PhD-FDEF-2020-04

The Faculty of Law, Economics and Finance

DISSERTATION

Defense held on 04 February 2020 in Luxembourg

To obtain the degree of

DOCTEUR DE L’UNIVERSITÉ DU LUXEMBOURG

EN SCIENCES ÉCONOMIQUES
by

TRAN Thi Thu Huyen

Born on 16 June 1990 in Nam Dinh (VIETNAM)

ESSAYS ON URBAN GREEN AMENITY

AND CITY STRUCTURE

Dissertation defense committee

Prof. Dr. Pierre M. Picard (supervisor, Université du Luxembourg)
Prof. Dr. Geoffrey Caruso (president of the jury, Université du Luxembourg / LISER)
Assoc. Prof. Dr. Antonio Cosma (vice president of the jury, Université du Luxembourg)
Prof. Dr. Jos van Ommeren (Vrij Universiteit Amsterdam)
Prof. Dr. Elisabet Viladecans Marsal (Universitat de Barcelona)
Preface

I could not have completed this PhD dissertation without the support of many col-
leagues, friends and family.

Pursuing a PhD in urban economics is not the path I envisioned following five years
ago, when I first came to the Toulouse School of Economics for the master’s program. I
became enthusiastic about urban economics when I did an internship at the Transport
Policies and Development Division at UN ESCAP in 2014. I researched the impact of
transport networks on economic activities, mostly in Southeast Asian countries. My ex-
perience at UN ESCAP inspired me about cities, geographies, transport and amenities.
My PhD journey officially began four years ago when I joined the SOS-BUGS project
on ‘Spatial Optima and the Social Benefits of Urban Green Space’ at the University of
Luxembourg.
I would first like to thank my primary supervisor, Prof. Dr. Pierre M. Picard, for his
guidance, support and encouragement for these past four years. Our discussions and
work together have taught me a great deal about research approaches, and his efforts
will undoubtedly shape the approaches I take to future projects. Thanks for recognizing
the potential in me, encouraging and guiding me while still providing me with enough
freedom to try on my own (and often make necessary mistakes).
I would also like to thank my second supervisor, Prof. Dr. Geoffrey Caruso. Thanks
to his generosity, I had the opportunity to interact and connect with researchers and
colleagues at Urban Planning and Geographical Centres for a semester. I learned a
great deal from his useful comments and advice on my research. His vast knowledge,
broad interests, and kindness will continue to inspire me in the future.
I am genuinely grateful for my third promoter, Assoc. Prof. Dr. Luisito Bertinelli.
Although we worked on relatively few projects together, I learned much from his help-
ful comments, attention to detail and chiefly from his availability to discuss and resolve
any issue that arose.
I was also fortunate to have been able to visit the Spatial Economic Department
at the Vrije Universiteit (VU) Amsterdam in 2017 for four months. Prof. Dr. Jos van
Ommeren, thanks for allowing me to spend some time at VU. Our discussions always
made me consider matters from different perspectives and significantly improved the
research. I genuinely appreciate his guidance and valuable remarks.

i
ii

I am truly grateful to my committee members, Prof. Dr. Elisabet Viladecans Marsal

and Assoc. Prof. Dr. Antonio Cosma, for their reading of the thesis and helpful com-
ments. I am particularly thankful to Dr. Evgenii Monastyrenko, who generously read
and gave considerable feedback on multiple drafts of the chapters in my dissertation.
The research center in economic and management (CREA) at the University of Lux-
embourg has been a great place to conduct research and to grow professionally. I thank
all CREA members who have contributed to such a great working environment. I am
also thankful to all my colleagues for creating a pleasant and friendly atmosphere and
the interesting talks over coffee breaks and beers at the doctoral school events. I would
particularly like to thank my office-mate, Ka Kit Iong, for easing my arrival at the uni-
versity, for tolerating the noise in our office, and for interesting political discussions.
His advice and conversions made the entire experience much better. I am also thankful
to Marina LeGrand, Elisa Gesellchen-Ferreira Da Silva, and Roswitha Glorieux from
the secretary office for providing me with such great support over the past few years.
Finally, I would like to thank my family. To my parents and my two little sisters,
Trang and Thuy, thanks for the love, encouragement and constant support. Without
them, this work would not have been possible.

Tran Thi Thu Huyen

Luxembourg, December 15, 2019
Abstract

Sustainable urban living is a growing issue in many regions worldwide. The debates
over compact city policies and the provision of green areas within urban landscapes are
increasing day-by-day as the number of urban dwellers has soared in recent decades.
This dissertation focuses on studying the optimal provision of green urban areas and
the welfare effects of a substantial change in green provision policies in the presence
of other types of land uses and adverse shocks. It comprises the following four papers
(chapters).

Green Urban Areas (joint with Pierre M. Picard). This paper studies the size and
location of urban green areas across city spaces. Urban green areas offer amenities
that affect residential choices, land consumption and land rent. This paper discusses
the socially optimal sizes and locations of urban green areas within a city and their
decentralized allocation through land markets. The main result is that the share of land
dedicated to urban green areas is a concave function of the distance to the city center.
This result is confirmed by the empirical study of urban structures in the 305 largest EU
cities.
Welfare Evaluation of Green Urban Areas (joint with Pierre M. Picard). This paper
quantifies the impacts of urban green areas on city structures under competitive land
markets for more than 300 of the most populous cities in Europe. We first introduce
the amenity of local and land-intensive green areas within an urban model. We utilize
two different classes of preferences: standard Cobb-Douglas preference and hyperbolic
preference. The theoretical equilibrium is then used to instruct empirical regressions for
residential land uses. Notably, we calibrate the model using data on detailed residential
land uses, green urban areas and population density. We back up the parameters and
run counterfactual simulations under the scenario that urban planners reduce green
areas by half. We consider the general equilibrium effects in households’ willingness
to pay for such a policy and the changes in the land market for both closed and open
cities with two land conversion policies. We show that the gross benefits of urban green
areas are substantial and that households are required to compensate up to 6.4 − 9%
of their yearly income on average to reach the same utility level before the changes.
However, the net cost of urban green areas when we allow for land conversion into
new residential land is negative, estimated at approximately 4% of annual household

iii
iv ABSTRACT

income.
Geographical Stratification of Green Urban Areas (joint with Pierre M. Picard).
This paper studies the provision of urban green areas in cities when residents have
preferences for the size of and access to those areas. At the optimum, the number of
urban green spaces is a non-monotone function of distance to the city centre, while
the sizes and distances to other urban green areas increase as one moves to the urban
fringe. This paper empirically investigates those properties for the 300 largest European
cities by using the GMES Urban Atlas database (European Environmental Agency).
The empirical analysis confirms the non-monotone relationship between the number
of urban green spaces and the distance to the city centre. The distance between two
parks also increases as one moves toward the urban fringe. Finally, richer cities are
associated with a denser network of urban green areas.
Urban Floods, Amenity Land Uses and City Structure. This paper investigates the
effects of concurrent negative shocks in the form of urban flood risks on the amenity
land use planning and city structures. We examine both the first-best and second-
best land use policy when parts of the cities are in flood-prone areas. We find a non-
monotone relationship of public mitigation infrastructures and amenity provision with
the level of flood risks. The outcomes of the first-best optimization are decentralized
through market equilibrium through locational taxation. However, an income tax used
to finance the flood prevention would distort the first-best allocation and induce more
housing developments with lower flood mitigation efforts per unit of land in flood-
prone areas compared to the number in the first-best scenario. For the empirical part,
we use multiple return periods of flood hazard maps in Europe together with the GMES
Urban Atlas to examine the actual land use in 305 of the most populated urban areas in
Europe. To our knowledge, this is the first paper that discusses both theoretically and
empirically the urban land uses under flood risk across all of Europe.
Contents

1 Introduction 1

2 Green Urban Areas 5

2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.2 Theoretical model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.2.1 Social optimum . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.2.2 Comparative statics . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.2.3 City border and population . . . . . . . . . . . . . . . . . . . . . . 12
2.2.4 Land regulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.2.5 Competitive land market equilibrium . . . . . . . . . . . . . . . . 14
2.2.6 Specific preferences . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.3 Empirical analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
2.3.1 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
2.3.2 Urban green area profiles . . . . . . . . . . . . . . . . . . . . . . . 25
2.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31

3 Welfare Evaluation of Green Urban Areas 41

3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
3.2 Literature review . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
3.3 Theoretical model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
3.3.1 Demand side . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
3.3.2 Competitive land market equilibrium . . . . . . . . . . . . . . . . 47
3.3.3 Specific preferences . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
3.4 Empirical analysis and welfare valuation . . . . . . . . . . . . . . . . . . 50
3.4.1 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
3.4.2 Estimation and valuation with Cobb-Douglas preferences . . . . 54
3.4.3 Estimation and valuation with hyperbolic preferences . . . . . . 65
3.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71

4 Geographical Stratification of Green Urban Areas 85

4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
4.2 The Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89

v
vi CONTENTS

4.2.1 Household preferences . . . . . . . . . . . . . . . . . . . . . . . . . 89

4.2.2 Optimal green urban spaces . . . . . . . . . . . . . . . . . . . . . . 91
4.2.3 Cobb-Douglas preferences . . . . . . . . . . . . . . . . . . . . . . . 92
4.3 Empirical analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
4.3.1 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
4.3.2 Descriptive statistics . . . . . . . . . . . . . . . . . . . . . . . . . . 100
4.3.3 Number of green urban spaces . . . . . . . . . . . . . . . . . . . . 102
4.3.4 Size of green urban spaces . . . . . . . . . . . . . . . . . . . . . . . 105
4.3.5 Distance to other green urban spaces . . . . . . . . . . . . . . . . . 106
4.3.6 Population density . . . . . . . . . . . . . . . . . . . . . . . . . . . 110
4.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114

5 Urban Floods, Amenity Land and City Structure 127

5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
5.2 Theoretical model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130
5.2.1 Model settings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130
5.2.2 First-best world . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132
5.2.3 Second-best world . . . . . . . . . . . . . . . . . . . . . . . . . . . 138
5.3 Empirical analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143
5.3.1 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143
5.3.2 Population density in floodplain areas . . . . . . . . . . . . . . . . 150
5.3.3 Housing construction density . . . . . . . . . . . . . . . . . . . . . 153
5.3.4 Flood prevention infrastructure systems . . . . . . . . . . . . . . . 156
5.3.5 Land use for local public goods within flooded zone . . . . . . . 159
5.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161

Bibliography 173
Chapter 1

Introduction

In recent centuries, there has been a massive increase in the number of urban dwellers.
Their share has grown from approximately less than 8 percent in 1800 to more than
50 percent in recent years (Glaeser, 2011; Glaeser, 2012; UN DESA, 2018). The emer-
gence of large metropolises such as Tokyo, New York, London, Mexico City and Mum-
bai has become a more familiar scene. There are many reasons for this phenomenon:
the agglomeration benefits of firm clusters (Glaeser and Gottlieb, 2009; Ellison et al.,
2010), thick and dynamic labor markets (Fallick et al., 2006), social interaction benefits
(Mossay and Picard, 2011; Picard and Tabuchi, 2013) or the provision of numerous con-
sumption amenities due to the availability of a sufficient mass of customers (Glaeser et
al., 2001; Rappaport, 2008). Due to this increasingly urbanized pattern, it is predicted
that the population share living in urban areas will increase to 70%. If this trend is ir-
reversible, then the next question should concern how we can make such urbanization
work. How can we improve urban quality of life and make city life sustainable?
One of the main components that make cities attractive to their residents is their
public park and garden systems. Green urban areas from a small community garden
to famous areas such as ‘Jardin du Luxembourg’ in Paris not only shape the face of the
city but are a quintessential aspect of the quality of life for local inhabitants. They offer
places for local recreation, beautiful views, cleaner air and many other advantages. Re-
cent research has validated the connection between urban parks and the well-being of
the city’s inhabitants. Bowler et al. (2010) and Manes et al. (2012) find a strong influ-
ence of urban natural ecosystems in mitigating air and noise pollution and urban heat
island effects. Heidt and Neef (2008) suggest economic benefits by showing a signif-
icant increase in nearby property values. Access to nearby urban parks helps reduce
stress, improve psychological well-being, and increase physical activity (Colding and
Barthel, 2013; Saldivar-Tanaka and Krasny, 2004; White et al., 2013).
Although green urban areas might seem to be meagre in comparison with other
natural ecosystems such as wetlands or forests, the value of the environmental, recre-
ational, and other services they offer is likely to be disproportionally high due to their

1
2 CHAPTER 1. INTRODUCTION

strategic locations. The benefits are at the local level and right at urban dwellers’
doorsteps. However, the provision of urban parks often entails high opportunity costs
because it demands land at locations where it is the most precious and scarce. Recent
studies report that some land development restrictions to preserve open spaces can dis-
tort land use allocations and reduce welfare (Cheshire and Sheppard, 2002; Turner et
al., 2014). To correctly identify optimal policies, it is necessary to quantify the trade-offs
between urban green areas and other types of land use as well as the welfare effects of
current green areas.

This dissertation includes four research papers that use theoretical, empirical and
simulation approaches to, first, explore the principles of optimal urban green provi-
sions. The theoretical models are then applied to real-life data on city land uses to
verify the predictions of the models and to estimate the welfare effects of a substantial
change in urban green areas.
Chapter two of the dissertation studies the optimal level of small and scattered
green urban areas in urban spaces in continuous fashion. Because green urban areas
are land-intensive and offer highly localized amenities, we find that the optimal share
of surface area devoted to green urban areas is a concave and non-monotonic function
of the distance to the city center. On the one hand, the opportunity cost of land is too
expensive in the city center for the planner to situate many green areas there. On the
other hand, residents are too sparse at the city edges to give planners the incentives
to invest in such highly localized amenities. The main contribution of the paper is to
confirm our results using the very detailed geographical database of the European En-
vironment Agency’s Urban Atlas data, which describes the land use of 305 EU cities
with more than 100,000 inhabitants. These data report land use and cover all of Eu-
rope using harmonized earth observations (EOs). The data represent a unique source
of reliable and comparable European urban planning data. To the best of our knowl-
edge, this is the first paper that uses both a theoretical model and empirical estimation
of European urban land uses. This result is robust to many variations of the land use
specifications, city structure specifications, and city and country characteristics.
The third chapter further utilizes the approach employed in the second chapter in a
setting with competitive land markets to estimate the welfare effects of a hypothetically
large change in green urban areas in European cities. We employ two different house-
hold preferences: Cobb-Douglas and hyperbolic linear utility. The marginal benefit of
urban green areas varies across locations within cities concerning their existing cover-
age in residential areas, the population density and the opportunity cost of land at each
location. The market equilibrium results in optimal residential land use at each loca-
tion across urban areas with a specified level of urban green areas. We then calibrate
the equilibrium of residential land using the data on detailed residential land cover,
green urban areas and population density. To control for the potential endogeneity of
 3

urban green provision from urban authorities, we employ the green areas in the past
as an instrument for the current green areas. We store the parameters and run counter-
factual simulations in a scenario in which urban planners reduce green areas by half.
By using this approach, we can account for the new equilibrium in the housing market
and households’ choices over housing consumption. We consider both the case of open
cities with free migration and closed cities with no migration. We show that the gross
benefits of urban green areas are substantial, and households require compensation of
up to 6.4 − 9% of their yearly income on average to reach the same utility level before
the changes. However, the net cost of urban green areas when we allow for land con-
version into new residential land is negative, estimated at approximately 4% of annual
household income.
In the fourth chapter, we examine the optimal geographical configurations of urban
parks in discrete fashion. We discuss the optimal numbers, sizes and distances of small
and scattered parks across different city locations. At the optimum, the number of
urban green spaces is a non-monotonic function of distance to the city center, while
the sizes and distances to other urban green areas increase as one moves to the urban
fringe. This paper empirically investigates those properties for the 300 largest European
cities by using the GMES Urban Atlas database. Chapter two uses the same database to
study the trade-off between residential space and green urban areas, and we highlight
a non-linear relationship between the land share of green urban areas and the distance
to the city center. By contrast, this chapter disentangles the effect of the latter on the
size, number and dispersion of parks. We find a non-monotonic relationship between
the number of urban green spaces and the distance to the city center and a dispersed
pattern in the optimal distance between two parks as one moves toward the urban
fringe. Furthermore, wealthier cities are associated with a denser network of urban
green areas.
The last chapter additionally investigates the negative shock of urban river flood
risks to amenity land use planning and city structures. We examine both the first-best
and second-best land use policy when parts of the cities are in flood-prone areas. We
find a non-monotonic relationship between amenity land and flood risk. Under the
first-best scenario, locations with higher flood risks in dense areas are more likely to
have higher amounts of amenity land, while the opposite is true in less dense areas.
The pattern is similar under the second-best scenario; however, an income tax used to
finance flood prevention would distort the first-best allocation and induce more people
to locate in the flood-prone area compared to the first-best. In the empirical analysis,
we use the 500-year flood hazard maps in Europe together with the GMES Urban Atlas
to examine the actual land use in the 305 most populated urban areas in Europe. To
the best of our knowledge, this is the first paper that discusses both theoretically and
empirically urban land uses under increased flood risk scenarios across Europe.
4 CHAPTER 1. INTRODUCTION

The findings presented in this dissertation contribute to our understanding of the

optimal land use planning and welfare effects of land-intensive public goods in the con-
text of a non-uniform population distribution. It also incorporates the adverse risk of
urban river floods into green provision and land use planning. The policies on green ar-
eas and urban flood risk are commonly mentioned and debated in many regions world-
wide. This discussion is also particularly relevant for the design and implementation
of future urban areas and policies. A better understanding of the potential channels in
which green areas and urban river flooding affect people’s choices and locations can be
used to examine future policy changes.
Chapter 2

Green Urban Areas

2.1 Introduction
Urban green areas play a crucial role in the debate on sustainable cities. They are an
important part of any urban area whose quality and quantity are prime concerns for
environmental sustainability. Recent research has confirmed the relationship between
urban parks and the well-being of city’s residents. Brack (2002) and Strohback et al.
(2012) find a strong influence of urban natural ecosystems in reducing air and noise
pollution and CO2 absorption in Australia and Germany. Heidt and Neef (2008) find
economic benefits by showing a significant increase in the nearby property values. Ac-
cess to nearby urban parks helps reduce stress and improve psychological well-being1
and increasing physical activities.2
Green urban areas are generally small, scattered and close to each other. As shown
in Table 1, on average, residents find a green urban area at less than about 350m from
their residences (about 6 minute walk) while more than 90% of green urban areas are
smaller than squares of about 300m length (i.e. surface smaller than 100,000 m2 ). Finally
the numbers of green urban areas with each city are roughly proportionate to the city
populations.
In this paper, we study the optimal level of small and scattered green urban areas
in urban spaces. Because green urban areas are land-intensive and offer very localized
amenities, we find that the relationship between the share of surface devoted to green
urban areas is a concave function of the distance to the city center. On the one hand, the
opportunity cost of land is too expensive in the city center for the planner to implement
many green areas there. On the other hand, residents are too sparse at the city edges
to give planners the incentives to invest in such very localized amenities. The main
contribution of the paper is to confirm our results using the very detailed geographical
1 See Berman et al. (2008), Hedblom et al. (2019), IFPRA (2013), Tzoulas et al. (2007), and van den Bosch
and Sang (2017)
2 See Cohen at al. (2007, 2016) and Evenson et al. (2013)

5
6 CHAPTER 2. GREEN URBAN AREAS

Table 2.1: Distance and Surface of Urban Green Areas (EU)

Small Cities Medium Cities Big Cities

Number of inhabitants (million) [0.1, .5) [0.5, 1) [1, ∞)
Closest distance between 2 urban green areas (m) 336 325 350
(std) (100) (62) (70)
Number of urban green areas within city border 96 384 784
(std) (74) (334) (513)
Average share of urban green areas with surface < 0.1 km2 94.1% 92.2% 90.6%
Note: Authors’ computation using GMES Urban Atlas database with the 305 largest European cities.
Adjacent green urban spaces are merged and counted for a single park. Distance between two parks is
the distance between the centroids of each park.

database of the European Environment Agency’s Urban Atlas data, which describes the
land use of 305 EU cities with more than 100,000 inhabitants. These data report land use
and cover across Europe using harmonized Earth Observations (EOs), which are com-
bined with Eurostat Urban Audit statistical data. The data represent a unique source
of reliable and comparable European urban planning data. As far as we know, this is
the first paper that uses both a theoretical model and empirical estimation of European
urban land use. This result is robust to many variations of the land use specifications,
city structure specifications, and city and country characteristics.

