.12 16.04 48.40 126.89 82.23 90.81 48.77 …… 76.94 76.43 47.84 53.07 51.22 48.61 9.49 0.00

32 / 38
Appendix 3-formulas for Z sore

The formulas derivation of Z score for Moran’I

I − E(I)
Z= � (8.1)
�var(I)

The value of Z that indicating no spatial autocorrelation is obtained by:

E (I) = −1�N − 1 (8.2)

The variance of Moran’I is

NS − S3 S5
Var(I)= 4 �(N − 1)(N − 2)(N − 3)(∑ ∑ w )2 (8.3)
i j ij

Where,

∑i ∑j (wij + wji )
S1 = �
2

S2= � (� wij + � wji )2

i j
j

N −1 ∑i ei 4�
S3 = (N−1 ∑i ei 2 )2

S4 = (N 2 − 3N + 3)S1 − NS2 + 3(� � wij )2

i j

S5 = 6(� � wij )2
i j

33 / 38
Appendix 4-variables correlation

The correlation between variables:

y persen~t dtobund6 dtocbd D1 D2 D3 D4 D5 D6 age T1 T2 T3 T4 noofla~r openyear totala~2 retai~im

y 1.00

persentage~t 0.99 1.00

dtobund6 -0.63 -0.63 1.00

dtocbd -0.52 -0.52 0.86 1.00

D1 0.44 0.49 -0.29 -0.13 1.00

D2 0.36 0.42 -0.10 -0.22 -0.09 1.00

D3 0.27 0.25 -0.31 -0.27 -0.17 -0.12 1.00

D4 0.14 0.12 -0.16 0.13 -0.12 -0.09 -0.17 1.00

D5 -0.31 -0.35 0.09 -0.22 -0.26 -0.19 -0.37 -0.26 1.00

D6 -0.60 -0.59 0.64 0.71 -0.16 -0.11 -0.22 -0.16 -0.34 1.00

age -0.04 -0.04 -0.12 -0.12 0.01 -0.06 0.16 -0.12 0.07 -0.14 1.00

T1 0.30 0.29 -0.22 -0.11 0.13 0.11 -0.09 0.21 -0.09 -0.15 -0.31 1.00

T2 0.04 0.00 -0.08 -0.15 -0.08 -0.07 0.17 -0.17 0.10 -0.05 0.32 -0.86 1.00

T3 -0.60 -0.50 0.16 0.05 -0.08 -0.06 -0.11 -0.08 0.06 0.20 0.09 -0.27 -0.16 1.00

T4 -0.26 -0.24 0.73 0.72 -0.05 -0.04 -0.08 -0.05 -0.12 0.35 -0.12 -0.19 -0.11 -0.04 1.00

nooflayer 0.35 0.35 -0.31 -0.26 0.30 0.11 -0.13 0.08 -0.02 -0.24 -0.02 0.13 0.01 -0.20 -0.17 1.00

openyear 0.04 0.04 0.12 0.12 -0.01 0.06 -0.16 0.12 -0.07 0.14 -1.00 0.31 -0.32 -0.09 0.12 0.02 1.00

totalareas~2 -0.30 -0.29 0.25 0.22 -0.14 -0.03 -0.27 0.09 0.15 0.15 -0.16 0.23 -0.32 0.12 0.08 0.11 0.16 1.00

retailgfas~m -0.35 -0.33 0.45 0.42 -0.13 -0.03 -0.18 0.04 -0.02 0.31 -0.15 -0.01 -0.22 0.09 0.58 -0.01 0.15 0.58 1.00

34 / 38
Appendix 5-estimate with OLS models

model1 model2(Robust) model3 model1(robusted) model2(robusted) model3

coefficient t-value coefficient t-value coefficient t-value coefficient t-value coefficient t-value coefficient t-value
lnTotalArea -0.02 -0.29 -0.03 -1.09 agetype1 -0.01 -0.61
lnRetailArea -0.02 -0.42 0.00 0.05 0.01 0.27 agetype2 -0.01 -0.40
age -0.01 -2.38 -0.01 -2.56 agetype3 0.09 0.48
nooflayer 0.01 0.65 0.00 0.42 0.00 -0.07 agetype4 -1.55 -0.63
DG1 0.28 1.10 TaD1 -0.42 -0.92
DG2 0.25 1.11 TaD2 -0.39 -0.74
DG3 0.18 0.84 TaD3 -0.41 -0.93
DG4 0.25 1.19 TaD4 -0.35 -0.76
DG5 0.30 1.40 TaD5 -0.41 -0.92
DG6 0.15 0.77 TaD6 -0.36 -0.80
DG7 0.35 1.60 TaType1 0.39 0.86
DG8 0.15 0.67 TaType2 0.37 0.83
DG9 0.06 0.28 TaType4 0.83 0.75
DG11 0.11 0.46 D1Type1 0.29 0.10
DG12 0.12 0.55 D1Type2 0.80 0.77
DG13 -0.07 -0.30 D2Type2 0.66 0.23
DG14 -0.03 -0.16 D3Type2 0.64 0.22
D2 0.02 0.42 D4Type1 -0.63 -0.21
D3 -0.22 -6.44 0.11 0.04 D5Type2 -2.17 -0.49
D4 -0.22 -5.78 D5Type1 -2.82 -0.83