Our contribution relates to several strands of the economics literature. First, green
urban areas share the nature of a local public good. Since Tiebout (1956), economists
have discussed the issue of resident mobility and democratic decision over local pub-
lic goods (voting with feet). Fujita (1986), Cremer et al. (1986) and Sakashita (1987)
discuss the problem of the optimal location of local public goods and find that local
public goods should spread to equidistant locations. Fujita (1989) studies the optimal
allocation of ”neighbourhood goods” (localized and congestible public good) and their
decentralization through Pigouvian taxes. As such goods do not use land as input, the
balance between amenity and land use cannot be discussed. Berliant et al. (2006) endo-
genize the public good provision and location in cities where households have inelastic
land use. Optimal public good providers are found again to be equidistant and to serve
basins of residents of the same size. Yet, in contrast to this paper, those studies are con-
ducted under the assumption of no land use in the production of local public goods
and/or no endogenous choice of residential land plots. Because green urban areas are
rather land-intensive, it is important to study how land use affects public goods. Fur-
thermore, the distribution of residents is not uniform across urban landscapes and not
exogenous to the local amenities given by green urban areas. This paper focuses on the
relationships between the endogenous distribution of residents and green areas.
Finally, this paper links to the urban economics literature regarding the effect of
open spaces on urban form. Wu and Plantinga (2003) investigate the effect of an open
space on the surrounding urban structure. They, however, treat the location and size of
2.2. THEORETICAL MODEL 7

this open space as exogenous. Warziniack (2010) considers voting on the location of a
single open space when the geographical distribution of households is exogenous. Lee
and Fujita (1997) and Yang and Fujita (1999) examine the effect of a greenbelt, which
has an exogenous location. Yang and Fujita (1999) consider the effect of open spaces
at the neighborhood level and conclude that the equilibrium open space provision is
uniform across the distance to the city center. Such results contrast with our empirical
analysis that shows that the share of green urban areas is not constant across the city
space. Our model with endogenous locations and choices of residential space enables
an explanation of this pattern.
Parallel to our question is the issue of unoccupied urban spaces, which are often
seen as green areas. In contrast to green urban areas, open spaces are not maintained
for human activities. Unoccupied land has been primarily justified by the leapfrog-
ging effect. Capoza and Helsley (1990) and followers root this effect in the commitment
of building decisions and the resulting option value of urban land. Turner (2005) ex-
plains unoccupied land by the negative externalities of dwellings in their direct neigh-
borhoods. Walsh (2007) discusses and estimates the protection and regulation of open
spaces in Wake County (California, USA), which expands the discussion beyond mono-
centric city frameworks. Caruso et al. (2007) simulate market equilibria with discrete
house slots and a fixed housing consumption, which lead to open spaces. In contrast,
our paper discusses a continuous model where households decide their locations and
slots and where open green areas are costly and planned as in many EU cities. Ur-
ban green areas, such as parks or trees planted in rows, have maintenance and land
opportunity costs that are incurred by society.
The organization of this paper is as follows: Section 2 presents the theoretical model
and discusses the social optimal allocation of green urban areas and the decentraliza-
tion through the land market. Section 3 is devoted to our empirical approach and re-
sults. We first provide evidence on the concave shape of the share of green urban areas,
then estimate residents’ land choice, and finally quantify the economic benefits of green
urban areas. The last section concludes the paper.

2.2 Theoretical model

We consider a circular monocentric city hosting a central business district (CBD) and a
mass N of individuals. We denote by b ∈ R0+ the distance between the CBD and the
city border. The population density is defined as the number of individuals in a unit of
area at distance r from the CBD and is denoted by the function n : [0, b] → R+ , which
varies across the city.
In this model, we focus on small green urban areas that spread across the city and
are closely accessible to the local community around its location.3 Green urban areas
3 Because of proximity of small parks (see Table 2.1), residents’ travel costs to parks can be neglected
8 CHAPTER 2. GREEN URBAN AREAS

provide quick and frequent access to greenery, quiet, children’s parks, socialization
areas, etc. We consider the few blocks in the vicinity of a green urban area as our
unit of area or patch and model the urban area in a continuous fashion. In a unit of
area at distance r from the CBD, green urban areas offers a service x : [0, b] → [0, x ],
x ∈ R0+ , to the local community living in the vicinity. This service brings a level of
amenity a = αx (r ), although it necessitates the use of a fraction of land βx (r ) and
maintenance costs γx (r ). The parameters α, β, γ ∈ R+ distinguish the amenity, land use
and maintenance factors that affect green urban areas. Hence, the fraction of land used
for residential purposes is given by 1 − βx, and the maximum service level x is bounded
by 1/β. We assume absentee landlords, and the outside opportunity value of land is
given by the agricultural land rent R A ∈ R+ . For simplicity, we consider that rural areas
beyond the city border consist of private properties that do not provide green urban
area service for city dwellers (e.g., private crop fields, fenced areas, etc.). We denote
the land supply at distance r from the CBD by ` : [0, b] → R+ (e.g., ` = 2πr if the city
lies in a plain disk). In summary, land at distance r from the CBD includes a surface
βx (r )`(r ) of maintained green urban area and a residential area [1 − βx (r )] `(r ) , and it
hosts n(r )`(r ) residents who all benefit from the green urban area amenity αx (r ).
Individuals consume a quantity z of nonhousing composite goods and a quantity
s of residential space, while they benefit from the amenity a of a green urban area.
They are endowed with the utility function U (z, s, a), which is assumed to be concave
and increasing for each variable. We assume that demands for nonhousing composite
goods, residential space and amenity are gross substitutes such that U has negative sec-
ond derivatives and positive cross derivatives.4 As individuals are homogeneous, they
work and earn the same income w ∈ R+ in the CBD. Workers incur a total commuting
cost t : [0, b] → R+ with t(0) = 0 and dt/dr > 0. The price of composite good z is
normalized to 1 without loss of generality. From this point on and whenever there is no
confusion, we dispense the functions a, `, n, s, t, x, z and R with reference to distance r.
We first study the social optimal allocation and then the land market equilibrium.
here as a first approximation. This contrasts with the literature on the location of scarce public goods
where the travel to public facilities must be considered (see Cremer et al., 1986 and followers).
4 We do not formally disentangle the land space used for house and private garden. First, our data

on urban land cover does not permit to identify each separate item. Second, the empirical literature is
ambiguous about the substitution effect of private gardens on the use of parks (also called ”compensation
effect”). For instance, Talen (2010) and Caruso et al. (2018) find no significant relationship between the
ownership of a private garden and the frequency of park visits. Finally, note that it can be shown that our
theoretical results are unaltered if individuals have utility function U (z, S, a) over a house service function
S that is a homothetic consumption bundle S(s g , sh ) of land consumptions for house and private garden sh
and s g . In this case, private gardens substitute for parks but the elasticity of substitution between private
gardens and parks is the same as the one between houses and parks.
2.2. THEORETICAL MODEL 9

2.2.1 Social optimum

In an ideal world, green urban areas and residential structures should be combined
to balance their social benefits and costs. Analysis of the social optimal structure of
residential and green urban areas provides urban planners with viable directions for
urban planning. Towards this aim, we assume a benevolent social planner who controls
residential and green urban plots across the city.
As in Herbert and Steven (1960), we assume that the planner desires to set the same
utility target u ∈ R for all urban residents.5 She (the planner) minimizes the cost in the
city
Z b
C= (tn + zn + R A + γx ) `dr,
0
subject to the target constraint U (z, s, αx ) = u and land use constraint sn = 1 − βx.
The total city population results from the accumulation of population density across
Rb
the city: N = 0 n `dr. The planner chooses the profiles of consumption (z, s) and
spatial allocations (n, x ) as well as the border b. Since wages w are exogenous, this is
equivalent to the maximization of total surplus S = wN − C. After substitution of the
population and land use constraints, this provides
Z b
w−t−z

S= (1 − βx ) − R A − γx `dr. (2.1)
0 s
The planner then chooses the variables (z, s, x, b) that maximize S s.t. U ≥ u.
The optimal consumptions (z, s) are given by the pointwise maximization of (2.1),
which is equivalent to the set the maximum of the residential land value:
w−t−z
V ≡ max s.t. U (z, s, a) ≥ u. (2.2)
z,s s
Since the objective function in this expression decreases with z and U increases with
it, the constraint is binding. We denote the consumption e z(s, a, u) as the solution of
U (z, s, a) = u. Because the utility function increases for all variables, we obtain e zs =
−Us /Uz < 0, e z a = −Ua /Uz < 0 and e zu = 1/Uz > 0, while concavity of utility yields
zss > 0, where the subscripts denote partial derivatives. Denoting an individual’s net
e
income (net of commuting cost) as

y ≡ w − t,

the problem simplifies to

y−e z(s, a, u)
V = max .
s s
The optimal use of residential space is given by the solution of the following first-order
condition:
z(s, a, u) − se
e zs (s, a, u) = y.
5 Thisassumption avoids Mirrlees’s discussion of the unequal treatment of equals, has a close link to
competitive land equilibrium and yields first and second welfare theorems (Fujita 1989).
10 CHAPTER 2. GREEN URBAN AREAS

Because e z − se
zs > 0 and (∂/∂s) (e z − sezs ) = −se zss < 0, this condition determines the
optimal residential space b s(y, a, u). We denote the optimal consumption of commodity
goods by the function b z(y, a, u) = e s(y, a, u), a, u] and the optimal bid land value as
z [b
V (y, a, u). Ceteris paribus, bs increases with lower y, and since e zs < 0, b
z increases with
higher y. By the envelop theorem, the residential land value V rises with increasing y
and a and decreasing u.
The planner’s problem can then be rewritten as
Z b
max S = [V (y, αx, u) (1 − βx ) − R A − γx ] `dr. (2.3)
x,b 0

Pointwise differentiation w.r.t. x provides the necessary condition for green urban area
service,
αVa (1 − βx ) − βV − γ = 0, (2.4)

where V is evaluated at (y, αx, u). Using the land use constraint, we obtain the follow-
ing optimality condition:
αsnVa = βV + γ. (2.5)

This condition expresses the planner’s balance between the benefit of green area ameni-
ties (LHS) and the costs of green urban land and its maintenance (RHS). Let x ∗ (y, u) be
the optimal profile of the green urban area service. Note that x ∗ never reaches its upper
bound x = 1/β. If it did, the population density n would fall to zero, and green urban
areas would lead to maintenance and land costs but no amenities (zero LHS in (2.5)). It
can nevertheless be that x ∗ = 0 if the LHS is smaller than the RHS for all x ∈ [0, 1/β).
In summary, x ∗ ∈ [0, 1/β). For the sake of conciseness, we assume in this section that
the second-order condition holds and concentrate our discussion on interior solutions.6
Interestingly, expression (2.5) can be recovered as Samuelson’s optimality condition
of public goods after some mathematical transformations (see Appendix A):

Ua Us
α n=β +γ (2.6)
Uz Uz
where Ua /Uz is the marginal rate of substitution between commodities and amenities
and Us /Uz that between commodities and residential spaces. The Samuelson’s opti-
mality condition states that the sum of marginal rates of substitution for green area
amenities equates the maintenance cost γ plus the marginal rate of substitution for res-
idential land. This last element is novel in the context of public goods theory. In our
context, it applies at the level of the patch because externalities are localized at this
level. A green urban area is a local public good because its amenities equally benefit
the residents localized in its patch, which has a population density n. The same local
park indeed serves many residents. Green urban areas differ from usual (spaceless)
6 The
second-order condition is given by α2 Vaa (1 − βx ) − (1 + α) βVa < 0. This condition holds pro-
baa < 0.
vided that V
2.2. THEORETICAL MODEL 11

local public goods in their land intensity. This has an impact on locations with lower
usage of residential spaces because low space consumption is usually associated with
higher marginal rates of substitution for space Us /Uz . To our knowledge, this tradeoff
has not been highlighted in the literature.
Fixing the variables z, s and n, we observe that a lower amenity parameter α, higher
land use parameter β and higher maintenance cost parameter γ entice the planner to
reduce the green urban area service x and also its land area αx. We also distinguish
between the effects of population density and use of space. On the one hand, areas
with low population density should accommodate smaller shares of green urban areas
because they benefit fewer people. On the other hand, areas with small residential plots
imply high marginal rates of substitution for space and should also be provisioned with
smaller shares of green urban areas. In general, population density is low at the city
edges, and residential plots are small near CBDs. Hence, the planner is enticed to set
smaller shares of green urban areas at both the city edges and CBDs and while it implements
larger shares in intermediate locations, which is the property that we will explore in the
empirical section.

2.2.2 Comparative statics

Comparative statics on the service of green urban areas can be obtained by totally dif-
ferentiating (2.4). Noting that Va > 0, it is easy to see that

dx ∗ dx ∗
, < 0,
dβ dγ

such that land and maintenance costs have negative impacts on the service of green
urban areas to residents. Other comparative statics are ambiguous as

dx ∗
> 0 ⇐⇒ ((1 + αVaa ) (1 − βx ∗ ) − βVa ) x ∗ > 0,
dα
dx ∗
< 0 ⇐⇒ αVay (1 − βx ∗ ) − βVy < 0,
dy
dx ∗
< 0 ⇐⇒ αVau (1 − βx ∗ ) − βVu < 0.
du
First, the effect of the amenity parameter is ambiguous. It can be shown from the
first expression that when green urban area services provide very few amenities and
use small land pieces (α, β → 0), the optimal service x ∗ and surface βx ∗ rise with the
amenity parameter α. In fact, at very low levels, a higher α gives the planner incentives
to raise service x ∗ because the parameter raises the effectiveness of the service. How-
ever, at high levels, a higher parameter α substitutes for service level x and entices the
planner to reduce it.
Net income y is given by a worker’s wage minus his/her commuting cost, which
increases with distance to the CBD. Therefore, comparative statics on y highlight the
12 CHAPTER 2. GREEN URBAN AREAS

effect of distance to the CBD. The effect of net income can be deduced from the second
expression above as follows:

dx ∗
   
∗ d Ua ∗ Ua d lnbs d Us
> 0 ⇐⇒ αn − αn −β > 0. (2.7)
dy d ln y Uz Uz d ln y d ln y Uz

The effect of higher net income on green area services depends on three factors: first,
on the income elasticity of demand for residential spaces d lnb s/d ln y; second, on the
marginal rates of substitution between commodities and green urban areas Ua /Uz ; and
finally on the reaction of those marginal rates to increases in income (d/d ln y) (Ua /Uz )
and (d/d ln y) (Us /Uz ). The latter reactions are related to the Engel curves in spaces
(z, s) and (z, a). It can be shown that the marginal rate of substitution rises (falls) with
income in those spaces if the Engel curves rise and bend upward (downward). In other
words, higher income raises more demand for residential spaces s and amenities a than
demand for commodities z. As a result, the effect of net income on green urban area ser-
vices depends on the balance between the income effects on the demands for amenities
and space. If income effects are identical (as under the Cobb-Douglas preferences), the
marginal rates of substitution are invariant to income, and the above inequality holds
for all net incomes. The optimal green urban area service then rises with net income.
As a result, this optimal service also rises with wages and falls with distance from the
CBD. In the end, the relative importance of each income effect is still an empirical issue
for which we have found no information in the literature.
A similar comparative exercise can be performed on the impact of utility target u.
One simply substitutes y for u in condition (2.7). The effect of target utility therefore
depends on how it affects the use of residential space and the above marginal rates of
substitution.

2.2.3 City border and population

Finally, the planner sets the city border so that the first-order condition w.r.t. b,

V (y, αx ∗ , u) (1 − βx ∗ ) − γx ∗ = R A , (2.8)

holds, where y and x ∗ are evaluated at r = b. Since x ∗ maximizes the LHS, the latter
should be no smaller than V (y, 0, u). A sufficient condition for an optimal border b∗ (u)
is that the LHS lies above zero at r = 0 and decreases for increasing r. That is,

V (w, 0, u) ≥ R A , (2.9)
 
dt
−Vy + αVa (1 − βx ) < βV + γ for r ∈ [0, b∗ ], (2.10)
dr

where the second line is evaluated at (y, αx ∗ , u). We assume that these conditions hold.
Comparative statics can be obtained by totally differentiating (2.8). Since (2.8) decreases
with x, the optimal border b∗ increases with the parameters that increase the value of
2.2. THEORETICAL MODEL 13

the LHS of (2.8) and reduce its RHS. Recalling that Vy , Va > 0 > Vu and y = w − t, it
follows that
db∗ db∗ db∗ db∗ db∗ db∗
, , , <0< , . (2.11)
dβ dγ du dR A dα dw
Hence, As in the literature, cities also expand with higher wages and cities spread when
green urban areas provide higher amenities, use smaller pieces of land and require lower main-
tenance costs. They shrink with higher agricultural rent and utility level.
R b∗ (u)
The city population is given by N = 0 [1 − βx ∗ (y, u)] /b
s(y, αx ∗ , u) `dr, with
y = w − t. Finally, the total surplus is given by
Z b∗ (u)
S ∗ (u) = [V (y, αx ∗ , u) (1 − βx ∗ (y, u)) − R A − γx ∗ (y, u)] `dr. (2.12)
0
R b∗ (u)
Therefore, by the envelop theorem, the change in surplus is given by Su∗ (u) = 0 Vu
∗
(1 − βx ) `dr, which is negative since Vu < 0. Higher utility targets u reduce the city
surplus.

2.2.4 Land regulation

We are now equipped to discuss the impact of migration restriction and land regula-
tion in cities. In practice, utility targets are determined by city planners (city officials
and representatives) through their land regulation and migration policies. On the one
hand, a city planner may opt for unrestricted migration so that u is determined by the
outside utility level, e.g., u ∈ R. The population densities and levels adapt to migration
pressure, and the city generates a surplus S ∗ (u), which we assume to be positive (oth-
erwise, the planner has no incentive to create the city without external funding). On the
other hand, a city planner may opt to restrict land use and population as to maximize
incumbent residents’ utility. Then, he/she targets the highest possible utility subject
that is compatible with a positive surplus: S ∗ (u) ≥ 0. As the surplus decreases with
increasing u, the highest utility, say u∗ , is reached for a zero surplus: S ∗ (u∗ ) = 0. In
this case, green urban areas can be self-financed by land value. Indeed, S ∗ (u∗ ) = 0 can
be written as
Z b∗ (u∗ ) Z b∗ (u)
(1 − βx ∗ ) (V − R A ) `dr = x ∗ (γ + βR A ) `dr,
0 0

which shows an exact balance between the aggregate differential residential land value
(V − R A ) and the land and maintenance costs of green urban areas (the functions V
and x ∗ being evaluated at (y, αx ∗ , u∗ )). This balance is a reminiscence of the Henry
George theorem, by which a confiscatory tax on land would by itself finance a city’s
public goods, provided that the city reaches the size that maximizes residents’ utility
(see e.g. Arnott and Stiglitz, 1979). This paper adds two new elements to the standard
14 CHAPTER 2. GREEN URBAN AREAS

version of this theorem: optimal green urban areas, first, are very localized public goods that
provide unequal amenities and, second, require uneven land areas through the city.7
It must be emphasized that self-financing takes place at the city level and not at
the patch level. There indeed exist cross-subsidies across urban dwellers. Indeed, at any
distance r from the CBD, residents create a value V of their residential land, while they
generate a green urban area maintenance cost γx and a land opportunity cost R A on the
patch. This approach yields a resident’s net value equal to S ≡ V (1 − βx ) − R A − γx,
which is the integrand of (2.3). Using the first-order condition (2.4), one readily obtains
that dS/dr = Vy dy/dr = −Vy dt/dr < 0. Hence, a resident’s net value falls with
increasing distance from the CBD. When the city management maximizes residents’
R b∗
utility and imposes a zero surplus such that S ∗ = 0 S `dr = 0, it is clear that residents
close to the city center bring positive net values, while those away from it bring negative
values. Hence, the central population subsidizes the green urban areas at urban edges.

2.2.5 Competitive land market equilibrium

In most modern cities, residents freely choose their residential locations and spaces.
They make their decisions according to the land rent values signaled in the urban land
market. Given that urban green areas generate externalities to residents, it is important
to highlight the conditions under which the competitive land market replicates the so-
cial optimum discussed in the previous sub-sections. We here discuss the equilibrium
allocations in a competitive land market for an exogenous profile of green urban ser-
vices x : [0, b] → [0, 1/β) and amenities a = αx. A household’s budget constraint is
given by z + sR + t ≤ w, where R : [0, b] → R+ is the land rent function of distance to
the CBD.
In a competitive land market equilibrium, each land slot is awarded to the highest
bidder, and individuals have no incentives to relocate within and out of a city. There-
fore, they reach the same utility level ue , where the superscript e refers to the equilib-
rium value. Households bid up to (w − z − t)/s for each unit of residential space. Their
bid rent ψ : [0, b] → R+ is a function of distance r from the CBD such that
y−z
ψ = max s.t. U (z, s, a) ≥ ue , (2.13)
s,z s
where net income y = w − t is a function of distance to the CBD. As individuals com-
pete for land, they raise their bids to make their participation constraint binding and
obtain the equilibrium utility level ue . Note that (2.13) is equivalent to the social opti-
mal consumption choice (2.2). Therefore, households’ optimal consumptions are given
7 Arnott and Stiglitz (1979) show that the Henry George theorem holds in the presence of a ”green belt”

in a model where the latter is located at the city border, provides a global pure public good to all residents,
uses no land and requires no maintenance. In this paper green urban areas spread across the city, provide
localized public goods and are intensive in land and maintenance.
2.2. THEORETICAL MODEL 15

by the functions b s(y, a, ue ) and b

z(y, a, ue ), and the bid rent, by ψ b (y, a, ue ) = V (y, a, ue ).
The bid rent inherits the properties of V. That is, ψ b a > 0, while ψ
by, ψ b u < 0.
A competitive land market equilibrium is defined as the set of functions (z, s, R, n)
and scalars (b, N, ue ) satisfying the following four conditions. First, individuals choose
their optimal consumptions: z = b z(y, a, ue ) and s = b s(y, a, ue ). Second, land is allocated
to the highest bidder: R = max{ψ b (y, a, ue ), R A }, with R = ψ b (y, a, ue ) if n > 0, and
R = R A if n = 0. Third, the land market clears everywhere: nb s(y, a, ue ) = (1 − g) if
Rb
n > 0. Finally, the total population conforms to its density: N = 0 n 2πrdr. Here, N is
taken as exogenous in a closed city model, while ue is exogenous in an open city.
Within a city, equilibrium land rents are given by the winning bids such that R =
b (y, a, ue ). Since bid rents ψ increase with net income y and amenities a, the equilibrium
ψ
land rent R falls with distance from the CBD but rises with the proportion of green ur-
ban area. Importantly, at the social optimal amenity a and utility level u, consumptions in
a competitive land market match the social optimal ones exactly, while land bid rents ψ and
land rents R match social land values V. This finding is reminiscent of the social optimum
property in Herbert and Steven’s (1960) model, where competitive land market equi-
libria are socially optimal. The land market is then allowed to decentralize the choices
of land and commodity consumption. However, this applies only if green urban areas
are optimally set in our framework.
The equilibrium population density is equal to ne = (1 − g)/b s(y, a, ue ) ≥ 0, while
the equilibrium population aggregates the population density across the urban area as
Z be
1 − βx
Ne = 2πrdr.
0 s(y, a, ue )
b
The following proposition summarizes the above discussion.

Proposition 1 The land and commodity consumption in the competitive land market equilib-
rium and social optimum coincide if green urban areas are provided at the socially optimal levels
and if the planned and equilibrium utility levels u and ue match.

A competitive land market is a powerful mechanism to decentralize consumption

decisions. It can be checked that land tax does not affect goods and land consumption
choices; thus, land taxes may be used by city planners to finance green urban areas.

To obtain more analytical results, we focus on narrower classes of preferences.

2.2.6 Specific preferences

In this subsection we present three classes of preferences for which the optimal share of
green urban space is a concave function of the distance to city center and the optimal
population density is a decreasing function of it. For two classes, the optimal share of
green urban area can be bell-shaped.
16 CHAPTER 2. GREEN URBAN AREAS

2.2.6.1 Hyperbolic preferences

We here assume that individuals are endowed with the utility function U (z, s, a) =
z − θ/ (2s) + a where preferences for residential space are represented by an hyperbolic
function parametrized by θ.8 Accordingly, we get the consumptions e z = u − a + θ/ (2s),
s = θ/ (y + a − u), b
b z = (y − a + u) /2 and land value V = (y + a − u)2 /(2θ ) (see
mathematical details in Appendix A). We focus on the case of positive consumption so
that y + a − u > 0. Then, the marginal benefit of larger green urban service (LHS of 2.4)
is proportionate to
α β
(y + αx − u) (1 − βx ) − (y + αx − u)2 − γ.
θ 2θ
This expression is nil at an interior solution x ∗ > 0 and negative at the corner solution
x ∗ = 0. The optimal green urban area service x ∗ can be shown to be equal to the interior
solution s
α 2 6θγ
  
∗ 1 2 α 1
x+ = − y−u+ + y−u+ − ,
β 3α β 3α β β
p
if y > 6θγ/β − α/β + u and to the zero corner solution otherwise. It can readily be
shown that x+ ∗ is a concave function of y and reaches a maximum at the net income

ymax =
p
8θγ/β − α/β + u. Hence, if net incomes are larger and then respectively
lower respectively than ymax at the city center and border, the green urban area increases
and then respectively decreases with larger net incomes. Since y falls with distance to
city center, the share of green urban area increases and then decreases with distance
from the city center.
The optimal population density is given by
 s s
2
α 2 6θγ
 
∗ β  α α γ
n+ = y−u+ + y−u+ − 6θ y−u+ −
9θα β β β β β

for the interior solution. It is simply equal to (y − u) /θ at the corner solution. In both
cases it increases with net income y and therefore decreases with distance from city
center.
The reason for the concavity in green urban area density then lies in the balance
between opportunity cost of land and density of residents as shown in the Samuelson
condition (2.6). In the vicinity of the CBD, the social value of land is high because
of shorter commutes. This entices the planner to increase population density at the
expense of access to green urban areas. By contrast, the social value of land is low
at the vicinity of the city border because of longer commutes and entices the planner
to compensate individuals with larger residential plots and lower population density.
In those low density locations, green urban areas yield smaller social benefit and the
8 This yields a demand for residential space that has a price elasticity between one and zero (see Mossay

and Picard 2011).

2.2. THEORETICAL MODEL 17

planner reduces their surfaces. So, it is the high land social value that refrains the
planner to organize large green urban areas about the CBD and it is the low population
density that refrains her to maintain large green urban areas at the city fringe.

Proposition 2 Under hyperbolic preferences, population density falls with distance from CBD
while the share of green urban area is a concave function of distance from CBD. The share of
green urban area peaks at a location that moves away from CBD as cities are richer and less
regulated.

2.2.6.2 Cobb-Douglas preferences

We here assume the Cobb-Douglas utility U = z1−φ− ϕ sφ e aϕ with φ, ϕ, (1 − φ − ϕ) ∈

1
(0, 1). We compute z̃ = (us−φ e−aϕ ) 1−φ− ϕ , which gives
1−ϕ−φ
z= y, (2.14)
1−ϕ
b
 1− ϕφ−φ  1
1−ϕ

−(1 − ϕ − φ) − aϕ φ
s= uy e , (2.15)
1−ϕ−φ
b
 − φ1
V = κ −1 uy−(1 − ϕ) e−aϕ (2.16)

1− ϕ
− 1− ϕφ−φ
where κ = (1 − ϕ) φ (1 − ϕ − φ ) . Condition (2.6) becomes
αϕ
x − 1−φ ϕ 1
e φ (αϕ − βφ − βαϕx ) = κγy uφ. (2.17)

As shown in Appendix A, there exists a unique interior optimal service level x ∗ > 0
if the green area amenities per surface unit are sufficiently large α/β > φ/ϕ, the main-
tenance cost γ is sufficiently low and the net income y is sufficiently high. Otherwise,
there is a corner solution x ∗ = 0.
It is also shown that for any interior solution x ∗ ,

dx ∗ d2 x ∗
> 0 and < 0.
dy dy2

Hence, since y = w − t(r ), the optimal share of green urban areas g∗ = βx ∗ increases
with wage w and decreases with distance from the CBD r. The optimal share is also a
concave function of wage and distance from the CBD. Finally, since b s falls in y and a
∗
and because a = αx increases with y, b ∗
s(y, αx , u) also falls with the latter. Then, the
population density increases with higher wages and falls with longer distances to the
CBD.

Proposition 3 Under the Cobb-Douglas preferences, population density is a decreasing func-

tion of distance from the CBD, while the share of green urban areas is a decreasing and concave
function of this distance.
18 CHAPTER 2. GREEN URBAN AREAS

2.2.6.3 Stone-Geary preferences

In this section we extend the Cobb-Douglas preferences with a Stone-Geary utility com-
ponent for residential space such that U (z, s, a) = e ϕa z1−φ− ϕ (s − s0 )φ with φ, ϕ ∈ (0, 1)
and 0 < φ + ϕ < 1. In this case, s0 > 0 is a minimum housing consumption that raises
the marginal utility for residential space disproportionally more at low residential con-
sumptions. Under this specification, the elasticity of substitution between residential
space and either composite good z or amenity a falls as the individual is given a resi-
dential surface closer to s0 . As we will see bellow, the planner is then unable to increase
population density to high levels in city centers and devote much less land for green
urban areas there.
The composite good consumption that gives the utility level u is given by z̃ =
 − ϕa  1
ue (s − s0 )−φ 1−φ− ϕ . The residential land value V = max [y − z̃(u, a, s)] /s yields
s
the optimal residential space given by the condition (2.6)

−
 
− 1−1− ϕ 1 ϕ − 1 a
ϕ
( s − s0 ) φ− ϕ s − s0 = yu 1−φ− ϕ e 1−φ− ϕ
1−φ−ϕ

which is a transcendental equation that nevertheless accepts a unique solution ŝ > s0 .

It can be shown that dŝ/dy < 0, dŝ/da < 0 and dŝ/du > 0. Hence, since y falls with
larger commuting cost t and distance from center r, the residential land consumption
decreases with smaller distance to city center. However land consumption is bounded
z = z̃(u, a, ŝ) and using a = αx, the optimal green urban area
from below by s0 . Setting b
service x is given by

αϕ (1 − βx ) βφ z
− −γ = 0
b
z
1−φ−ϕ 1 − φ − ϕ ŝ − s0
b
ŝ

An interior maximum for x ∗ exists when the LHS of this expression falls to a zero value.
A corner solution x ∗ = 0 takes place when the LHS of this expression is negative for
all x ∈ [0, 1/β). The optimal green urban area service can only be found numerically.
Figure 1 displays an example of a concave profile of x ∗ : as one moves away from the
city center, x ∗ first increases and then decreases. Note that under this specification, x ∗
abruptly falls to zero at some distance from the city center. For higher opportunity
costs of land R A , this profile is truncated from the right so that the profile is concave
and increasing. Our numerical exploration found only concave profiles for sufficiently
high s0 . For sufficiently small s0 , the profile resembles to the decreasing and concave
functions as the ones found under Cobb-Douglass preferences.

Proposition 4 Under Stone-Geary preferences, the optimal residential space increases with
distance from CBD. Numerical exercises show that the share of green urban area is a concave
and bell-shaped function of distance from CBD for sufficiently high s0 .
2.3. EMPIRICAL ANALYSIS 19

Figure 1: Share of green urban space under Stone Geary preferences

The preferences reviewed in the last three subsections show that green urban space
profiles are generally concave function of distance from the center. The profile can
be bell-shaped if the elasticity of substitution between residential space and either the
composite good or amenity decreases with smaller residential space. The following sec-
tion shows that such theoretical results are not inconsistent with the data on European
cities.

2.3 Empirical analysis

In this section, we compare the model prediction to the actual green urban patterns in
European cities.

2.3.1 Data

In this paper, we use the dataset on urban land use from the Urban Atlas 2006, imple-
mented by the Global Monitoring for Environment and Security (GMES) service and
provided by the European Environment Agency (EEA), for the time period 2005-2007.
The dataset offers a high-resolution map of land use in European urban areas, contain-
ing information derived from Earth observations and backed by other reference data,
such as navigation data and topographic maps. The Urban Atlas uses Earth observa-
tion satellite images with 2.5 m spatial resolution.9 According to the GMES, the dataset
covers the functional urban areas (FUAs) of the EU cities with at least 100,000 inhabi-
9 GMES maps have a 100 times higher resolution than traditional maps in the Corine Land Cover in-
ventory produced since 1990.
20 CHAPTER 2. GREEN URBAN AREAS

tants.10 FUAs include land with both commuting distance and time below the critical
levels defined by Eurostat.11 The dataset includes all capital cities and covers nearly
300 of the most populous towns and cities in Europe (EU 27).12 Figure 2 displays the
urban areas covered by these cities.
The Urban Atlas provides a classification of city zones that allows for a compari-
son of the density of residential areas, commercial and industrial zones and extent of
green areas. In this paper, we use the data on ”green urban areas” (class 14100), which
are defined as artificial nonagricultural vegetated areas. They consist of areas with
planted vegetation that is regularly worked and/or strongly influenced by humans.
More precisely, first, green urban areas include public green areas used predominantly
for recreational use (gardens, zoos, parks, castle parks, cemeteries, etc.). Second, sub-
urban natural areas that have become and are managed as urban parks are included as
green urban areas. Finally, green urban areas also include forest and green areas that
extend from the surroundings into urban areas with at least two sides being bordered
by urban areas and structures and containing visible traces of recreational use. Impor-
tantly, for our study, green urban areas do not include private gardens within housing
areas, buildings within parks, such as castles or museums, patches of natural vegeta-
tion or agricultural areas enclosed by built-up areas without being managed as green
urban areas. It must be noted that green urban areas belong to the Urban Atlas’ class
of ”artificial surfaces”, which include all nonagricultural land devoted to human activi-
ties.13 This class is distinguished from the agricultural, seminatural areas and wetlands,
forest areas and water areas devoted to nonurban activities.
We select the (oldest) town hall locations as the CBDs. Then, we create a set of an-
nuli (rings) around each CBD at 100 m intervals. We define the ”annulus land area” as
the intersection of the annulus and the land within the urban zone area reported by the
GMES. This area includes artificial surfaces, agriculture, seminatural areas, wetlands
and forest but does not include water areas because those seas and oceans are not ap-
propriate for potential human dwellings. We then compute the share of green urban
area as the ratio of the surface of green urban area to the total land in the annulus land
area for each annulus . Figure 3 displays the annuli and the land use of green urban
areas (green color) for Dublin.

Whereas urban theoretical models usually assume a neat frontier between residen-
10 See the definition in the Urban Audit database and European Environmental Agency, GMES Urban
Atlas Metadata. Link: https://land.copernicus.eu/local/urban-atlas (accessed on Jan 25, 2018).
11 See the definition in the Urban Audit in EEA, 2015, and the details in Appendix B.
12 Austria, Belgium, France, Germany, Bulgaria, Cyprus, the Czech Republic, Denmark, Estonia, Finland,

Greece, Hungary, Ireland, Italy, Latvia, Lithuania, Luxembourg, Malta, Netherlands, Poland, Portugal,
Romania, Slovakia, Slovenia, Spain, Sweden, and the United Kingdom.
13 In addition to green urban areas, artificial surfaces include urban areas with dominant residential use,

innercity areas with central business district and residential use, industrial, commercial, public, military
and private units, transport units, mines, dump and construction sites, and sports and leisure facilities.
2.3. EMPIRICAL ANALYSIS 21

Figure 2: GMES 2006 maps

Figure 3: Dublin Land Use maps

22 CHAPTER 2. GREEN URBAN AREAS

Figure 4: Dublin Green Urban Space

tial and nonresidential spaces, urban data do not provide a clear separation between
residential locations and agricultural areas and forests. In this paper, we choose to fix
the city borders to the annulus for which the ratio of residential surfaces over the annu-
lus land area falls below 20%. Residential surfaces include urban areas with dominant
residential use and innercity areas with central business district and residential use.
They are shown in red in Figure 3. As shown in the sequel and Appendix B, the use
of other thresholds does not lead to qualitative differences in our empirical results. We
define the distance from the CBD, dist, as the distance from the CBD of the annulus
and the relative distance, rdist, as this distance from the CBD divided by the distance
between the CBD and the city border. Figure 4 plots the share of green urban areas in
each annulus of Dublin city as a function of distance. It also includes the quadratic and
cubic fit function. The non-monotone relationship is quite apparent. Figure 5 presents
a box-plot of those shares as a function of the relative distance for all 305 cities. The
non-monotone is also apparent although there is now a lot of heterogeneity related to
country and city specificities.

We include several controls that do not depend on the relative distance to the CBD.
A country dummy vector accounts for a country’s specific urban regulations and wealth.
We also divide the sample into three city groups: small cities, with a population below
a half million; medium-sized cities, with a population between a half and one million;
and large cities, with a population exceeding one million. The dummy vector city size
includes the fixed effect on each city group.
2.3. EMPIRICAL ANALYSIS 23

Figure 5: Green Urban Space in 305 EU Cities

In addition to the GMES, we use the population density from the European pop-
ulation grid.14 We calculate the population mass at each distance to the city center
and redistribute the population to the residential area in each annulus.15 Because the
Eurostat population grid does not cover Cyprus, we exclude Lefkosia, Cyrus. Other
measures of cities’ exogenous geographical characteristics are taken from the E-OBS
database.16 We control for these exogenous geographical characteristics because they
may affect residential choices. We finally measure the city populations as the number
(millions) of inhabitants living in the city and greater city (CGC) areas, as defined and
reported in the Eurostat databases.17 Table 2.2 presents the summary statistics.

On average, green urban areas account for nearly 6% of the total surface of city
areas. Cities have a rather heterogeneous share of green urban areas. Figure 6 displays
a histogram of the shares of green urban areas across the studied cities. There is a
large dispersion in the average green urban areas across EU cities. In Figure 6, the
city with lowest share of green urban area (0.62%) is Limerick, Ireland, and the one
with the highest share (42.6%) is Karlovy Vary, Czech Republic. The latter is a spa
resort city, which offers many green areas to its visitors. The former city includes few
land surfaces classified as green urban areas because it also has many agricultural and
14 For more information on the European population grid, please check the technical report of the GEO-
STAT 1A project from Eurostat.
15 Residential areas are called ’urban fabrics’ in the GMES.
16 The E-OBS database provides detailed data on the daily temperature, daily precipitation, sea level

pressure and elevation across Europe. We acknowledge the E-OBS dataset from the EU-FP6 project
ENSEMBLES (http://ensembles-eu.metoffice.com) and the data providers in the ECA&D project
(http://www.ecad.eu).
17 For more details, see metadata files for urb esms in the Urban Audit database of the Eurostat website.
24 CHAPTER 2. GREEN URBAN AREAS

Table 2.2: Summary statistics

average sd min max observations

City border (km) 4.3 3.2 1.0 24.0 305
City area (km2 ) 84 174 1 1809 305
Number of annuli 43 32 10 240 305
Population in FUA (millions) 0.79 1.29 0.06 12.10 301
Population in CGC (millions): all cities 0.44 0.79 0.03 8.17 305
small cities 0.20 0.11 0.03 0.50 240
medium sized cities 0.64 0.13 0.51 0.98 36
large cities 2.25 1.66 1.01 8.17 29
Total share of green urban area in city (UGS) (%) 6.5 4.3 0.6 42.6 305
Highest share of UGS (%) 19.2 8.8 2.4 70.0 305
Highest share of UGS (%) (kernel smoothing) 9.5 5.0 0.4 42.2 305
Distance of highest share of UGS (km) 1.3 1.5 0.1 15.8 305
Distance of highest share of UGS (km) (kernel smoothing) 1.7 1.8 0.1 15.5 305
GDP per capita (e1000/hab.) 26.88 13.06 6.00 83.70 305
Household income (e1000/hab.) 15.46 5.63 3.70 30.90 304
City Geographical Controls
Elevation (m) 212 210 −2 1,614 305
Average temperature at Jan 01 (o C) 2.30 4.62 −8.48 15.57 305
Average temperature at July 01 (o C) 19.18 12.27 12.27 28.70 305
Average daily precipitation (mm/day) 1.91 0.60 0.48 4.45 305
Share of Urban Green Land (%) 6.57 4.38 0.62 42.69 304
Share of Residential Land (%) 34.03 4.99 12.85 47.42 304
Note: The GMES database released on May 2010 reports only 301 FUAs for the time period 2005-2007. Cities
without FUAs reporting are Wrexham and Derry (UK), and Gozo (Malta). Aix-en-Provence shares the same
FUA as Marseille. We use the Nadaraya-Watson Gaussian Kernel to smooth variations of annuli values.18
GDP per inhabitants and Household income are taken from Regional Economic Accounts from Eurostat at
NUTS3 and NUTS2 level respectively. Note that in Eurostat database, household income level exists only
at NUTS2 level. In eurostat database for household income at NUTS2, there is no data for Luxembourg
(NUTS2 code LU00); therefore, there is only 304 cities instead of 305 cities. The number of inhabitants in
each annuli is calculated based on Eurostat Population Grid. As Eurostat Population Grid 2006 does not
cover Cyrus; hence, we also drop the city cy001l lefkosia in our database. The total number of annuli are
calculated for 303 cities excluding lu001l luxembourg and cy001l lefkosia. For city geographical controls,
we take into account the average for period 1995-2010 for each city.

seminatural lands that can be used for urban green amenities. These outliers do not
affect our results. Spots with the highest densities of urban green areas are located,
on average, at 1.3 − 1.7 kilometers from the CBD. City elevation also varies greatly,
from two meters below sea level in Amsterdam, Netherlands, to 1, 614 meters above
sea level in the mountainous city of Innsbruck, Austria. European cities belong to a
mild climate zone, with temperatures varying between −8 and +28 degrees Celsius at
the lowest and highest day temperature in winter and summer (measured on January
1 and July 1, respectively, for the period 1995-2000).19 The average population density
is approximately 4,400 inhabitants per square kilometer and ranges from 1,000 to 9,800
inhabitants per square kilometer. These numbers are reasonable because the database
concentrates on the core of urban areas with no agricultural or seminatural land use.
19 We use observations from the E-OBS databases from the EU-FP6 project (for details, see the references).

Our samples do not contain some northern European cities in Iceland and Norway.
2.3. EMPIRICAL ANALYSIS 25

Figure 6: Average Share of Green Urban Space Distribution

2.3.2 Urban green area profiles

In this subsection, we compare our theoretical predictions with the empirical proper-
ties of green urban areas in EU cities. According to the theoretical model, the social
optimal land share devoted to green urban areas is a concave and possibly nonmono-
tonic function of distance to the CBD; it first increases and then decreases as one moves
away from the CBD. This pattern reflects the tradeoff between the high land values in
the center, which make green urban areas too costly, and the too sparse population in
the periphery, which associates green urban areas with too low social benefits. The aim
of this subsection is to test the concavity of the land share of green areas in the studied
European cities. In addition, we test whether this share is nil at the CBD. We propose
the following reduced form model:

gijc = η1 rdistijc + η2 rdist2ijc + η3 rdist3ijc + X j + eijc , (2.18)

where gijc is the land share of green urban areas in annulus i (ranked according to dis-
tance to the CBD) of city j in country c. We study both quadratic and cubic models,
where the coefficient η3 is constrained to zero in the first case. To allow for comparison
across cities, we define the covariate of the relative distance of an annulus rdistijc and
add the city fixed effects X j as controls. Urban green areas are likely to lie on adja-
cent annuli such that green urban area densities may not be independent observations,
which biases the estimation. Additionally, more distant annuli aggregate more surface;
thus, the estimation may suffer from heteroskedasticity. Hence, we report the Huber-
26 CHAPTER 2. GREEN URBAN AREAS

White heteroskedastic-consistent estimation of the standard residual errors. Table 2.3

presents the results from the regression of equation (2.18).