35 / 38
D5 -0.46 -11.08 2.63 0.61 D6Type1 -2.37 -1.24
D6 -0.65 -14.28 1.38 0.40 D6Type2 -1.62 -0.47
type1 -0.48 -1.29 0.22 2.85 _cons -0.56 -1.25 -1.03 -5.44 -1.27 -0.44
type2 -0.53 -1.44 0.21 2.74 -0.50 -0.17 Number of obs 85.00 85.00 85.00
type3 -1.20 -3.06 -0.37 -3.04 Root MSE 0.18 0.09 0.09
dtobund6 -2.91 -3.28 0.28 0.62 F-value 7.91 206.01 28.86
dtocbd -0.05 -0.04 -0.30 -0.52 -0.12 -0.22 R-squared 0.74 0.92 0.95
F-value for B-P
test before
ageD1 0.01 0.76 robust 1.46 5.27 1.03
ageD2 0.01 0.22 VTF(average) 9.1 8.43 18429.40
Z for Moran’I
(distance
ageD3 0.00 0.12 inverse) 0.02 0.02 0.02
Z for
Moran’I(square
distance
ageD6 0.00 -0.02 inverse) 1.65 1.55 1.52

36 / 38
Appendix 6-VIF test results

Model(3.1) Model(3.2) Model(3.3) Model(3.3)

Variable VIF 1/VIF Variable VIF 1/VIF Variable VIF 1/VIF Variable VIF 1/VIF

type1 53.3 0.0 type1 31.6 0.0 TaType1 70270.6 0.00001 D4Type1 8721.0 0.0001

DG4 13.9 0.1 type2 29.0 0.0 TaD5 60157.3 0.00002 D1Type1 8620.0 0.0001

type3 11.7 0.1 dtocbd 10.1 0.1 TaType2 55717.9 0.00002 D6Type2 6332.0 0.0002

dtobund6 10.5 0.1 dtobund6 9.1 0.1 D5 46960.7 0.00002 D6Type1 2813.3 0.0004

dtocbd 9.8 0.1 lnTotalArea 7.2 0.1 TaD3 46344.5 0.00002 agetype3 2682.4 0.0004

DG8 9.3 0.1 lnRetailArea 6.9 0.1 TaType4 45160.6 0.00002 D2Type2 1149.7 0.0009

DG7 9.2 0.1 type3 6.1 0.2 agetype4 41522.1 0.00002 D1Type2 289.8 0.0035

DG2 8.2 0.1 D5 4.5 0.2 TaD6 38251.3 0.00003 agetype2 99.3 0.0101

lnRetailArea 8.1 0.1 D6 3.6 0.3 TaD4 29736.6 0.00003 ageD3 61.3 0.0163

lnTotalArea 7.9 0.1 D3 3.3 0.3 TaD1 27431.1 0.00004 ageD5 59.0 0.0170

DG6 7.5 0.1 D4 2.0 0.5 D5Type2 25761.7 0.00004 ageD6 57.5 0.0174

DG5 6.7 0.1 D2 2.0 0.5 type2 22911.3 0.00004 agetype1 56.3 0.0178

DG12 5.9 0.2 nooflayer 1.5 0.7 TaD2 21997.0 0.00005 ageD1 31.1 0.0321

DG13 5.8 0.2 age 1.2 0.8 D3 18500.3 0.00005 ageD2 23.6 0.0424

DG3 5.7 0.2 Mean VIF 8.4 D6 17870.6 0.00006 lnRetailAr 13.1 0.0764

DG1 5.2 0.2 D5Type1 17335.6 0.00006 dtocbd 9.7 0.1028

DG9 5.0 0.2 D3Type2 9648.5 0.00010 nooflayer 2.6 0.3831

DG11 2.6 0.4 Mean VIF 18429.4

DG14 1.8 0.6

nooflayer 1.7 0.6

age 1.5 0.7

Mean VIF 9.1

Appendix 7-formulas for RMSE test

The formulas for the root mean square error (RMSE) test are shown as following:

𝑌𝑌𝑖𝑖 = 𝑎𝑎 + ∑𝑘𝑘𝑛𝑛 =1 𝛽𝛽1𝑛𝑛 𝑋𝑋𝑛𝑛𝑛𝑛 + 𝜀𝜀1𝑖𝑖 ,i∈ ［1 , T］ (A.4.1)

Where, 𝑌𝑌𝑖𝑖 and 𝑋𝑋𝑛𝑛𝑛𝑛 are the dependent and independent variables for observation i,
the number of the independent variables is n. And 𝛽𝛽1𝑛𝑛 is the estimated coefficient for
independent variable𝑋𝑋𝑛𝑛𝑛𝑛 , 𝜀𝜀1𝑖𝑖 is the residual for observation i in the model (A.4.1).
The total number of the observation is T.
Then the sub sample are selected randomly, let the number of the sub sample be S.
New model is estimated based on the sub sample as below:
37 / 38
𝑌𝑌𝑖𝑖 = 𝑎𝑎 + ∑𝑘𝑘𝑛𝑛 =1 𝛽𝛽2𝑛𝑛 𝑋𝑋𝑛𝑛𝑛𝑛 + 𝜀𝜀2𝑖𝑖 ,i∈ ［1 , S］ (A.4.2)

Then the errors for the rest observations (T-S) by the model (A.4.2) are estimated as:

𝜀𝜀3𝑖𝑖 = 𝑌𝑌𝑖𝑖 − 𝑎𝑎 − ∑𝑘𝑘𝑛𝑛 =1 𝛽𝛽2𝑛𝑛 𝑋𝑋𝑛𝑛𝑛𝑛 ,i∈ ［1 , (T − S)］ (A.4.3)

Finally, the RMSE statistics equals to the average deviations of model (A.4.3) in the
(T-S) sample.
2 (𝜀𝜀 )2
3𝑖𝑖
𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅 = �(𝑇𝑇−𝑆𝑆) , i∈ ［1 , (T − S)］ (A.4.4)

38 / 38