Table 2.3 reports a negative and significant correlation with the square of the dis-
tance to the CBD, which suggests that the hypothesis of a concave profile for the share
of green urban areas g should not be rejected. Columns (1) to (3) report the results with
and without country and city controls. As can be seen, these controls do not have a
large effect on the amplitude and significance of results. Columns (4) to (6) present the
results for the subsamples of small cities (population below a half million), medium-
sized cities (population between a half and one million) and large cities (population
exceeding one million), respectively. The signs are not altered, which corroborates the
idea of concave profiles. A formal testing of concavity requires an examination of the
p-value of a one-sided t-test of the respective coefficients. In Appendix B, we show
that the p-values are very low; thus, the joint hypothesis of η1 > 0 and η2 < 0 cannot
be rejected at the 99% confidence. We also show that the results are robust to various
observation weightings, which suggests that misspecification issues can be excluded.
Columns (4) to (6) also show that the (absolute values of) amplitudes of the coef-
ficients increase with city size. The shares of green urban areas reach higher levels
in larger cities. Indeed, we can compute the average shares of green urban areas in
the CBD (rdist = 0), at its peak location (rdist = −η1 / (2η2 )) and at the city border
(rdist = 1). The average shares of green urban areas in the CBD are given by the in-
tercepts of the regression models, which are computed as the following averages of the
city fixed effects in Columns (4) to (6): 0.067∗∗∗ , 0.059∗∗∗ , and 0.058∗∗∗ . With this in-
formation, we deduce that, on average, the shares of green urban areas in small cities
(Column (4)) rise from 6.7%, peak at 7.7% at rdist = 33% of the border and fall back to
3.6% at the city border. The shares in medium-sized cities (Column (5)) rise from 5.9%,
peak at 9.0% at 46% of the border and fall back to 4.6%. Finally, large cities (Column (6))
have shares that rise from 5.8%, reach their peaks at 9.9% at 45% of the border and fall
back to 3.7%. Finally, since the share of green urban areas over an entire city is given
R1
by 0 η0 + η1 ξ + η2 ξ 2 dξ = η0 + 21 η1 + 13 η2 , we can compute the average shares to be

6.7%, 7.7% and 8.1% for small, medium-sized and large cities, respectively.
Our model predicts a concave profile for the share of green urban areas rather than
the quadratic profile in Columns (1) to (6). Columns (7) to (9) present the same re-
gression analysis as in Columns (1) to (3) for a cubic regression model. One can ob-
serve that the coefficients are significantly different from zero. The results suggest a
hump-shaped profile for the share of green urban areas, with the coefficients for the
square of distance being significantly negative. In this cubic model, local concavity is
given by the sign of the second derivative g00 = 2η2 + 6η3 × rdist, with rdist ∈ [0, 1].
Since η2 is negative in Table 2.3, the share of green urban areas is certainly a con-
cave function in areas sufficiently close to the CBD. However, because η3 is positive
 Table 2.3: Profile of green urban spaces (EU27): baseline results

Dependent variable: Share of Urban Green Space

Quadratic Cubic
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12)
Distance 0.102∗∗∗ 0.102∗∗∗ 0.102∗∗∗ 0.062∗∗∗ 0.134∗∗∗ 0.181∗∗∗ 0.297∗∗∗ 0.292∗∗∗ 0.294∗∗∗ 0.232∗∗∗ 0.299∗∗∗ 0.424∗∗∗
2.3. EMPIRICAL ANALYSIS

(0.016) (0.016) (0.016) (0.018) (0.040) (0.037) (0.038) (0.038) (0.039) (0.047) (0.082) (0.089)
Distance square −0.127∗∗∗ −0.127∗∗∗ −0.127∗∗∗ −0.093∗∗∗ −0.147∗∗∗ −0.202∗∗∗ −0.598∗∗∗ −0.588∗∗∗ −0.591∗∗∗ −0.499∗∗∗ −0.550∗∗∗ −0.802∗∗∗
(0.014) (0.014) (0.014) (0.015) (0.035) (0.034) (0.080) (0.079) (0.080) (0.097) (0.174) (0.196)
Distance cubic 0.307∗∗∗ 0.300∗∗∗ 0.302∗∗∗ 0.263∗∗∗ 0.264∗∗ 0.396∗∗∗
(0.048) (0.047) (0.048) (0.058) (0.108) (0.120)
Country FE X X
City FE X X X X X X X X
Sample of cities All All All Small Medium Large All All All Small Medium Large
Observations 13,091 13,091 13,091 7,982 2,106 3,003 13,091 13,091 13,091 7,982 2,106 3,003
Adj. R2 0.046 0.148 0.369 0.367 0.338 0.383 0.056 0.158 0.379 0.374 0.346 0.402
df 13,088 13,063 12,785 7,741 2,068 2,972 13,087 13,062 12,784 7,740 2,067 2,971
F Stat. 79.63∗∗∗ 20.50∗∗∗ 26.13∗∗∗ 20.32∗∗∗ 30.05∗∗∗ 63.00∗∗∗ 54.35∗∗∗ 19.60∗∗∗ 27.13∗∗∗ 20.82∗∗∗ 30.37∗∗∗ 66.07∗∗∗
Note: Significance levels are denoted by ∗ for p<0.1, ∗∗ for p<0.05 and ∗∗∗ for p<0.01. Standard errors are clustered at city level and reported in parentheses. The row ”df” reports
the degree of freedom. The table shows results from the regression of share of green urban space within each annulus on the relative distance and its square to the city center
using control variables and a constant. For comparison purpose, the relative distance of an annulus is given by its distance from city center divided by the distance of the farthest
annulus in the city. Columns (2) to (3) sequentially includes controls on country and city fixed effects. Columns (4) to (6) restrict the sub-samples to small cities (population lower
than half million), medium size cities (population lower between half and one million) and large cities population larger than one million). Same strategies are applied to cubic
regressions (7) to (12).
27
28 CHAPTER 2. GREEN URBAN AREAS

in Table 2.3, concavity fails at greater distances from the CBD. Using the results in Col-
umn (9), the convexity coefficient is equal to g00 (0) = 2η2 = −1. 182 in the CBD and
g00 (1) = 2η2 + 6η3 = 0.630 at the city edge. Hence, the profiles are increasing and
concave near the CBD but convex at city edges. However, such convexity is not incon-
sistent with our theory because the share of green urban areas must have convex kinks
when reaching zero. Finally, using Column (9), the share of green urban areas reaches
a maximal value at 33% of the distance between the CBD and city border (solution of
g0 = η1 + 2η2 × rdist + 3η3 × rdist2 = 0).
We run a series of robustness checks, and the results are presented in Table 2.4. We
study variations around the setup of Column (3) in Table 3, with both city and coun-
try fixed effects. First, as Figure 4 shows, there exists substantial serial correlation in
the share of green urban areas, which questions the assumption of homoskedatisticity.
Column (1) presents similar results under OLS without correction for heteroskedas-
ticity and therefore suggests that heteroskedasiticy is not an important issue. Second,
we check issues of truncation and size observation units. The surface areas of the an-
nuli rise linearly with distance to their center. Therefore, the annuli about a CBD mea-
sure green urban areas within smaller surfaces and may have much more variability, as
would be the case if an identically sized park were randomly dropped on the annuli.
Furthermore, small-surface annuli are supposed to include the true city center but may
miss this objective if true city centers are slightly away from the city hall locations used
as the city centers. The observed share of green urban areas close to CBDs can then be
more volatile and biased. To check for this issue, we aggregate the three most central
100-meter annuli into one larger central ring and the next two annuli in another ring,
while we leave the other annuli intact. The results are presented in Column (2) and
do not qualitatively differ from the baseline model. In Column (3), we also include the
robustness check for urban green profiles, where all large parks are excluded. Since our
model focuses on the effect of local urban green areas within a neighborhood, it might
not apply to very large parks that have global effects on city inhabitants. Therefore,
we further exclude all the parks exceeding one square kilometer,20 and the regression
results do not significantly different from our baseline regression.
Third, in Table 2.3, city borders have been defined by the locations where the share
of residential space reaches a threshold of 20%. This definition resulted from our trade-
off between theory and data. In theory, a city border is well defined and has a zero
residential density, although it is not well defined in the data where the residential den-
sity never reaches zero. The use of a threshold that is too high certainly undershoots the
actual distances between the CBD and city borders. Therefore, we extend the definition
of a city border with a lower threshold of 15%. This extension is shown in Column
(4), where the number of observations rises to 16,851 annuli. The results remain qual-
20 We also checked other criteria for this size threshold for large parks in our model, such as the 99th
percentile level; however, the results were rather similar.
2.3. EMPIRICAL ANALYSIS 29

itatively the same. However, for this threshold value, the random variations in the
share of residential spaces lead to fluctuations and downward biases in city border val-
ues. Column (5) displays the results obtained when the share of residential spaces is
smoothed (with the same kernel smoothing as in Table 2.2). The number of observed
annuli rises to 22,549. The results remain qualitatively the same, except for the coeffi-
cient of the linear distance term, which becomes not significant. This result reflects a
decrease in the slope of the measured share of green urban areas and is explained by
the fact that Column (5) includes new observations with no green urban areas. These
new observations with zeros at far distance from the CBD reduce the slope of the share
of green urban areas.
Fourth, the dataset may not match the monocentric city hypothesis of the theoretical
model because it includes polycentric cities and contiguous cities. The next columns
of Table 2.4 reduce this mismatch. Column (6) reports results for the set of 27 small
and medium-sized EU cities with populations less than 1 million individuals (keeping
fixed country effects). This approach eliminates the largest cities that are prone to host
multiple subcenters. Column (7) focuses on monocentric cities using the OECD study
on metropolitan urban polycentricity, keeps the 21 countries that are common to our
GMES database and excludes reported polycentric cities (see Appendix B for details).
Similarly, in our theory, cities are spatially separated, which is not always the case in the
data. Columns (8) and (9) report results with the subsamples of cities that are at least
50 km and 100 km apart, respectively, keeping fixed country effects. The first distance
usually corresponds to the extent of urban labor market areas. The second distance
makes sure that daily commuting between cities is unattractive. As can be seen in
Table 2.4, Columns (6) to (9) do not qualitatively deviate from the baseline results.
One may question to which land functionality green urban areas should be com-
pared. Our theoretical model discusses the split between green urban areas and resi-
dential land. In the above empirical model, we have extended the areas for residential
functionality to all human dwellers’ functionalities. Accordingly, our above baseline
empirical analysis used a measure of the share of green urban areas consisting of the
ratio of the area of green urban areas as the numerator and the area of artificial surfaces,
agriculture, seminatural areas, wetlands and forest as the denominator. Therefore, the
denominator includes many potential land functions. The last columns of Table 2.4
present the results on alternative measures for this denominator, which increasingly
narrow the comparison down to residential areas. Column (10) displays the share of
green urban areas when we keep only the artificial surfaces in this denominator. This
approach eliminates agricultural areas, wetlands and forests. Next, we compare green
urban areas to the land used exclusively for human activities. Column (11) reports the
results with the denominator measuring the land for residences, offices and green urban
areas (i.e., urban fabrics, industrial, commercial, public, military and private units and
green urban areas). This approach eliminates roads, railways, ports, airports, mines,
CHAPTER 2. GREEN URBAN AREAS

Table 2.4: Profile of green urban spaces: robustness

Dependent variable: Share of Urban Green Space
OLS Wider central Exclude Border 15% Monocentric cities Contiguity Other measures Medieval
annulli big parks Raw Smooth’d <1Mo OECD 50km 100km cities
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10 (11) (12) (13)
Distance 0.102∗∗∗ 0.068∗∗∗ 0.092∗∗∗ 0.065∗∗∗ 0.005 0.078∗∗∗ 0.102∗∗∗ 0.080∗∗∗ 0.055 0.108∗∗∗ 0.087∗∗∗ 0.126∗∗∗ 0.112∗∗∗
(0.006) (0.019) (0.015) (0.015) (0.014) (0.017) (0.029) (0.021) (0.043) (0.020) (0.027) (0.030) (0.016)
Distance square −0.127∗∗∗ −0.100∗∗∗ −0.120∗∗∗ −0.116∗∗∗ −0.078∗∗∗ −0.105∗∗∗ −0.128∗∗∗ −0.112∗∗∗ −0.089∗∗ −0.105∗∗∗ −0.094∗∗∗ −0.111∗∗∗ -0.129∗∗∗
(0.005) (0.016) (0.013) (0.014) (0.013) (0.014) (0.025) (0.018) (0.036) (0.018) (0.024) (0.027) (0.014)
Med.*Dist sqr 0.002
(0.008)
City FE X X X X X X X X X X X X X
Sample of cities All All All All All Small+med. EU21 50km 100km All All All All
Observations 13,091 12,179 13,091 16,819 22,507 10,088 5,094 8,634 2,826 13,091 13,091 13,091 13,045
Adj. R2 0.369 0.435 0.349 0.399 0.442 0.363 0.300 0.387 0.301 0.340 0.306 0.326 0.368
df 12,785 11,873 12,785 16,513 22,201 9,811 5,011 8,425 2,758 12,785 12,785 12,785 12,740
F Stat. 26.13∗∗∗ 31.73∗∗∗ 22.42∗∗∗ 37.65∗∗∗ 59.54∗∗∗ 21.79∗∗∗ 27.58∗∗∗ 27.23∗∗∗ 19.18∗∗∗ 23.10∗∗∗ 19.89∗∗∗ 21.75∗∗∗ 26.03∗∗∗
Note: Significance levels are denoted by ∗ for p<0.1, ∗∗ for p<0.05 and ∗∗∗ for p<0.01. Standard errors are clustered at city level and reported in parentheses. The row ”df” reports
the degree of freedom. The table shows results from the regression of share of green urban space within each annulus on the relative distance and its square to the city center using
control variables and a constant. The relative distance is normalized to one for farthest annulus for comparison purpose. Columns (1) includes simple OLS without corrections;
other columns are controlled for heteroskedasticity. Column (2) merges the central annuli in wider areas. Column (3) excludes all big parks with the size above 1 kilometre square.
Columns (4) and (5) define city borders with a lower threshold (15%) on the un-smoothed and smoothed profile of residential density. Columns (6) and (7) isolate monocentric
cities respectively by using cities with less than one million inhabitants and by excluding the polycentric cities reported by OECD. Columns (8) and (9) exclude contiguous cities
with inter-city distances lower than respectively 50 and 100 km. Column (10) measures the share of green urban space with respect to only artificial surfaces. Column (11) further
excludes transport infrastructure, etc. Column (12) measures it with respect to only urban fabrics.# Colunm (13) includes a row with the effect of cities with medieval history on
the square of distance. A city has a medieval history if it existed before year 1500 (Bairoch et al. 1988).#
30
2.4. CONCLUSION 31

construction sites, land with no use and sports and leisure facilities. Finally, Column
(12) is even more restrictive by concentrating on only urban fabrics and green urban
areas. The regression coefficients remain stable despite important variations in the def-
inition of the share of green urban areas.
We also check the impact of the medieval history of each city on the share of green
urban areas. One may argue that medieval cities included walls, moats and hunting
places that have been preferably converted to parks by planners when the cities ex-
panded in the 19th and 20th centuries. Accordingly green urban areas would be placed
in annuli between the center and periphery of existing cities, which would explain the
non-monotone relationship between the share of green urban space and distance to city
center. To test this hypothesis, Column (13) shows the same regression as in the baseline
regression (Table 2.3 Column (3)) with an additional cross-effect of a medieval history
dummy on the square of the distance to center. Cities are defined as medieval if they
are built before 1500. As it can be seen, this cross-effect is not statistically significant,
indicating no difference between the groups of cities with medieval and non-medieval
history.
Finally, we also run the regression (2.18) for each city in our sample and count
the number of cities for which the concavity property holds. We observe only 4.59%
(10.49%) of cities where we cannot reject convexity (η2 > 0) for p-value < 0.01 (p-value
< 0.1).

2.4 Conclusion
In this paper, we discuss the patterns of urban green areas in cities from theoretical
and empirical perspectives. Urban green areas mainly include green areas maintained
for recreational purposes by non-private human institutions (typically, municipalities).
Green urban areas provide residents with amenities that have the property of local pub-
lic goods and high land intensity. We find that the optimal provision of urban green
areas is a nonmonotonic concave function of the distance to CBDs. It results from the
balance between the higher opportunity cost of land near CBDs and the lower popula-
tion density at city edges.
This property is confirmed by our study of the urban land use in the 305 most pop-
ulous urban EU areas. We use detailed maps of urban land use from the European
Environment Agency (GMES) to study the spatial configurations of urban green areas.
Our study shows a concave and hump-shaped profile of urban green areas with re-
spect to distance to the CBD. The result is robust to many variations in the land use
specifications, city structure specifications, and city and country characteristics.
32 APPENDIX

Appendix A: Proofs
Samuelson Rule We can transform expression (2.5) in the Samuelson’s opti-
mality condition of public goods. Plugging the optimal residential space condition
z − se
e zs = y in the land value V = (y − ez) /s gives V = −e zs . At the same time, applying
the envelop theorem on V yields Va ≡ −e s > 0. Finally, using e
z a /b zs = −Us /Uz and
z a = −Ua /Uz , expression (2.5) can be written as (2.6).
e

Hyperbolic preferences We here assume that individuals are endowed with the
utility function
U (z, s, a) = z − θ/ (2s) + a
where preferences for residential space are represented by an hyperbolic function parametrized
by θ.21 Accordingly, we get the consumptions e z = u − a + θ/ (2s), b
s = θ/ (y + a − u),
2
z = (y − a + u) /2 and land value V = (y + a − u) /(2θ ). We focus on the case of
b
positive consumption so that y + a − u > 0. Then, condition for optimal green urban
area service (2.4) simplifies to the identity
α β
(y + αx − u) (1 − βx ) − (y + αx − u)2 − γ = 0, (2.19)
θ 2θ
where the LHS expresses the net benefit of green urban service and is a concave quadratic
function of x. If the inequality
s
α 6θγ
y−u+ > (2.20)
β β

does not hold, the LHS is negative so that there are no benefits from green urban areas
for any net revenue y. If it holds, the identity accepts two solutions, but only the highest
one, x = x+∗ , determines an interior maximum. Hence, the optimal interior solution for

land use of green urban area is given by

  s 2
∗ ∗ 2β α β α 6θγ
g+ ≡ βx+ = 1− y−u+ + y−u+ − ,
3α β 3α β β
Using the previous relationships, we can also infer the optimal interior solution for
population density by
 s s
2
α 2 6θγ
 
∗ β  α α γ
n+ = y−u+ + y−u+ − 6θ y−u+ − .
9θα β β β β β

If (2.20) holds at the same time as g+ ∗ > 0 and n∗ > 0, then the allocation ( g∗ , n∗ ) =
+
∗ , n∗ ) is a socially optimal one. Otherwise, the allocation ( g∗ , n∗ ) = (0, max{0, ( y − u ) /θ })
( g+ +
is socially optimal.
21 This yields a demand for residential space that has a price elasticity between one and zero (see Mossay

and Picard 2011).

APPENDIX 33

The objective of this paper is to understand the profile of the share of green urban
area. We successively get the following comparative statics result:
 
∗
dg+ β  1 
= r − 2
,
dy 3α  1 −  6θγ
2
β y−u+ αβ
∗
d2 g+ 2γθ
=− < 0.
d2 y
r 2
3/2 6θγ
α y−u+ α
β − β

The share of green urban area g+ ∗ is therefore a concave function of y − u, and therefore,

since y = w − t and dt/dr > 0, it is a concave function of wage w and distance to CBD,
r. Stronger land regulation reflected by higher u also have non monotone effect on green
urban area. By contrast, it can readily be observed that the population density n∗+ rises
with y so that it increases with wage w but falls with distance to the CBD r. The planner
compensates longer commuting distances with larger residential space. The reason for
the concavity in green urban area density then lies in the balance between opportunity
cost of land and density of residents as shown in the Samuelson condition (2.6). In
the vicinity of the CBD, the social value of land is high because of shorter commutes.
This entices the planner to increase population density at the expense of access to green
urban areas. By contrast, the social value of land is low at the vicinity of the city border
because of longer commutes and entices the planner to compensate individuals with
larger residential plots and lower population density. In those low density locations,
green urban areas yield smaller social benefit and the planner reduces their surfaces.
So, it is the high land social value that refrains the planner to organize large green urban
areas about the CBD and it is the low population density that refrains her to maintain
large green urban areas at the city fringe.
It can be further checked that the share of green urban area reaches the maximum
∗ /dy = 0). Hence, the
p p
value 1 − 2θγβ/ (9α2 ) for y − u + α/β = 8θγ/β (where dg+
share of green urban area reaches a peak that rises with higher amenity α, lower green
urban area land use β, lower maintenance cost γ and lower preferences for residential
space θ. All these properties are intuitive. The maximum value is however independent
of wage w and regulation policies through changes in u. To know the peak location, we
p
can use y = w − t and write t(r ) = w − u + α/β − 8θγ/β. Since dt/dr > 0, we
infer that the peak location shifts away from the CBD with higher amenity α, lower
maintenance cost γ and weaker preference for residential space θ. It also shift away
from CBD with higher wages w and weaker land regulation (smaller u).
To fix ideas, consider cities in the same country where labour mobility is high, ur-
ban land regulations are rather similar and residents’ preferences are the same. Hence,
cities differ only with respect to per capita incomes w, which are driven by first nature
and second nature advantages, such as the presence of specific factors, harbour, steel in-
34 APPENDIX

dustry, financial center, etcetera. Then, the location of peak in the share in green urban
area increase with earnings and therefore city population sizes N. This reflects the fact
that land rent becomes more expensive about the CBD and entices the planner to shift
them away from it. Under these preferences, the amplitude of this peak is unrelated to
earnings and city sizes. We will come back to this relationship in the empirical analysis.

Proposition 5 Under hyperbolic preferences, population density falls with distance from CBD
while the share of green urban area is a concave function of distance from CBD. The share of
green urban area peaks at a location that moves away from CBD as cities are richer and less
regulated.

One can also check the following comparative statics on population density:

dn∗+ dn∗+ dn∗+ dn∗+ dn∗+

, , <0< , (2.21)
dγ dθ du dα dy

(see Appendix A). As standard in the urban economic literature, the population density
falls with stronger preferences for residential space θ while, because y = w − t(r ), it also
rises with larger income w but falls with distance from CBD. Population density also
rises with larger green space amenity parameter α and lower maintenance cost γ. This is
because green urban area services bring more utility to residents, who may be allocated
to smaller residential plots. Also, since lower maintenance costs entice the planner to
enlarge green urban areas, residents obtain higher amenity and can be offered smaller
residences.
The optimal city border b∗ is determined by the first order condition w.r.t. b (2.8)
evaluated at V (y, a, u) = (y + a − u)2 /(2θ ) and r = b∗ . It is instructive to first study
the case where there is no green urban area at the city border so that x ∗ (b) = 0 and
√
n∗ (b) = [y(b) − u] /θ. The above border condition simplifies to y(b) − u = 2θR A ,
√
which determines the city border as the solution of t(b∗ ) = w − u − 2θR A . Therefore,
the city spreads with higher earnings and shrinks with higher agricultural land rents,
preference for residential space and also with stronger land regulation (through lower
u). The city border is however independent of the preferences and costs for green ur-
ban area. In the presence of green urban area at city border b (g+ ∗ ( b ) > 0), condition

(2.8) does not accept a closed form solution and depends on preferences for green ur-
ban area and maintenance cost. We can totally differentiate (2.8) and get the following
comparative statics (see Appendix A):

db∗ db∗ db∗ db∗ db∗ db∗ db∗

, , , , <0< , . (2.22)
dβ dγ dθ du dR A dα dw

This confirms the comparative statics (2.11) obtained under more general preferences.
In addition, it is here shown that the optimal city shrinks with stronger preferences
APPENDIX 35

for residential space θ. This is because the land value falls and lies above R A within a
smaller area when θ gets larger.
R b∗
The optimal city population is given by N ∗ = 0 n∗ `dr, which rises with larger city
border b∗ and higher population density n∗ . Using (2.21) and (2.22), it readily comes
that
dN ∗ dN ∗ dN ∗ db∗ dN ∗ dN ∗
, , , <0< ,
dγ dθ du dR A dα dw
In particular, city population rises with smaller maintenance cost and higher amenity
parameter for green urban area.

Finally, we discuss the difference between green urban areas and local public goods
in the case of hyperbolic preferences. Urban green space are local public goods as they
provide the same amenity to the residents in its close vicinity and incur maintenance
costs that are independent of the number of users. The difference lies in the land in-
tensity of green urban areas. The urban economic literature presents local public good
as spaceless amenities that do not use space (see Fujita Thisse, 2004, for synthesis). In
the case of spaceless local public good (β = g = 0), land supply is fully supplied
to residents so that the land market clears: sn = 1. Condition (2.19) simplifies to
α (y + αx − u) − θγ = 0, in which the LHS measures the marginal surplus of local
public good and rises in x. This implies that the surplus is convex in x. The plan-
ner must therefore choose between the minimum level x = 0 and the maximum level
x = x o ≡ (y + u) /α that is reached when individual is asked to consume no compos-
z = (y − αx + u) /2 = 0. Thus, because the marginal surplus increases
ite good, i.e. b
with net income y, the surplus is a convex function of y and there exits a net income y
above which the surplus is larger at x = x o than at x = 0. Hence, the planner sets the
spaceless local public good to x ∗ = x o if y ≥ y and to x ∗ = 0 otherwise. Population
density is respectively given by no = (y + αx o − u) /θ = 2y/θ and (y − u) /θ. Since net
income y falls with distance from CBD, the planner provides lower and lower levels of
local public goods to locations farther from the center and none beyond the distance y.
Population density falls with distance from CBD. It falls half as fast beneath y.

Proposition 6 Under hyperbolic preferences, services offered by spaceless local public goods
fall with distance from CBD and vanish at some distance from it.

Proofs Comparative statics on population density n∗+ : Most comparative statics

d log n∗
are trivial except the ones on α and β. Yet, one can show dα + > 0 ⇐⇒ 1 + 6θγβ +

 2  
1 − β2 y − u + αβ + 2 y − u + αβ > 0. Comparative statics on β are unfortunately
inconclusive.
To compute the comparative statics on the city border b∗ , let us denote (??) as
F ( x (b) , α, β, γ, θ, R A ) = 0. Hence, db∗ /dα = − (∂F/∂α) / (dF/db) has the sign of
36 APPENDIX

∂F/∂α and the same argument applies for other parameters. At the optimal city border,
we compute ∂F/∂α = x ∗ Va > 0, ∂F/∂β = − x ∗ V < 0, ∂F/∂γ = − x ∗ < 0, ∂F/∂θ =
x ∗ ∂V/∂θ < 0, ∂F/∂R A = −1 < 0, ∂F/∂w = Vy > 0, ∂F/∂r = −Vy (1 − αx ∗ ) dt/dr < 0
and ∂F/∂u = Vu (1 − αx ∗ ) < 0.

Cobb Douglas preferences Under extreme value theorem, there always exists a
global maximum for a continuous function on a compact set; therefore, there always
exists solution for city planner as x is always within the domain [0, β1 ].
We denote the marginal welfare by F ( x ) − G (y) where F ( x ) = e Ax ( B − Cx ) and
− 1− ϕ 1
G (y) = κγy φ u φ > 0 are the LHS and RHS of the FOC condition (2.17) while x ∈
[0, 1/β], A = αϕ/φ > 0, B = αϕ − βφ, and C = αβϕ > 0. If B ≤ 0, then F ( x ) − G (y) < 0
for x ∈ [0, 1/β], so that the optimal service is the corner solution: x ∗ = 0. If B > 0 then,
F (−∞) = 0, F (0) > 0 and F (∞) = −∞ while F ( x ) has a unique maximum at x̄ such
that F ( x̄ ) ≥ F ( x ) and F ( x̄ ) > 0. It can be checked that F (1/β) < 0. As a result F ( x ) has
a single root for x < 1/β. If AB < C ⇐⇒ x̄ < 0, F ( x ) is a decreasing function and
there exists a single root x 0 for F ( x ) − G (y) = 0 iff F (0) > G (y). The optimal service
is then x ∗ = x 0 . If AB > C ⇐⇒ x̄ > 0, F ( x ) is a bell-shaped function and there exist
two roots for F ( x ) − G (y) = 0 iff F ( x̄ ) > G (y). The highest root x 00 has F 0 ( x 00 ) < 0 and
gives the optimal service: x ∗ = x 00 . To sum up, there exists an interior optimal service
x ∗ > 0 if F (max(0, x̄ )) > G (y). At the interior optimal service x ∗ , F 0 ( x ∗ ) < 0.
To get the comparative statics w.r.t. y on the interior optimum x ∗ , we denote K ( x, y) ≡
F ( x ) − G (y). So, Kx ( x ∗ , y) = F ( x ∗ ) < 0, Kxy ( x ∗ , y) = 0, Ky ( x ∗ , y) = − G 0 (y) > 0, and
Kyy ( x ∗ , y) = − G 00 (y) < 0. Then, we get dx ∗ /dy = − Fy /Fx > 0 and d2 x ∗ /dy2 =
− Kyy Kx − Ky Kxy / (Kx )2 = −Kyy /Kx < 0.


APPENDIX 37

Appendix B: Data and robustness check for urban green space

profiles
Definitions

We first summarize the definitions of urban area entities. Eurostat uses three level of
spatial units that are based on clusters of high density grid cells and urban cores. Fol-
lowing Eurostat a high density grid cells are defined as population grid cells of one kilo-
meter square with at least 1, 500 inhabitants. A cluster of high density cells is a set of high
density grid cells that are each surrounded by at least five other high density cells (over
the eight surrounding cells). Clusters exclude the high density grid cells that are not
connected or isolated. An urban core is a cluster of high density cells that totals at least
50, 000 inhabitants. The first level of spatial unit is the City. It is related to an urban core
and defined by the local administrative boundary so that more than 50% of inhabitants
live inside the associated urban core.22 The second level of spatial unit is the Greater
City, which is created when the urban population resides far beyond the local admin-
istrative boundaries. Greater cities like Greater Manchester, Greater Nottingham and
Greater Paris have been defined with alternative but close definitions. In most cases,
a Greater City contains a single City. The City, Greater City (CGC) includes those two
levels. A Functional Urban Area (FUA) combines the city area and its commuting zone,
as defined in the EU-OECD functional urban area definition (OECD, 2013). A FUA
includes the ”working catchmen area” of a city and is defined as the collection of all
surrounding municipalities with at least 15% of their employed residents working in
the associated urban core. Figure B1 presents the three different levels of spatial units
for Dublin.

Figure B1: Eurostat levels of spatial units (Source: Eurostats)

22 There are some exception to this rule when the geography is disrupted by a river, a lake, fjords, or
steep slopes etcetera, making it hard to recover the urban core. In this case, the City can be added to cover
this urban centre.
38 APPENDIX

Hypothesis testing of Table 2.3

Table B1 reports the p-value associated to Table 2.3 columns (1) to (6). Those values are
extremely small, corroborating the strong non-monotone relationship.

Table B1: P-values for one-sided t-test for baseline models

(1) (2) (3) (4) (5) (6)

H0 : β 0 > 0 1.23 × 10−58 1.39 × 10−6 2.01 × 10−160 2.00 × 10−24 3.55 × 10−10 1.28 × 10−24
H0 : β 1 > 0 4.16 × 10−11 4.99 × 10−11 7.88 × 10−11 2.29 × 10−4 4.31 × 10−4 4.90 × 10−7
H0 : β 2 < 0 5.14 × 10−20 5.60 × 10−20 1.31 × 10−19 5.11 × 10−10 1.55 × 10−5 2.68 × 10−9
Note: The covariance matrix for residuals used for hypothesis testing is clustered at city level and
heteroskedastic-robust.

Polycentric cities

In the GMES EU27 database, some cities develop in a polycentric way and host several
urban cores that are physically separated but economically connected. Column (6) in
Table 3 excludes EU polycentric cities using an OECD study on polycentric cities. In the
latter, OECD assigns a polycentric status to a FUA using the percentage of residential
population commuting from one urban core to another. A polycentric city includes
two or more urban cores, which are connected and attract at least 15% of each other’s
population as workforce (OECD, 2013).
Furthermore, the column (6) of Table 2.4 focuses on the 21 countries that are com-
mon to the 27 countries in GMES EU27 database and the 23 countries in the OECD
study. This excludes six EU27 countries (Bungaria, Lithuania, Latvia, Malta, Roma-
nia and Cyrus) and 2 OECD countries (Switzerland and Norway). Table B2 reports
the population statistics of monocentric and polycentric cities for those 21 countries.
There are 18 FUAs with two urban centres, and 6 FUAs with more than two urban
cores (Barcelona, Paris, Lyon, Amsterdam, Stockholm and London). We will exclude
all those FUAs in our sample. Finally, since the OECD study concentrates on cities with
more 500,000 inhabitants, we keep only the cities with same sizes in the robustness of
column (6) of Table 2.4. This excludes the GMES 27 ’small’ sized cities that we used in
other robustness analyses.
Table B3 provides further robustness analysis on the city border determination. It
reports regressions coefficient when the city border is defined by various shares of res-
idential density: 15%, 20%, 25% and 30%. Coefficient signs are unaffected by those
definitions in both the quadratic and cubic models.
Table B4 presents regressions with alternative weighting schemes for observations.
Larger cities are less numerous but include larger populations and larger numbers of
annuli, which could influence the results. Table B4 reports the regression coefficients
corresponding to Table 2.3 with weights proportional to the city population (columns
APPENDIX 39

Table B2: Descriptive statistics for EU metropolitan forms in OECD database

average sd min max obs.

Monocentric Cities
Population of metro area (thousands) 1235 890 445 4399 87
Population of city area (thousands) 772 633 90 3467 87
Duocentric Cities
Population of metro area (thousands) 1713 1609 561 7079 18
Population of city area (thousands) 1187 1274 314 5264 18
Polycentric Cities (with more than three centers)
Population of metro area (thousands) 5786 5030 1096 12401 6
Population of city area (thousands) 4487 4105 1331 9942 6
Note: This table includes 21 countries: All EU27 except Bungaria, Lithuania, Latvia,
Malta, Romania and Cyrus. Metro population is computed from Census 2010 accord-
ing to OECD metropolitan boundary maps in 2001. The metropolitan population is
very similar to the FUA population provided in Urban Audit database of Eurostat. For
more details, see OECD Metropolitan Explorer database, version June 2016 (OECD,
2016).

Table B3: Profile of green urban areas (EU27): Robustness with different levels of cut-off

Dependent variable: Share of Green Urban Area

Quadratic Cubic
(1) (2) (3) (4) (5) (6) (7) (8)
Border cut-off 15% 20% 25% 30% 15% 20% 25% 30%
Dist. 0.065∗∗∗ 0.102∗∗∗ 0.119∗∗∗ 0.111∗∗∗ 0.303∗∗∗ 0.294∗∗∗ 0.264∗∗∗ 0.260∗∗∗
(0.015) (0.016) (0.018) (0.020) (0.037) (0.039) (0.041) (0.045)
Dist. square −0.116∗∗∗ −0.127∗∗∗ −0.123∗∗∗ −0.098∗∗∗ − 0.697∗∗∗ −0.591∗∗∗ −0.470∗∗∗ −0.449∗∗∗
(0.014) (0.014) (0.016) (0.018) (0.079) (0.080) (0.085) (0.093)
Dist. cubic 0.380∗∗∗ 0.302∗∗∗ 0.225∗∗∗ 0.226∗∗∗
(0.048) (0.048) (0.052) (0.056)
Constant 0.099∗∗∗ 0.100∗∗∗ 0.098∗∗∗ 0.093∗∗∗ 0.078∗∗∗ 0.082∗∗∗ 0.085∗∗∗ 0.080∗∗∗
(0.003) (0.004) (0.004) (0.005) (0.004) (0.005) (0.005) (0.006)
City FE X X X X X X X X
City Sample All All All All All All All All
Obs. 16,819 13,091 10,188 8,156 16,819 13,091 10,188 8,156
Adj. R2 0.399 0.369 0.344 0.327 0.416 0.379 0.349 0.331
df 16,513 12,785 9,882 7,850 16,512 12,784 9,881 7,849
F Statistic 37.65∗∗∗ 26.13∗∗∗ 18.50∗∗∗ 13.97∗∗∗ 40.09∗∗∗ 27.13∗∗∗ 18.84∗∗∗ 14.21∗∗∗
Note: This table shows results from the regression of share of green urban area within each annulus on the relative distance and its square to the
city center using control variables and a constant. The relative distance is normalized to one for farthest annulus for comparison. All columns
include controls for country and city size. The percentage of cut-off points is defined as the annuli with percentage of urban fabric areas over
the total area of the city within the annuli smaller than the percentage of the cut-off points indicated. Standard errors are clustered at city level
and reported in parentheses. The row ”df” reports the degree of freedom. Significance levels are denoted by ∗ for p<0.1, ∗∗ for p<0.05 and ∗∗∗
for p<0.01.
40 APPENDIX

Table B4: Profile of green urban areas (EU27): Robustness with weighted OLS

Dependent variable: Share of Urban Green Space

Quadratic Cubic
(1) (2) (3) (4) (5) (6)
Distance 0.166∗∗∗ 0.074∗∗∗ 0.070∗∗∗ 0.344∗∗∗ 0.267∗∗∗ 0.278∗∗∗
(0.011) (0.016) (0.019) (0.029) (0.040) (0.047)
Distance square −0.187∗∗∗ −0.065∗∗∗ −0.057∗∗∗ −0.625∗∗∗ −0.523∗∗∗ −0.551∗∗∗
(0.009) (0.016) (0.019) (0.061) (0.094) (0.111)
Distance cubic 0.288∗∗∗ 0.295∗∗∗ 0.318∗∗∗
(0.038) (0.063) (0.074)
Constant 0.084∗∗∗ 0.079∗∗∗ 0.079∗∗∗ 0.068∗∗∗ 0.062∗∗∗ 0.060∗∗∗
(0.005) (0.005) (0.005) (0.006) (0.006) (0.007)
City Size FE X X X X X X
City Sample All All All All All All
Weighting City Pop. Pop. Dens. 1 Pop. Dens. 2 City Pop. Pop. Dens. 1 Pop. Dens. 2
Observations 13,117 13,117 13,117 13,117 13,117 13,117
Adj. R2 NA 0.111 0.115 NA 0.118 0.123
df 13086 13086 13086 13085 13085 13085
F Stat. 69.30∗∗∗ 27.99∗∗∗ 27.47∗∗∗ 68.17∗∗∗ 28.77∗∗∗ 28.20∗∗∗
Note: We include different weighting strategy. Weights are proportional to the city population (columns (1)
and (4)). Columns (2) and (5) include population density measured by dividing total city population with
area of whole city. Columns (3) and (6) measure city population by dividing city population to areas of
artificial urban fabric, which mostly used for residential purpose. Standard errors are clustered at city level
and reported in parentheses. The row ”df” reports the degree of freedom. Significance levels are denoted by
∗ for p<0.1, ∗∗ for p<0.05 and ∗∗∗ for p<0.01.

(1) and (4)) and the two different measures of city population density (columns (2)-(3)
and (5)-(6)). Coefficients are invariant to weighting specifications, which suggests low
risk of misspecification.
Chapter 3

Welfare Evaluation of Green Urban

Areas

3.1 Introduction
Urban green area is one of the keys for building resilient and sustainable urban systems.
These areas provide many valuable local services to urban dwellers such as mitigating
air pollution and urban heat islands (Bowler et al., 2010; Manes et al., 2012), support-
ing ecological learning (Barthel et al., 2010), and enhancing human health (White et al.,
2013). They are also an essential block of urban social life as they help to bond urban
social ties, providing a place for recreational opportunities and other social benefits
(Colding and Barthel, 2013; Saldivar-Tanaka and Krasny, 2004; White et al., 2013). It
is no doubt that high-quality urban green spaces can bring many relevant services to
urban areas, and their locations constitute a crucial component of urban planning in
the quest for sustainable cities and communities (Tyväinen et al., 2005). This fact poses
a critical question for city planners: how do the residents value them? Such informa-
tion is vital to design more desirable residential communities and to thrive by limiting
development.
There are three main approaches to evaluating the value of green and open areas.
The first approach is the contingent evaluation (CV) method, which requires house-
holds to directly specify their willingness to pay (WTP) for the preservation or im-
provement of specific urban green areas. However, this method depends on hypo-
thetical scenarios, and it does not reflect the actual behaviors of all households. The
second approach is the hedonic pricing (HP) method, which reveals the marginal WTP
of households through the premium that they give to housing prices through market
transactions. The HP method adequately reflects the actual evaluation of households;
however, it does not measure the nonmarginal changes for green areas. It also suffers
from the ‘Tiebout bias’ (see Kuminoff et al., 2013), as households can move to other
districts in response to a policy change. The third approach is the equilibrium sorting

41
42 CHAPTER 3. A WELFARE EVALUATION OF GREEN URBAN AREAS

(ES) method, which takes into account the sorting processes of households in the mar-
ket by simulating a new equilibrium. The gains from the ES method come at a cost,
as it is necessary to specify the preference structure and the choice sets available for
consumers. Additionally, the housing supply is treated as fixed to allow for simple cal-
ibration (Freeman et al., 2014; Kuminoff et al., 2013). A detailed survey of the literature
is reported in Section 2.
In this paper, we utilize a different approach. We first introduce the urban green
areas inside an urban land market model and formalize the channel in which the green
areas enter households’ decisions and land prices. We employ two different house-
holds’ preferences: Cobb-Douglas and hyperbolic linear utility. The marginal benefit of
urban green areas varies across locations within cities concerning their existing covers
in the residential areas, the population density and the opportunity cost of land at each
location. The market equilibrium results in the optimal residential land use at each lo-
cation across urban areas with a specified level of urban green areas. We then calibrate
the equilibrium of residential lands using the data on detailed residential land covers,
green urban areas and population density. To control for the potential endogeneity of
urban green provision from urban authorities, we employ the green areas in the past as
an instrument for the current green areas. We back up the parameters and run counter-
factual simulations under the scenario that urban planners reduce green areas by half.
By using this approach, we can account for the new equilibrium in the housing market
and households’ choices over housing consumption. We consider both the case of open
cities with free migration and closed cities with no migration.
We consider only small and scattered green urban areas in this paper. It is because
95% of urban parks have a surface lower than 0.1 square kilometer, and the average dis-
tance between the two parks is lower than 400 meters in our sample of 305 of the most
populated cities in Europe. We estimate residential land use and take the estimated
parameters to study the value of green urban areas implied by our theoretical urban
land market model. For the parameter estimation, we use the geographical database of
the European Environment Agency’s Urban Atlas data, which describes the land use of
305 EU cities with more than 100,000 inhabitants. These data report land use and cover
across Europe using harmonized Earth observations (EOs), which are combined with
Eurostat Urban Audit statistical data. The data represent a unique source of reliable
and comparable European urban planning data. To our best knowledge, this is the first
paper that uses the geographical land covers in both a theoretical model and empirical
estimation of European urban land uses.
Through counterfactual exercises where 50% of urban green areas are removed, we
show that open cities lose more than 6% of their population if those areas are left un-
used. The total loss for landlords is approximately e150 million per city under the
Cobb-Douglas preference and more than e300 million in the case of the hyperbolic
preference if green areas are not converted into residential land. Converting those ur-
3.2. LITERATURE REVIEW 43

ban green areas into residential land, however, would increase residential surfaces and
increase the total housing market value by approximately e50 million in the Cobb-
Douglas preference and a loss of less than e10 million in the hyperbolic preference. In
closed cities, residents lose utility, which can be restored by a subsidy of nearly 6-9% of
their net annual income (it is equivalent to a negative willingness to pay from house-
holds). Those results are robust for both types of preferences. This exercise, based on
a dramatic decline in green urban areas, suggests that those areas provide highly valu-
able amenities to residents. However, when we allow for the land conversion of the
erased green areas into new residential covers, the net costs of the green areas are up to
4.2% of annual household income, suggesting that the land opportunity costs of urban
green areas are substantial. Our results are in line with those of Cheshire and Sheppard
(2002), who also find that the net cost of restrictive land use planning can reach 3.9% of
the household’s income.
Our contributions are twofold. First, we utilize new datasets, including the geo-
graphical land covers, and population grids to estimate the parameters of our theoreti-
cal models. We also account for the potential endogeneity in green provision by using
the green areas from the past to instrument for the current level of green provision.
With these data, we account for the actual residential choices of households in reality
instead of hypothetical questions as in the CV method. Second, the use of an urban
model together with two different land conversion policies allows us to account for
the change in residential land supply and enables us to compute the general equilib-
rium willingness to pay of households for a significant change in green provision. This
approach, compared with the HP model, addresses the nonmarginal changes in green
areas, and it also relaxes the restrictions of fixed housing supply in the sorting equilib-
rium method. Our tractable and clean models and counterfactual exercises also come
with the cost of imposing a specific preference for households. We address this issue
partially by considering two types of utility functions: Cobb-Douglas and hyperbolic.
Second, we only consider one class of household for each city. The extension for mul-
tiple types of households in this model and the calibration could be a helpful direction
for future research.

3.2 Literature review

There is a large and growing literature quantifying the value of open and green spaces
in urban systems. The first branch of literature is to evaluate green areas through con-
tingent valuation (CV). This approach designs surveys with hypothetical scenarios for
environmental goods and asks survey respondents to specify their willingness to pay
(WTP) for the preservation or improvement of specific urban green areas. Researchers
aggregate those values and estimate the monetary value of this environmental asset.
For Europe, Elsasser (1996) and Tyrväinen (2001) use the CV method for the city of
44 CHAPTER 3. A WELFARE EVALUATION OF GREEN URBAN AREAS

Hamburg in Germany and Joensuu and Salo in Finland. They find that the mean WTP
was approximately e42 per person per annum in Germany1 and e60-144 in Finland2 .
Schindler et al. (2018) find that the reported willingness to pay for green area access
increases with household’s average income. Even though the CV method provides
valuable information on how much value each resident places on the usage of green
areas, it is based on hypothetical scenarios and does not reflect their actual valuations
and behaviors. The WTP measured in other studies revealed the preference approach,
in that hedonic pricing or equilibrium sorting reports a much higher value. The WTP
measured in our paper also has a higher value of approximately 6-9% of annual house-
hold income. Nevertheless, the WTP estimated under the Cobb-Douglas preference in
our model also increases with income quantiles, which is similar to the pattern reported
in Schindler et al. (2018).
The second approach is hedonic pricing (HP), which evaluates the value of green
amenity through its capitalization in the housing market. The premise underlying this
method is that if households enjoy green amenity in a location, they will pay a premium
for it. Most of the studies use housing transaction data, and they find that the proxim-
ity to parks and open spaces induces an increase in home prices (Bolitzer and Netusil
2000; Geogrehan et al., 1997; Lutzenhiser and Netusil, 2001; Smith et al., 2002). Ander-
son and West (2006) further reveal that the value of proximity to open spaces is higher
in dense and high-income areas. Luttik (2000) find that a pleasant view leads to an in-
crease of 6-12% in house price in the Dutch land market, and Tyrväinen and Miettinen
(2000) estimate a 5% increase in home price in Finland. Tyrväinen and Miettinen (2000)
estimate that a one-hectare green area attributes to an increase of e11-15 million in the
total value of surrounding houses in dense central districts and e4-5 million in single-
home and less dense areas. In a competitive land market, this information helps us to
infer the marginal willingness to pay for green amenity. Even though this information
is undeniably valuable, it does not capture nonmarginal changes in urban green areas.
A unique departure from this framework is Cheshire and Sheppard (2002). They use
hedonic pricing methods to estimate the implicit price for accessible open spaces and
then embed them into a monocentric model and calibrate the data for the city of Read-
ing, UK. The use of an urban model allows them to incorporate the general equilibrium
effects in their analysis. They estimate that the gross benefit of public open spaces is
£2,424 per household per annum (in 1984 price). However, when accounting for the
cost of restrictive residential lands3 , the net loss from planning could be as high as 3.9%
of annual household income. In our paper, when urban planners remove half of the
urban green areas, we find a WTP of e970 - 1300 per year per habitant to avoid such a
1 in
terms of the 1995 price, as cited by Tyrväinen et al. (2005), pg. 102.
2 in
terms of the 2000 price, also per person and per annum
3 They compare between the current level of 37.9% of residential land to a less restrictive scenario of

42.5% of residential land

3.3. THEORETICAL MODEL 45

policy. However, when we account for the reconversion of those lands into residential
uses, the net cost of these green areas is estimated at 3.2%-4.2% of annual household
income, which is in line with the outcomes from Cheshire and Sheppard (2002).
The third approach is the equilibrium sorting (ES) approach, which seeks to evalu-
ate the large changes in public policies. In contrast to the HP estimation, which assumes
that a substantial change in amenity level does not affect household locational choice,
the ES model seeks to characterize the household’s sorting process in response to a
policy change. It is then straightforward to measure both marginal and nonmarginal
changes in urban amenity policies by simulating a new sorting equilibrium.4 Wu and
Cho (2003) apply this method to the environmental amenities in Portland, Oregon, and
find that a 1.2% increase in the share of parks and open spaces is equivalent to a $100 in-
crease in educational expenditure in household choices for communities. Walsh (2007)
uses this sorting method to evaluate the open space protection policy in Wake County,
North Carolina, and finds that an open space protection policy in denser areas can lead
to a 11.4% increase in house price and approximately 8.14% in less dense areas. Our
paper also reports a larger effect in denser areas near the city core and a smaller effect
in land price near urban fringes. Klaiber and Paneuf (2010) use a similar approach to
estimate household preference for open space, and they find that the general and par-
tial equilibrium becomes diverse with the magnitude of the change in policy. However,
this ES method is not without constraints. One constraint is that it is limited in choice
sets for housing types due to the difficulties in increasing dimensions for calibration.
Hence, most of the papers use the sorting process through the neighborhood instead
of house types. They also have to specify the distribution of preference heterogeneities
(Freeman et al., 2014). Second, the housing supply is treated as fixed to allow for simple
calibration (Freeman et al., 2014; Kuminoff et al., 2013). In our paper, we allow for the
land conversion policies, as discussed in the above section, which relaxes the housing
supply constraints in a new equilibrium.

3.3 Theoretical model

The model considers a circular monocentric city. The economy consists of a mass N of
individuals living within city boundary and a central business district (CBD) where all
households commute to work. We denote b ∈ R0+ as the distance between the CBD
and the city border. The population density is defined as the number of individuals in
a unit of area at distance r from the CBD and is denoted by the function n : [0, b] → R+ ,
which varies across the city.
In this model, we focus on small green urban areas that spread across the city and
are closely accessible to the local community around its location.5 Green urban areas
4 For a comprehensive review, see Kuminoff et al. (2013).
5 Because of proximity of small parks, residents’ travel costs to parks can be neglected here as a first
46 CHAPTER 3. A WELFARE EVALUATION OF GREEN URBAN AREAS

provide quick and frequent access to greenery, quiet, children’s parks, socialization
areas, etc. We consider the few blocks in the vicinity of a green urban area as our
unit of area or patch and model the urban area in a continuous fashion. In a unit of
area at distance r from the CBD, green urban areas offers a service x : [0, b] → [0, x ],
x ∈ R0+ , to the local community living in the vicinity. This service brings a level of
amenity a = αx (r ), although it necessitates the use of a fraction of land βx (r ) and
maintenance costs γx (r ). The parameters α, β, γ ∈ R+ distinguish the amenity, land
use and maintenance factors that affect green urban areas. Hence, the fraction of land
used for residential purposes is given by 1 − βx, and the maximum service level x is
bounded by 1/β. We assume absentee landlords, and the outside opportunity value
of land is given by the agricultural land rent R A ∈ R+ . For simplicity, we consider
that rural areas beyond the city border consist of private properties that do not provide
green urban area service for city dwellers (private crop fields, fenced areas, etc.). We
denote the land supply at distance r from the CBD by ` : [0, b] → R+ (e.g., ` = 2πr if the
city lies in a plain disk). In summary, land at distance r from the CBD includes a surface
βx (r )`(r ) of maintained green urban area and a residential area [1 − βx (r )] `(r ) , and it
hosts n(r )`(r ) residents who all benefit from the green urban area amenity αx (r ).
Individuals consume a quantity z of nonhousing composite goods and a quantity
s of residential space, while they benefit from the amenity a of a green urban area.
They are endowed with the utility function U (z, s, a), which is assumed to be concave
and increasing for each variable. We assume that demands for nonhousing composite
goods, residential space and amenity are gross substitutes such that U has negative sec-
ond derivatives and positive cross derivatives.6 As individuals are homogeneous, they
work and earn the same income w ∈ R+ in the CBD. Workers incur a total commuting
cost t : [0, b] → R+ with t(0) = 0 and dt/dr > 0. The price of composite good z is
normalized to 1 without loss of generality. From this point on and whenever there is no
confusion, we dispense the functions a, `, n, s, t, x, z and R with reference to distance r.
approximation. This contrasts with the literature on the location of scarce public goods, where the travel
to public facilities must be considered (see Cremer et al., 1986 and followers).
6 We do not formally disentangle the land space used for houses and private gardens. First, our data

on urban land cover do not permit to identify each separate item. Second, the empirical literature is am-
biguous about the substitution effect of private gardens on the use of parks (also called the ”compensation
effect”). For instance, Talen (2010) and Caruso et al. (2018) find no significant relationship between the
ownership of a private garden and the frequency of park visits. Finally, note that it can be shown that our
theoretical results are unaltered if individuals have utility function U (z, S, a) over a house service function
S that is a homothetic consumption bundle S(s g , sh ) of land consumption for houses and private gardens
sh and s g . In this case, private gardens substitute for parks, but the elasticity of substitution between
private gardens and parks is the same as that between houses and parks.
3.3. THEORETICAL MODEL 47

3.3.1 Demand side

In most modern cities, residents freely choose their residential locations and spaces.
They make their decisions according to the land rent values signaled in the urban land
market. Given that urban green areas generate externalities to residents, it is impor-
tant to highlight the conditions under which the competitive land market replicates the
social optimum. We here discuss the equilibrium allocations in a competitive land mar-
ket for an exogenous profile of green urban services x : [0, b] → [0, 1/β) and ameni-
ties a = αx. A household’s budget constraint is given by z + Rs + t ≤ w, where
R : [0, b] → R+ is the land rent function of distance to the CBD. Households, there-
fore, face the following optimization problem.

max U (z, s, a) s.t. w − t ≥ z + Rs (3.1)

s,z

In line with the literature, we assume that the land rent is paid to absentee landlords
(Fujita (1989), Fujita and Thisse (2012)). Denote y ≡ w − t as the disposable income after
subtracting the commuting cost. The maximization problem (3.1) yields the demand for
the composite good
z(y, R, a) = y − Rs( R, y, a) (3.2)

and the housing demand s( R, y, a), which is the unique solution of the following equa-
tion:
Us (y − Rs, s, a) − RUz (y − Rs, s, a) = 0. (3.3)

We then obtain the indirect utility for each household

V (y, R, a) ≡ U (z( R, y, a), s( R, y, a), a) (3.4)

However, housing price or land rent is rarely exogenous for households. The higher
the desirability is of one location, the higher is the demand, and it leads to higher land
rent. To investigate the equilibrium land price, we utilize the bid rent approach in the
following section.

3.3.2 Competitive land market equilibrium

In a competitive land market equilibrium, each land slot is awarded to the highest bid-
der, and individuals have no incentives to relocate within and outside of a city. There-
fore, they reach the same utility level ue , where the superscript e refers to the equilib-
rium value. Households bid up to (w − z − t)/s for each unit of residential space. Their
bid rent ψ : [0, b] → R+ is a function of distance r from the CBD such that
y−z
ψ = max s.t. U (z, s, a) ≥ ue . (3.5)
s,z s
As individuals compete for land, they increase their bids to make their participation
constraint binding and obtain the equilibrium utility level ue . To solve the problem,
48 CHAPTER 3. A WELFARE EVALUATION OF GREEN URBAN AREAS

we first denote z̃(s, a, ue ) as the solution of U (z, s, a) = ue . As household utility is an

increasing and concave function of composite good z, it is apparent that z̃ exists and is
unique. The bidding problem in equation (3.5) can be rewritten as follows
y − z̃(s, a, ue )
ψ = max (3.6)
s s
The equilibrium housing slot size is given by the first-order condition

z̃(s, a, ue ) − sz̃(s, a, ue ) = 0 (3.7)

We denote the solution for equation (3.7) as b s(y, a, ue ), which is also the households’
optimal consumption.7 Therefore, households’ optimal consumption are given by the
functions b s(y, a, ue ) and b z(y, a, ue ) as well as the bid rent ψ b (y, a, ue ). Ceteris paribus, we
have the equilibrium housing slot b s that increases with distance to CBD (or lower y)
and decreases with amenity level a. By the envelope theorem, the bid rent ψ b increases
with disposable income y and amenity a and decreases with reservation utility ue ; that
is, ψ b a > 0, while ψ
by, ψ b u < 0.
A competitive land market equilibrium is defined as the set of functions (z, s, R, n)
and scalars (b, N, ue ) satisfying the following four conditions. First, individuals choose
their optimal consumption: z = b z(y, a, ue ) ≡ z(ψb (y, a, ue ), y, a) and s = b s(y, a, ue ) ≡
s(ψb (y, a, ue ), y, a). Second, land is allocated to the highest bidder: R = max{ψ b (y, a, ue ), R A },
with R = ψ b (y, a, ue ) if n > 0, and R = R A if n = 0. Third, the land market clears ev-
erywhere: nb s(y, a, ue ) = (1 − g) if n > 0. Finally, the total population conforms to its
Rb
density: N = 0 n 2πrdr. Here, N is taken as exogenous in a closed city model, while
ue is exogenous in an open city. Within a city, equilibrium land rents are given by the
winning bids such that R = ψ b (y, a, ue ). Since bid rents ψ increase with net income y
and amenities a, the equilibrium land rent R falls with distance from the CBD but rises
with the proportion of green urban area.
In equilibrium, land rents must exceed R A for any location r ∈ [0, b) and be equal
to it at the equilibrium city border be . To simplify the exposition, we assume that Re (r )
crosses R A from above at r = be , which occurs if ψ b (y, a, ue ) lies above R A in the CBD
and falls in r. A sufficient condition is given by

b (w, 0, ue ) > R A ,
ψ
dt b a dx < 0.
−ψ
by + αψ
dr dr
After some reshuffling, this gives

z(w, 0, ue ) + R A b
w>b s(w, 0, ue ), (3.8)
dt dx
> −αe
za . (3.9)
dr dr
7 We
have the second-order condition (∂/∂s)(z̃ − sz̃s ) = −sz̃ss < 0 due to the concavity of the utility
condition.
3.3. THEORETICAL MODEL 49

These sufficient conditions imply that urban productivity is sufficiently high for a city
to exist in the absence of green urban areas and that green urban areas do not have ex-
cessively steep density profiles or do not yield too much spatial variation in amenities.
Sufficiently high wages w and a low amenity parameter α guarantee these conditions.
Under conditions (3.8) and (3.9), a spatial equilibrium exists. The equilibrium city bor-
der be is given by the unique solution of the land arbitrage condition: R(be ) = R A .

Proposition 1 Suppose that conditions (3.8) and (3.9) hold. Then, a competitive land market
equilibrium exists and is unique.

The equilibrium population density is equal to ne = (1 − g)/b

s(y, a, ue ) ≥ 0, while
the equilibrium population aggregates the population density across the urban area as
Z be Z be
e 1 − βx 1 − βx
N = 2πrdr = 2πrdr.
0 s(y, a, ue )
b 0 s(ψ(y, a, ue ), y, a)
b

To obtain more analytical results, we focus on narrower classes of preferences.

3.3.3 Specific preferences

In this subsection, we characterize the equilibrium land rent and population density for
two classes of preferences: Cobb-Douglas and hyperbolic. Cobb-Douglas is a standard
and classic homothetic preference, while the hyperbolic preference is a relatively less
used preference and is nonhomothetic.

3.3.3.1 Cobb-Douglas preferences

We consider the following Cobb-Douglas utility

U (z, s, a) = z1−φ− ϕ sφ e ϕa (3.10)

with φ, ϕ, (1 − φ − ϕ) ∈ (0, 1). Maximizing (3.10) with respect to z and s subject to the
budget constraint z + Rs ≤ y leads to the linear expenditure system:

1−ϕ−φ
z= y, (3.11)
1−ϕ
φ y
s= (3.12)
1−ϕR

and the indirect utility function

V = (φκ )φ y1− ϕ R−φ e ϕa (3.13)

− 1− ϕ 1− ϕ − φ
where κ = (1 − ϕ) φ (1 − ϕ − φ) φ .
Our objective now is to determine the land rent equilibrium at each location in the
city area. Since the housing slot size is given for the highest bidder, what makes a
50 CHAPTER 3. A WELFARE EVALUATION OF GREEN URBAN AREAS

specific location attractive is both its closeness to the working place in the city center
and the green amenity level. Solving the bidding problem in (3.5), we obtain the land
rent at equilibrium:
 − 1
b = κ ue y−(1− ϕ) e− ϕa φ .
ψ (3.14)

3.3.3.2 Hyperbolic preferences

In this section, we utilize a new class of model, which is not homothetic as in the Cobb-
Douglas preference. We assume that individuals are endowed with the utility function
as given below:
θ
U (z, s, a) = z − +a (3.15)
2s
where preferences for residential space are represented by a hyperbolic function parametrized
by θ.8 Accordingly, we obtain the following expenditure system:
r
θR
z = y− (3.16)
2
r
θ
s= (3.17)
2R
and the indirect utility function
√
V = y− 2θR + a. (3.18)

As we can see from the previous part, the consumption of composite goods under
Cobb-Douglas is constant and does not change with the new equilibrium of the housing
market. However, with the hyperbolic function, the new equilibrium in the housing
market will affect the consumption of the nonhousing good: higher land rent results in
lower consumption of z. Solving the bidding problem in (3.5), we obtain the land rent
at equilibrium for the hyperbolic preference as

e 2
b = (y + a − u ) .
ψ (3.19)
2θ
The following section applies those theoretical results to the data on European cities.

3.4 Empirical analysis and welfare valuation

In this section, we estimate the relationship among population density, green urban
spaces, distance to the city center and other controls, and we apply the estimates to our
theoretical model. We then estimate the effect of altering the surfaces of green urban
areas.
8 This yields a demand for residential space that has a price elasticity between one and zero (see Mossay

and Picard 2011).

3.4. EMPIRICAL ANALYSIS AND WELFARE VALUATION 51

3.4.1 Data

We utilize two main datasets in this paper. The first dataset is the population grid from
Eurostat. The second dataset is the Urban Atlas implemented by the Global Monitoring
for Environment and Security (GMES) service and provided by the European Environ-
ment Agency (EEA), for the time period 2005-2007. The Urban Atlas provides a clas-
sification of city zones that allows for a comparison of the density of residential areas,
commercial and industrial zones and extent of green areas. The Urban Atlas uses Earth
observation satellite images with 2.5 m spatial resolution.9 According to the GMES, the
dataset covers the functional urban areas (FUAs) of the EU cities with at least 100,000
inhabitants.10 FUAs include land with both commuting distance and time below the
critical levels defined by Eurostat.11 The dataset includes all capital cities and covers
nearly 300 of the most populous towns and cities in Europe (EU 27).12
In this paper, we use the data on “green urban areas” (class 14100), which are de-
fined as artificial nonagricultural vegetated areas. They consist of areas with planted
vegetation that is regularly worked and/or strongly influenced by humans. More pre-
cisely, first, green urban areas include public green areas used predominantly for recre-
ational use (gardens, zoos, parks, castle parks, cemeteries, etc.). Second, suburban nat-
ural areas that have become and are managed as urban parks are included as green
urban areas. Finally, green urban areas also include forest and green areas that extend
from the surroundings into urban areas with at least two sides being bordered by ur-
ban areas and structures and containing visible traces of recreational use. Importantly,
for our study, green urban areas do not include private gardens within housing areas,
buildings within parks, such as castles or museums, patches of natural vegetation or
agricultural areas enclosed by built-up areas without being managed as green urban
areas. It must be noted that green urban areas belong to the Urban Atlas’ class of ”arti-
ficial surfaces”, which includes all nonagricultural land devoted to human activities.13
This class is distinguished from the agricultural, seminatural areas and wetlands, forest
areas and water areas devoted to nonurban activities.
We select the (oldest) town hall locations as the CBDs. Then, we create a set of an-
nuli (rings) around each CBD at 100 m intervals. We define the ”annulus land area” as
9 GMES maps have a 100-times higher resolution than that of traditional maps in the CORINE Land
Cover inventory produced since 1990.
10 See the definition in the Urban Audit database and European Environmental Agency, GMES Urban

Atlas Metadata. Link: https://land.copernicus.eu/local/urban-atlas (accessed on Jan 25, 2018).

11 See the definition in the Urban Audit in EEA, 2015, and the details in Appendix B.
12 Austria, Belgium, France, Germany, Bulgaria, Cyprus, the Czech Republic, Denmark, Estonia, Finland,

Greece, Hungary, Ireland, Italy, Latvia, Lithuania, Luxembourg, Malta, the Netherlands, Poland, Portugal,
Romania, Slovakia, Slovenia, Spain, Sweden, and the United Kingdom.
13 In addition to green urban areas, artificial surfaces include urban areas with dominant residential use,

inner city areas with a central business district and residential use, industrial, commercial, public, military
and private units, transport units, mines, dump and construction sites, and sports and leisure facilities.
52 CHAPTER 3. A WELFARE EVALUATION OF GREEN URBAN AREAS

Figure 1: Berlin Land Use maps

the intersection of the annulus and the land within the urban zone area reported by the
GMES. This area includes artificial surfaces, agriculture, seminatural areas, wetlands
and forest but does not include water areas because those seas and oceans are not ap-
propriate for potential human dwellings. We then compute the share of green urban
area as the ratio of the surface of green urban area to the total land in the annulus land
area for each annulus. Figure 1 displays the land uses of green urban areas (green color)
for Berlin. Whereas urban theoretical models usually assume a neat frontier between
residential and nonresidential spaces, urban data do not provide a clear separation be-
tween residential locations and agricultural areas and forests. In this paper, we choose
to fix the city borders to the annulus for which the ratio of residential surfaces over the
annulus land area falls below 20%. Residential surfaces include urban areas with dom-
inant residential use and inner city areas with a central business district and residential
use. They are shown in red in Figure 1.14 We define the distance from the CBD, dist.
In addition to the GMES, we use the population density from the European popu-
lation grid.16 We calculate the population mass at each distance to the city center and
14 Theuse of other thresholds does not lead to qualitative differences in our empirical results.
16 Formore information on the European population grid, please check the technical report of the GEO-
STAT 1A project from Eurostat. Link: http://ec.europa.eu/eurostat/documents/4311134/4350174/
ESSnet-project-GEOSTAT1A-final-report_0.pdf/fc048569-bc1c-4d99-9597-0ea0716efac3 (Ac-
cessed on May 30, 2018).
3.4. EMPIRICAL ANALYSIS AND WELFARE VALUATION 53

Table 3.1: Summary statistics

average sd min max obs.

City border (km) 4.3 3.2 1.0 24.0 305
City area (km2 ) 84 174 1 1809 305
Number of annuli 43 32 10 240 305
Population in FUA (millions) 0.79 1.29 0.06 12.10 301
Population in CGC (millions): all cities 0.44 0.79 0.03 8.17 305
GDP per capita (e1000/hab.) 26.88 13.06 6.00 83.70 305
Household income (e1000/hab.) 15.46 5.63 3.70 30.90 304
Density (hab./100m2 ) 0.44 0.28 0.10 1.96 303
Residential Space (100m2 ) 0.98 0.44 0.21 2.47 304
City Geographical Controls
Elevation (m) 212 210 −2 1,614 305
Average temperature at Jan 01 (o C) 2.30 4.62 −8.48 15.57 305
Average temperature at July 01 (o C) 19.18 12.27 12.27 28.70 305
Average daily precipitation (mm/day) 1.91 0.60 0.48 4.45 305
Share of Urban Green Land (%) 6.57 4.38 0.62 42.69 304
Share of Residential Land (%) 34.03 4.99 12.85 47.42 304
Share of Industrial and Public Land (%) 16.61 6.02 2.26 47.57 304
Share of Sport and Leisure Land (%) 3.76 3.21 0.00 12.79 304
Share of Forest Land (%) 5.14 6.23 0.00 32.57 304
Share of Agricultural Land (%) 16.29 10.75 0.00 52.13 304
Share of Forest Land within 100m buffer (%) 1.43 2.04 0.00 13.42 304
Share of Agricultural Land within 100m buffer (%) 6.15 5.66 0.00 33.19 304
Note: The GMES database released on May 2010 reports only 301 FUAs for the time period 2005-2007. We
use the Nadaraya-Watson Gaussian Kernel to smooth variations of annuli values.15 GDP per inhabitants
and Household income are taken from Regional Economic Accounts from Eurostat at NUTS3 and NUTS2
level respectively. Note that in Eurostat database, household income level exists only at NUTS2 level.
In eurostat database for household income at NUTS2, there is no data for Luxembourg (NUTS2 code
LU00); therefore, there is only 304 cities instead of 305 cities. The number of inhabitants in each annuli is
calculated based on Eurostat Population Grid. As Eurostat Population Grid 2006 does not cover Cyrus;
hence, we also drop the city cy001l lefkosia in our database. The total number of annuli are calculated
for 303 cities excluding lu001l luxembourg and cy001l lefkosia. For city geographical controls, we take
into account the average for period 1995-2010 for each city.

redistribute the population to the residential area in each annulus.17 Because the Euro-
stat population grid does not cover Cyprus, we exclude Lefkosia, Cyrus. For the income
level of a city, we use the household net income at the NUTS2 level, as reported in the
Eurostat’s Regional Economic Accounts, which provides the finest detail on household
net income. Our results are robust to the use of the city’s per capita GDP at the NUTS3
level.18 Other measures of cities’ exogenous geographical characteristics are taken from
the E-OBS database.19 We control for these exogenous geographical characteristics be-
cause they may affect residential choices. We finally measure the city populations as
the number (millions) of inhabitants living in the city and greater city (CGC) areas,
17 Residential
areas are called ’urban fabrics’ in the GMES.
18 SeeSupplementary Material.
19 The E-OBS database provides detailed data on the daily temperature, daily precipitation, sea level

pressure and elevation across Europe. We acknowledge the E-OBS dataset from the EU-FP6 project
ENSEMBLES (http://ensembles-eu.metoffice.com) and the data providers in the ECA&D project
(http://www.ecad.eu).
54 CHAPTER 3. A WELFARE EVALUATION OF GREEN URBAN AREAS

as defined and reported in the Eurostat databases.20 Table 3.1 presents the summary
statistics.
In this paper, we mostly use the household income that measures the per capita in-
come net of all income taxes and at the NUTS2 level. Household incomes vary greatly
across EU cities, from e3,700 per inhabitant to e30,900 per inhabitant. The average in-
come is e15,460. Household income represents slightly more than one-half of the per
capita production value (NUTS3), which reflects the high tax wedge between produc-
tion cost and net income in the EU. The share of urban green, on average, accounts for
approximately 6.5% of the total surface of city areas. Cities have a rather heterogeneous
share of green urban areas. In our sample, the city with lowest share of green urban area
(0.62%) is Limerick, Ireland, and the city with the highest share (42.6%) is Karlovy Vary,
Czech Republic.21 City elevation also varies greatly, from two meters below sea level
in Amsterdam, the Netherlands, to 1, 614 meters above sea level in the mountainous
city of Innsbruck, Austria. European cities belong to a mild climate zone, with tem-
peratures varying between −8 and +28 degrees Celsius at the lowest and highest day
temperature in winter and summer (measured on January 1 and July 1, respectively,
for the period 1995-2000).22 The average population density is approximately 4,400 in-
habitants per square kilometer and ranges from 1,000 to 9,800 inhabitants per square
kilometer. These numbers are reasonable because the database concentrates on the core
of urban areas with no agricultural or seminatural land use.
In the next section, we estimate residents’ land use under the specification of Cobb-
Douglas preferences. Those are convenient and popular in urban economics. We then
present the same analysis for the hyperbolic preferences.

3.4.2 Estimation and valuation with Cobb-Douglas preferences

We suppose the presence of observable heterogeneity χ in the preference for land plots
or specific characteristics of locations as well as unobservable heterogeneity or mea-
surement errors e. The utility function becomes

U (s, a, z, χ, e) = eχz1−φ− ϕ sφ (e a ) ϕ .

We assume that the transport cost t = w(1 − e−τ (r) ); thus, the net income is given
by y = w − t = we−τ (r) , where τ (r ) is a function of distance to the CBD. For simplicity,
we assume the quadratic form τ (r ) = τ1 × r + τ2 × r2 . Green area amenities are given
by a = αx, which can be written as a function of green urban areas g as a = αg/β,
20 Formore details, see metadata files for urb esms in the Urban Audit database of the Eurostat website.
21 The latter is a spa resort city, which offers many green areas to its visitors. The former city includes
few land surfaces classified as green urban areas because it also has many agricultural and seminatural
lands that can be used for urban green amenities. These outliers do not affect our results.
22 We use observations from the E-OBS databases from the EU-FP6 project (for details, see the references).

Our samples do not contain some northern European cities in Iceland and Norway.
3.4. EMPIRICAL ANALYSIS AND WELFARE VALUATION 55

where α and β are green amenity and land use intensity parameters, respectively. We
can further standardize the amenity value α = 1 and consider β as land use intensity to
provide one unit for green amenities. We obtain the housing slot size in equilibrium as
below:
 1− ϕφ−φ  1
1−ϕ

u  −τ (r) −(1− ϕ−φ) − ϕ βg φ
s= we e (3.20)
1−ϕ−φ
b
eχ
Taking the natural logarithm of (3.20) and adding the heterogeneity and error terms,
we obtain the following residential land use:

 1− ϕ − φ
1−ϕ−φ − φ 1−ϕ−φ

ϕ
s = ln
ln b + ln β − ln w
1−ϕ φ φ
1−ϕ−φ 1−ϕ−φ 2 ϕ 1 1 1
+ τ1 r + τ2 r − g + ln ū − ln χ − ln e.
φ φ φβ φ φ φ

Accordingly, residents have larger land plots for cities with smaller incomes w,
larger distances between residences and the CBD r, smaller green urban areas g, higher
outside utility u, and smaller observable characteristics χ.

3.4.2.1 Estimation

From these results, we build a regression model of residential land use

ln(sijc ) = ϑ0 + ϑ1 ln w jc + ϑ2 distijc + ϑ3 dist2ijc + ϑ4 gijc + ϑ5c Ic + ϑ6 X jc + ϑ7 Aijc + ε ijc

for the observations of annulus i in city j of country c. We measure the city wage w jc by
the per capita household net wage in the NUTS2 areas23 and the green urban areas gijc
by the land share of green urban areas. Given the language, cultural and administrative
barriers, we consider that individuals freely move across cities only within the same
country. Thus, the country utility level is captured by the vector of country dummies Ic .
Finally, vector X jc controls for observed city characteristics, such as elevation, rainfall
and temperature. Vector Aijc controls for observed amenities in each annulus, such as
the shares of sport leisure facilities and industrial lands and the shares of forest and
agricultural lands within a 100 m distance from the residential areas.
A potential endogeneity issue arises because the choices for residents’ land use and
planers’ green urban areas are intertwined. Indeed, urban planners are expected to or-
ganize green urban areas as a function of surrounding population densities and there-
fore residents’ land use. To control for such a reverse causality, we use the historical
23 Inthis text, city wages are measured by the incomes net of taxes at the NUTS2 level. Net incomes
closely reflect the budget constraints faced by residents in their land use choices. However, NUTS2 en-
compasses larger areas than the cover of many cities, which may downward bias city income values. In
Appendix C, we perform the same analysis with the production value at the NUTS3 level, which includes
taxes. The results are similar except the values should be interpreted differently.
56 CHAPTER 3. A WELFARE EVALUATION OF GREEN URBAN AREAS

level of urban green areas as indicative of the current levels. The main idea behind
using historical urban green area information is that once an urban green area is devel-
oped, it is rarely changed. In fact, many urban green areas in Europe were provided
decades ago and have remained intact. Examples are Hyde Park in London, created in
approximately the 16th century by Henry VIII and originally intended for hunting, and
the ‘Jardin du Luxembourg’, which was first built as a private garden of Queen Marie
de Medici in the early 17th century. Both private parks were later converted to public
green areas by public authorities. Thus, we can mitigate the reverse causality using
the data for old parks to predict the locations of current public green areas. Toward
this aim, we use the CORINE Land Cover 1990 database, which unfortunately does not
cover all GMES Urban Atlas countries. As a result, the regression results exclude cities
in the UK, Sweden and Finland.24 The details for the first-stage regression are reported
in Appendix C, which confirm that the historical levels of urban green areas are a good
predictor of the current levels.
The results are reported in Table 3.2. Columns (1) to (4) display OLS estimates with-
out instrument variables. In all columns, the coefficient estimates are consistent with
our model predictions: residents use larger land plots for smaller city income, larger
distance between residences and the CBD and smaller green urban areas. The results
are robust after controlling for country fixed effects, city geographical conditions, such
as elevation, rainfall and temperature (see Columns (2) and (3)), and different types of
amenities within annuli (see Column (4)). To account for the above endogeneity issue,
Columns (5) to (8) display the same estimates after instrumenting the share of urban
green areas with its historical value from the CORINE Land Cover 1990 database. The
IV regression reports slightly stronger effects of urban green areas on the residential slot
size than those of the OLS regressions, which is intuitive because the historical level of
urban green areas was lower than the current level. We also apply the Wu-Hausman
test for endogeneity (reported in Appendix A). The Wu-Hausman test coefficient is not
significant at the 90% confidence level, which supports the alternative hypothesis of
no endogeneity at the 90% confidence level and suggests that endogeneity may not
be a critical issue in our analysis. Both the OLS and IV results show significant coeffi-
cients for the share of green urban area amenities g for approximately 2.24 − 2.25 before
including the control and 1.69 − 1.86 after including all other controls. This finding im-
plies that, ceteris paribus, residents in annuli with no green urban areas use 14% more
24 CORINE Land Cover (CLC) 1990 does not cover the UK, Sweden and Finland. The database cov-
ers Austria, Belgium, Bulgaria, Croatia, the Czech Republic, Denmark, Estonia, France, Germany, Greece,
Hungary, Ireland, Italy, Latvia, Lithuania, Luxembourg, Malta, Montenegro, the Netherlands, Poland,
Portugal, Romania, Serbia, Slovakia, Slovenia, Spain, and Turkey, out of which 23 countries are in-
cluded in our data. For details, see CORINE Land Cover 1990 Metadata: https://land.copernicus.eu/
pan-european/corine-land-cover/clc-1990?tab=metadata (Accessed May 02, 2018). To our knowl-
edge, CLC 1990 is the oldest land use database that systematically covers all of European cities.
3.4. EMPIRICAL ANALYSIS AND WELFARE VALUATION 57

land than those residing in annuli with a 7% share of green urban area.25 Population
densities are reduced in the same proportions. According to this empirical estimation,
green urban areas are an important factor explaining the use of residential land and
population density.

3.4.2.2 Welfare valuation

In this subsection, we use the previous regression model to quantify the value of green
urban areas with a focus on the general equilibrium effect. We recover all parame-
ters of our theoretical model and run several counterfactual analyses. In particular, we
build counterfactuals where half of the green urban areas are deleted in every annulus
and are either left unused or converted to new residential land. We can then evaluate
the changes in the residential land use and consumption of goods, population density,
land rents and utility levels for each city. To express utility changes more intuitively,
we measure the cost to residents by their incentives to leave the city and compensating
variation wage (subsidy or tax) that they must receive to keep their utility levels. The
difference between the status quo wage level and the new compensating variation in-
come is precisely the willingness to pay for each household (Sieg et al. (2004)). By the
same token, we discuss the distribution of the effect of green urban areas between cities
and within them. We consider this analysis a useful exercise because it informs policy
makers about the impact of urban green areas on city structures and sizes.
We recover the model parameters from the estimated coefficient of residents’ land
use using the values of ϑ0 , ϑ1 , ϑ2 , ϑ3 and ϑ4 from Column (8) in Table 3.2. Country util-
ity levels are recovered from the parameters ϑ5c and the constant term ϑ0 . Our baseline
model and counterfactuals use the observed distance to the city center, city and country
characteristics and local (nongreen) amenities. The baseline model simulates the vari-
ables under study using those estimated parameters and the observed characteristics
(distance to CBD, wage, green urban areas, etc.). The counterfactual exercises investi-
gate the impact of canceling 50% of the urban green areas in each annulus of each city,
keeping the same observed characteristics. Counterfactual exercise 1 considers open
cities where utility levels and unobserved heterogeneity are maintained. This helps us
discuss a long-term and unregulated perspective, where urban planners do not impose
restrictions on workers’ mobility within and between cities. Counterfactual exercise
2 considers closed cities with exogenous city populations. The study of closed cities
can be appropriate in evaluating policy changes that occur simultaneously in all cities,
such as changes in EU policies.26 Here, our aim is to discuss a midterm or regulated
perspective, where urban planners are able to restrict workers’ mobility between cities
but allow residents’ land use to change. To give a relevant measure of utility change,
0
25 We compute sijc /si0 jc = eϑ3 ( gijc − gijc ) , with gijc = 0 and 0.07.
26 Cheshire and Shepard (2002) also use the closed city model to analyze the welfare effects of policy

changes in the UK.

CHAPTER 3. A WELFARE EVALUATION OF GREEN URBAN AREAS

Table 3.2: Residents’ land use

Dependent variable: Ln Residential Space
OLS IV
(1) (2) (3) (4) (5) (6) (7) (8)
Constant −0.507∗∗∗ −1.448∗∗∗ −1.517∗∗ −0.896 −0.505∗∗∗ −1.420∗∗∗ −1.478∗∗ −0.868
(0.143) (0.356) (0.631) (0.556) (0.145) (0.360) (0.635) (0.560)
Ln Household Income −0.074 −0.758∗∗∗ −0.784∗∗∗ −0.735∗∗∗ −0.074 −0.752∗∗∗ −0.778∗∗∗ −0.730∗∗∗
(0.071) (0.173) (0.167) (0.151) (0.072) (0.174) (0.168) (0.151)
Distance to CBD 1.502∗∗∗ 1.412∗∗∗ 1.455∗∗∗ 0.885∗∗∗ 1.497∗∗∗ 1.415∗∗∗ 1.448∗∗∗ 0.879∗∗∗
(0.237) (0.220) (0.212) (0.213) (0.237) (0.221) (0.212) (0.213)
Distance to CBD square −0.607∗∗∗ −0.549∗∗∗ −0.553∗∗∗ −0.314∗∗ −0.604∗∗∗ −0.556∗∗∗ −0.551∗∗∗ −0.312∗∗
(0.130) (0.130) (0.133) (0.129) (0.130) (0.130) (0.133) (0.127)
Share of Urban Green −2.241∗∗∗ −2.052∗∗∗ −1.941∗∗∗ −1.695∗∗∗ −2.253∗∗∗ −2.274∗∗∗ −2.153∗∗∗ −1.866∗∗∗
(0.400) (0.254) (0.251) (0.214) (0.431) (0.285) (0.273) (0.237)
Country FE No Yes Yes Yes No Yes Yes Yes
City Geographical Controls No No Yes Yes No No Yes Yes
Annuli Amenity Controls No No No Yes No No No Yes
Sample All All All All All All All All
Observations 10,853 10,853 10,853 10,853 10,853 10,853 10,853 10,853
Adjusted R2 0.207 0.468 0.483 0.579 0.197 0.464 0.480 0.576
Residual Std. Error 0.574 0.470 0.464 0.418 0.578 0.472 0.465 0.420
df 10,848 10,827 10,823 10,819 10,848 10,827 10,823 10,819
F Statistic 19.35∗∗∗ 26.02∗∗∗ 24.35∗∗∗ 34.07∗∗∗ 18.23∗∗∗ 24.60∗∗∗ 24.45∗∗∗ 34.67∗∗∗
Note: Here, we use the Household income taken from Regional Economic Accounts from Eurostat at NUTS2 level with adjustment to purchasing
power standard (PPS) as the proxy for city income level, and it is measured in e100,000. The distance to CBD is measured in 10 kilometres. The
inverse of residential space is calculated by dividing the number of inhabitants in each annuli with annulus areas net of its urban green space. We
exclude Cyrus and Luxembourg as the Eurostat population grid database does not cover Cyrus and the household income data at NUTS2 of Eurostat
does not cover Luxembourg. United Kingdom and Finland are also excluded as they are not covered by Corine Land Cover 1990. City boundary
is chosen at 20% cut-off point. For city control, we take into account the elevation, average rain fall, average temperature in Jan 01 and average
temperature in July 01 for period 1995-2010. City amenity controls include the share of industrial, sport and leisure land use as well as the share of
forest and agriculture land within 100 meters buffer from residential area. Standard errors are clustered at city level. The row ”df” reports the degree
of freedom. Significance levels are denoted by ∗ for p<0.1, ∗∗ for p<0.05 and ∗∗∗ for p<0.01.
58
3.4. EMPIRICAL ANALYSIS AND WELFARE VALUATION 59

Table 3.3: Counterfactual analysis: open and closed cities

Cities Composite Housing Income Residential Green

Goods (Z) Rent (R × s) (W) Area Area (GA)
number (e1000) (e1000) (e1000 ) (km2 ) (km2 )
Spatial heterogeneity 264 5.05 6.92 15.22 25.29 4.52
(1.87) (2.56) (5.87) (48.91) (9.87)
No spatial heterogeneity 264 5.03 6.89 15.22 25.29 4.52
(1.86) (2.55) (5.87) (48.91) (9.87)
Note: Household income is taken from Eurostat at NUTS2 level and is measured on purchasing power
standard (PPS) at 1000e. More details on PPS measure, please check Eurostat technical documents.
The standard deviation is reported in the parenthesis.

we also compute the compensating variation wage as the city wage that maintains the
baseline utility level when we remove green urban areas and the corresponding will-
ingness to pay. Counterfactual exercises 1 and 2 hinge on the assumption that empirical
model residuals reflect land heterogeneity that is unobserved to the econometricians
but observed and used by residents in their land plot size choices. Such heterogeneity
is reported in the counterfactual results. This assumption may be strong, as it imposes
strong information on behalf of residents. Therefore, we also take the opposite view in
counterfactuals 3 and 4 where the residuals consist of measurement errors that can be
observed neither by the econometricians nor the residents. The details of the counter-
factual analysis are relegated to Appendix B.
The results of the baseline model are displayed in Table 3.3. The first column reports
the number of cities in the baseline exercise. Every other column reports the average
and standard deviation over the city averages imputed from the baseline model. The
second and third columns show the consumption of composite goods and housing by
households, while the fourth column displays the net income. The difference between
this column and the sum of the two previous columns accounts for commuting costs.
On average, individuals have e15,220 as disposable income and spend e6,920 for hous-
ing expenses, which accounts for approximately 45% of their net income. Such a value
is slightly above the average housing costs in European cities, which are approximately
a quarter of the household income for both European rural and urban areas. The lit-
erature reports a range between 18% and 32%, with higher levels for urban areas and
renters (Fahey et al, 2004, Davis and Ortalo-Magne, 2011). Our model differs from this
literature because we do not take into account housing furniture and maintenance (5%
of housing costs in Eurostats, 2015), we consider city cores, which have more expen-
sive housing locations, and finally, we do not model the construction process. The last
two columns of Table 3.3 report the average residential area and green urban area across
cities, the latter being approximately one-fifth of the former. The rows address the cases
when we consider the regression model error as a spatial amenity for the residents (spa-
tial heterogeneity) and when we do not (no heterogeneity). The difference between the
two cases is not large.
60 CHAPTER 3. A WELFARE EVALUATION OF GREEN URBAN AREAS

Our main results are displayed in Table 3.4, which shows the baseline model (first
row) and the counterfactual exercises (other rows) for cases with and without spatial
heterogeneity. The table structure is the same as that of Table 3.3. Every column re-
ports the average and standard deviation of the city averages over annuli imputed in
the baseline and counterfactual exercises. Consider the first row, which presents our
baseline model and permits comparison with the literature. The first five columns dis-
play the imputed residential surfaces, land rents on units of residential plots and green
urban areas, population and relative utility. The average city size of 0.31 million in-
habitants is consistent with the statistics that most European cities are medium-sized
(European Commission and UN-Habitat (2017), Urban Audit, Eurostat). Residents’ av-
erage use of space is approximately 95 m2 ; the measure is reasonable given that we
consider the core of the most populous cities in the EU, which are the densest ar-
eas of the most urbanized parts of the EU.27 The land rent per square meter is 93.41
e/m2 /year on average for all cities. On average, the land values of green urban areas
(102.32 e/m2 /year) are higher than residential land prices. These values are imputed
from the residential land price associated with each annulus. Because urban green ar-
eas are concentrated at close and intermediate distances to CBDs, they are surrounded
by more expensive residential land plots.
Consider, now, counterfactual exercise 1, where one removes 50% of the green areas
in every annulus of open cities. In open city systems, the utility of city inhabitants
is exogenous, but the change in green amenities affects the urban structures. Suppose,
furthermore, the case where the removed land is not (re)used, as indicated in the second
row of Table 3.4. To keep the same utility level, residents must compensate for the
decrease in urban green amenities by larger residential plots, which implies that a share
of the population must migrate out of the city. On average, city residents increase their
land use from 95 to 100 m2 (a rise of 5.3%), and the city population falls from 0.31 to
0.29 million (a loss of 6.5%). Land rents fall from 93.41 to 87.11 e per m2 and year (a
fall of 6.74%). We compute the total loss in the land market to be approximately e147
million for an average city.28
Suppose now that 50% of green urban areas are converted into residential land, as
shown in the third row of Table 3.4. In an average city, there is a new land supply of 2.26
km2 (half of 4.52 km2 ) on top of the baseline residential land supply of 25.29 km2 , an
increase of 8.9% (see Table 3.3). This increase is slightly more than the 5.3% space com-
pensation that residents required without land conversion. As a result, the additional
land supply attracts new city dwellers, and the city population rises to 0.32 million on
average. Residential land rents rise slightly to 87.64 e per m2 and year because the new
27 They are more densely populated than US cities.
28 Toestimate the total land rent loss, we multiply the city residential area of each annulus with the per-
square-meter land rent loss between the baseline and counterfactual models, aggregate over the city, and
compute the average over all cities.
3.4. EMPIRICAL ANALYSIS AND WELFARE VALUATION 61

land is supplied at more central locations with higher values. We compute that, com-
pared to the baseline model, the housing market increases its total value by nearly e55
million per year and city when we take into account converted areas.
Counterfactual exercise 2 allows us to discuss the impact of reducing urban green
areas by half in closed cities where city planners prohibit migration. As predicted by
our theoretical model, when there is no conversion of land, the utility of all residents
decreases once we reduce the level of urban green amenities. Specifically, the average
utility decreases from 1 to 0.94. Residential land rents decrease only by a small amount
from 93.41 e/m2 /year to 93.36e/m2 /year, providing a total loss of e1.82 million per
year. The decrease in utility combined with the change in land rent requires an increase
in the baseline annual net income of e15, 200 to the compensating-variation income of
e16, 650, an increase of e1, 430 (9.4%). This is also the measure of our general equilib-
rium WTP as indicated in section 3.3 for canceling half of the urban green areas without
land conversion in the city area. Multiplying this figure by the city population, we ob-
tain a subsidy of e580 million for an average city. For reference, the general equilibrium
WTP for closing all publicly accessible open spaces in Cheshire and Sheppard (2002) is
£2,424 per household per year for the city of Reading, UK. Our average measure is not
too far in magnitude given the 50% closing of green urban areas in our samples.
Suppose now that green urban areas are converted into residential land. Then, the
residential land supply increases, land rents decreases, and residents can use more land
to compensate for the loss of green area amenities. Specifically, the land rents drop by
9.6% to 84.47 e/m2 /year, and the total loss in housing market increases to e3.45 million
per year. However, city residents enjoy both lower land rent and larger residential land
plots, which increase their average utility level. They obtain a slightly higher average
utility level (increase by 3%) and require a smaller compensating-variation wage of
e14, 570 per year to maintain their level of utility, which is equivalent to an income
reduction of e650 per year. Those results are also in line with those of Cheshire and
Sheppard (2002), who also found that the net effect of land use planning is negative
and that allowing more residential developments is welfare improving.
The bottom panel of Table 3.4 replicates the above analysis when we replace the
assumption of unobserved spatial heterogeneity by that of measurement errors. It can
be seen that most effects are similar. The main differences lie in the level of averages
and standard deviations of our variables of interest. On the one hand, the standard
deviations are naturally smaller because residents are no longer assumed to consider
spatial amenity variations in their choices. However, the reduction in the standard
deviation is not drastic, which shows that the model is already well explained by the
independent variables of the regression model. On the other hand, the assumption of
measurement errors also alters the average values of imputed variables. For instance,
in the baseline model of the bottom panel, the residential populations, land uses and
land rents are smaller than in the baseline model of the top panel. This is because
 Table 3.4: City structure under closed and open scenarios g = 0.5 × g0
CHAPTER 3. A WELFARE EVALUATION OF GREEN URBAN AREAS

Total loss in
Cities s Land Rent Green value Pop. U/Ū b Comp. W housing market wage comp. pop.
(number) (100m2/hab) (e/m2 /y) (e/m2 /y) (mil.) (e1000) (e mil.) ( e mil.) 1, 000hab.
Baseline 0 (spatial heterogeneity) 264 0.95 93.41 102.32 0.31 1.00
(0.44) (68.95) (75.71) (0.69) (0.36)
g = 0.5 × g0
Counterfactual 1: Open City
Urban green conversion No 264 1.00 87.11 92.73 0.29 1.00 146.61 0 20.38
(0.46) (63.11) (66.70) (0.64) (0.36) (407.11) (52.53)
Yes 264 1.00 87.64 92.73 0.32 1.00 −54.74 0 -8.24
(0.46) (63.46) (66.70) (0.71) (0.36) (126.62) (18.39)
Counterfactual 2: Closed City
Urban green conversion No 264 0.95 93.36 99.38 0.29 0.94 16.65 1.82 580.56 0
(0.44) (69.00) (72.98) (0.64) (0.34) (6.46) (16.97) (1926.01)
Yes 264 1.03 84.47 89.26 0.32 1.03 14.57 3.45 −198.56 0
(0.47) (60.75) (63.28) (0.71) (0.38) (5.71) (15.30) (544.27)
Baseline 0 (no spatial heterogeneity) 264 0.88 93.70 100.81 0.28 1.00
(0.34) (59.34) (61.55) (0.50) (0.36)
g = 0.5 × g0
Counterfactual 3: Open City
Urban Green Conversion No 264 0.93 87.37 91.53 0.29 1.00 130.20 0 18.13
(0.35) (54.04) (54.44) (0.64) (0.36) (311.97) (41.07)
Yes 264 0.93 87.60 91.53 0.32 1.00 −50.81 0 -7.58
(0.35) (54.03) (54.44) (0.71) 0.36) (115.20) (16.41)
Counterfactual 4: Closed City
Urban Green Conversion No 264 0.88 93.64 98.01 0.29 0.94 16.65 2.26 511.77 0
(0.34) (59.40) (59.64) (0.64) (0.34) (6.47) (13.72) (1483.03)
Yes 264 0.96 84.72 88.58 0.32 1.03 14.61 2.88 −183.74 0
(0.36) (51.84) (52.31) (0.71) (0.37) (5.70) (11.47) (484.95)
Note: Household income is taken from Eurostat at NUTS2 level and is measured on purchasing power standard (PPS) at 1000e. More details on PPS measure, please check Eurostat technical documents. Population is calculated in
1/φ̂
Ujc Ujc u jc
million inhabitants. Utility is measured relatively to baseline average. Basically, we measure = = 1/φ̂ where u jc is calculated as in the above section. The standard deviation is reported in
Ū baseline 1 1 
∑ U baseline ∑ ubaseline
N jc jc N jc c
62

the parenthesis.
3.4. EMPIRICAL ANALYSIS AND WELFARE VALUATION 63

those variables are nonlinear functions of the error term ε under the Cobb-Douglas
preferences.29
What type of cities are more sensitive to removing green urban areas? Where in
the city are the changes more important? To answer these questions, we compare the
impact of removing or converting green urban areas between cities of different incomes
and population sizes as well as between within-city locations at different distances to
the CBD. Toward this aim, Table 3.5 reports the baseline wage (first row), the changes
in the compensating-variation wages to sustain constant utility (next four rows), the
baseline land rent to landlords (sixth row), and the landlords’ losses (last four rows)
when we group cities by income quartiles (first four columns), by population size quar-
tiles (next four columns) and by quartiles of relative distances to the CBD (last four
columns). The positive changes in compensating-variation wages can be interpreted as
subsidies required to maintain the residents at their baseline equilibrium levels. Table
3.5 presents the results for open and closed cities with and without the conversion of
green urban areas to residential land. All figures are aggregated from the same coun-
terfactual exercise with unobserved spatial heterogeneity and with the 50% reduction
in the green urban areas presented in Table 3.3 and 3.4.
Let us consider first the case of open cites (counterfactual exercise 1). The reduction
in green urban area amenities harms residents who partly leave the city. Because of free
migration, city residents keep the same utility as in the countryside and ask for no com-
pensation to stay in cities, which is why the second and third rows in Table 3.5 display a
set of zeros for the change in compensating-variation wages. By contrast, landlords lose
money. If urban green areas are not converted to residential land, they lose e3.27 and
e8.22 per m2 and year in cities belonging to the bottom and upper income quartiles,
respectively (see seventh row). Similarly, they lose e5.21 and e7.90 per m2 and year in
cities belonging to the bottom and upper population quartiles, respectively. This result
is explained by the fact that land value, city size and income are positively correlated.
Landlords also lose e11.24 per m2 and year in the central city quartile but only e3.89
per m2 and year in the city periphery quartile, indicating that land rents decrease with
distance from the CBD. This pattern remains approximately the same if urban green
areas are converted to residential land (eighth row). In this case, the above figures
decrease by approximately e1 per m2 and year in the lowest income and population
size cities but only slightly for those with the highest income and population size. The
conversion of green urban areas mitigates the conclusions only to a small extent.
Let us now consider closed cities in which migration is restricted and half of the
green urban areas is removed (conterfactual exercise 2). Suppose, initially, that there is
no land conversion (fourth row in Table 3.5). To stay in the city, residents require an
29 The expression of land useŝ(w, r, g, u, e) includes a term eε that is a convex function of ε. Similarly, the
expression for the population density 1/ŝ(w, r, g, u, ε) includes a term e−ε , which is also a convex function
of ε.
CHAPTER 3. A WELFARE EVALUATION OF GREEN URBAN AREAS

Table 3.5: Counterfactual analysis: between- and within-city statistics

Between cities Within cities
City Income Quartiles City Population Size Quartiles Distance to CBD Quartiles
0-0.25 0.25-0.50 0.50-0.75 0.75-1 0-0.25 0.25-0.50 0.50-0.75 0.75-1 0-0.25 0.25-0.50 0.50-0.75 0.75-1
Baseline wage (e1000/hab/y) 8.19 12.27 17.71 22.93 12.36 14.97 15.85 17.72 16.71 16.62 16.64 16.54
(2.03) (1.23) (1.16) (2.23) (4.47) (5.91) (6.19) (5.57) (5.92) (5.91) (5.92) (5.92)
WTP (e1000/hab/y)
Counterfactual 1: Open City
Urban green conversion No 0 0 0 0 0 0 0 0 0 0 0 0
Yes 0 0 0 0 0 0 0 0 0 0 0 0
Counterfactual 2: Closed City
Urban green conversion No −0.77 −1.18 −1.67 −2.09 −1.28 −1.37 −1.33 −1.71 −1.60 −1.85 −1.47 −1.12
(0.54) (1.12) (0.97) (0.92) (1.29) (0.89) (0.78) (1.09) (1.75) (1.53) (1.27) (1.31)
Yes 0.36 0.66 0.71 0.90 0.83 0.64 0.51 0.64 0.29 0.43 0.61 0.92
(0.41) (0.84) (0.77) (0.54) (1.04) (0.60) (0.42) (0.49) (0.83) (0.67) (0.63) (1.03)
Land rent baseline (e/m2 /y) 49.73 94.59 104.09 126.23 73.45 85.55 100.08 114.56 176.38 114.34 80.06 63.85
(25.29) (56.73) (93.81) (57.47) (43.00) (52.39) (73.99) (90.54) (131.70) (85.43) (66.05) (63.93)
Landlord loss (e/m2 /y)
Counterfactual 1: Open City
Urban green conversion No 3.27 6.30 7.49 8.22 5.21 5.53 6.54 7.90 11.24 8.69 5.23 3.89
(2.95) (6.21) (9.40) (7.31) (5.86) (4.79) (8.09) (8.80) (16.55) (10.29) (7.55) (9.75)
Yes 2.97 5.35 6.79 8.09 4.01 5.02 6.22 7.83 11.24 8.69 5.23 3.89
(2.69) (5.20) (8.57) (8.41) (4.43) (4.35) (7.41) (9.58) (16.55) (10.29) (7.55) (9.75)
Counterfactual 2: Closed City
Urban green conversion No 0.05 0.08 0.04 0.01 0.01 0.03 0.06 0.09 0.08 1.27 −0.06 −0.45
(0.08) (0.17) (0.21) (0.43) (0.12) (0.18) (0.16) (0.43) (14.95) (7.05) (3.34) (5.14)
Yes 4.69 9.04 10.51 11.67 8.43 7.87 8.83 10.73 15.47 11.71 7.49 5.92
(5.15) (10.02) (13.61) (10.99) (10.58) (7.24) (8.83) (10.73) (19.39) (13.30) (10.42) (13.85)
Note: Here, we report only the case with heterogeneity; for non-heterogeneity case (like in counterfactual exercises 3 and 4), the results are similar. The loss for landlord is
accounted as the difference between new land rent equilibrium and land rent in baseline model 0. The standard deviations are reported in the parenthesis.
64
3.4. EMPIRICAL ANALYSIS AND WELFARE VALUATION 65

increase in compensating-variation wages (negative willingness to pay), or a subsidy,

of e770 per year for the bottom city income quartile and e2, 090 per year for the top
quartile. This increase represents up to 9.1% and 9.6% of the baseline net incomes.
These subsidies also increase with city population. One can check that larger cities
require proportionally higher subsidies, which result from the higher losses incurred
by the residents in larger cities. The subsidy is not monotonic with distance to the city
center: it first increases from e1, 600 to e1, 850 per year when one moves from the first
to the second distance quartiles and then drops to e1, 120 for the last distance quartile.
This pattern reflects the geographical distribution of the share of green urban areas. It
can finally be seen that landlords are not substantially harmed by the reduction in green
urban areas when cities are closed and land is not converted (ninth row).
Finally, suppose that the green urban areas are converted to residential land (fifth
row), which increases the residential land supply and compensates residents for the
lack of green area amenities. The negative changes in the compensating-variation wages
indicate that residents are better off in this situation. Low-income cities would accept
lower compensating-variation wages and would therefore pay a tax of e360 per year
in the lowest city income quartiles and e900 per year in the highest. As shown in the
table, this benefit is larger for peripheral residents. Finally, landlords are negatively
affected by the additional supply of residential land (see tenth row). They are more
impacted in the richest and the largest cities and at the most central locations.

3.4.3 Estimation and valuation with hyperbolic preferences

Similar to the case of Cobb-Douglas preference, we suppose the presence of observable

heterogeneity χ in the preference for land plots or specific characteristics of locations
as well as unobservable heterogeneity or measurement errors e. The hyperbolic utility
function becomes
θ
U (s, a, z, χ, e) = z − +a+χ+e
2s
Accordingly, we obtain the consumption b s = θ/ (y + a + χ + e − u), b
z = (y − a − χ − e + u) /2
2
and land value Ψ = (y + a + χ + e − u) /(2θ ). We focus on the case of positive con-
sumption so that y − a − χ − e + u > 0.
We assume that the transport cost is given by t = w × τ (r ), where τ (r ) is a function
of distance to the CBD. For simplicity, we assume the quadratic form τ (r ) = τ1 × r +
τ2 × r2 . Green area amenities are given by a = αx, which can be written as a function
of green urban areas g as a = αg/β, where α and β are green amenity and land use
intensity parameters, respectively. We can further standardize the amenity value α = 1
and consider β as land use intensity to provide one unit for green amenities. We obtain
the inverse of housing slot size in equilibrium as given below:
 
1 1 α
= w − w × τ (r ) + g − u + χ + e (3.21)
s
b θ β
66 CHAPTER 3. A WELFARE EVALUATION OF GREEN URBAN AREAS

3.4.3.1 Estimation

We run the following regression:

1

2

= ϑ0 + ϑ1 wic + ϑ2 rijc × wic + ϑ3 rijc × wic + ϑ4 gijc + ϑ5 Ic + ϑ6 X jc + ϑ7 Aijc + eijc
sijc
where i is the location of the annuli within the city border, j is the city and c is the coun-
try where the city belongs to. Here, we use household income from NUTS2 regions as
the proxy for the wage level; rijc is the distance of annuli i of city j to its CBD. Ic , X jc and
Aijc are the vectors of country dummies, city geographic controls and annuli amenity
controls variables, respectively. We introduce the nonlinearity in transportation cost by
adding the quadratic term as explained in Cobb-Douglas preference sections.
Table 3.6 reports the regression results from both OLS and IV regressions with differ-
ent levels of controls. We utilize a similar IV strategy as in the Cobb-Douglas case and
the first stage regression, as well as the Wu-Hausman test for endogeneity, as reported
in Appendix A. Columns (1) to (4) display OLS estimates without instrument variables.
In all columns, the coefficient estimates are consistent with our model predictions: resi-
dents use larger land plots for smaller city income, a larger distance between residences
and the CBD and smaller green urban areas. The results are robust after controlling for
country fixed effects, city geographical conditions, such as elevation, rainfall and tem-
perature (see Columns (2) and (3)) and different types of amenities within annuli (see
Column (4)). After controlling for all other amenities, the square of the distance to the
CBD loses its significance, suggesting the linearity in transportation cost. Columns (5)
to (8) display the same estimates after instrumenting the share of urban green areas
with its historical value from the CORINE Land Cover 1990 database. The IV regres-
sion reports a similar magnitude of the effects of urban green areas on the residential
slot size. The Wu-Hausman test coefficient (reported in Appendix A) is not significant
at the 90% confidence level, which supports the alternative hypothesis of no endogene-
ity at the 90% confidence level and implies that endogeneity may not be a severe issue
in our analysis. Both the OLS and IV results show significant coefficients for the share
of green urban area amenities g for approximately 2.9 − 3.2 before including the control
and 2.4 − 2.7 after including all other controls. This finding implies that, ceteris paribus,
residents in annuli with no green urban areas use approximately 19 − 22% more land
than those residing in annuli with a 7% share of green urban area.30 Those changes are
slightly higher than those in the Cobb-Douglas preference (approximately 14%).

3.4.3.2 Welfare valuation

In this section, we examine the changes in city structure and land rent when we cancel
50% of green urban space within the city boundary. We proceed with similar steps as
30 We 1
− 1 0 ), with g
= ϑ4 ( gijc − gijc 0
compute sijc 0
sijc ijc = 0.07 and gijc = 0 and on average sijc is at 0.95

(100 m2 /hab.).
 Table 3.6: Hyperbolic residents’ land use regression results

Dependent variable: Inverse of space

OLS IV
(1) (2) (3) (4) (5) (6) (7) (8)
Household Income 3.149∗∗∗ 9.908∗∗∗ 10.557∗∗∗ 9.145∗∗∗ 3.171∗∗∗ 9.842∗∗∗ 10.507∗∗∗ 9.096∗∗∗
(0.813) (2.011) (1.968) (1.785) (0.822) (2.030) (1.975) (1.791)
Distance to CBD × Income −10.202∗∗∗ −9.302∗∗∗ −9.716∗∗∗ −5.323∗∗∗ −10.200∗∗∗ −9.263∗∗∗ −9.693∗∗∗ −5.302∗∗∗
(1.783) (1.713) (1.714) (1.626) (1.780) (1.707) (1.704) (1.609)
Distance to CBD square × Income 4.344∗∗∗ 3.645∗∗∗ 3.718∗∗∗ 1.771 4.333∗∗∗ 3.635∗∗∗ 3.716∗∗∗ 1.767
(1.162) (1.346) (1.414) (1.288) (1.142) (1.333) (1.400) (1.264)
Share of Urban Green 3.190∗∗∗ 2.930∗∗∗ 2.749∗∗∗ 2.463∗∗∗ 2.930∗∗∗ 3.228∗∗∗ 3.034∗∗∗ 2.727∗∗∗
(0.701) (0.405) (0.414) (0.384) (0.735) (0.467) (0.461) (0.435)
Country FE No Yes Yes Yes No Yes Yes Yes
City Geographical Controls No No Yes Yes No No Yes Yes
Annuli Amenity Controls No No No Yes No No No Yes
Sample All All All All All All All All
Observations 10,853 10,853 10,853 10,853 10,853 10,853 10,853 10,853
Adjusted R2 0.131 0.455 0.474 0.550 0.118 0.452 0.471 0.548
Residual Std. Error 0.946 0.749 0.736 0.680 0.953 0.752 0.738 0.682
df 10,848 10,827 10,823 10,819 10,848 10,827 10,823 10,819
F Statistic 33.38∗∗∗ 24.61∗∗∗ 23.96∗∗∗ 36.54∗∗∗ 37.55∗∗∗ 24.63∗∗∗ 23.34∗∗∗ 36.37∗∗∗
3.4. EMPIRICAL ANALYSIS AND WELFARE VALUATION

Note: Here, we use the Household income taken from Regional Economic Accounts from Eurostat at NUTS2 level with adjustment to purchasing power
standard (PPS) as the proxy for city income level, and it is measured in e100,000. The distance to CBD is measured in 10 kilometres. The inverse of space is
calculated by divided number of inhabitants in each annuli with the areas within the annuli minus the areas using as urban green (in 100 meters). We exclude
Cyrus and Luxembourg as the Eurostat population grid database does not cover Cyrus and the household income data at NUTS2 of Eurostat does not cover
Luxembourg. United Kingdom and Finland are also excluded as they are not covered by Corine Land Cover 1990. City boundary is chosen at 20% cut-off
point. For city control, we take into account the elevation, average rain fall, average temperature in Jan 01 and average temperature in July 01 for period
1995-2010. City amenity controls include the share of industrial, sport and leisure land use as well as the share of forest and agriculture land within 100 meters
buffer from residential area. Standard errors are clustered at city level. The row ”df” reports the degree of freedom. Significance levels are denoted by ∗ for
p<0.1, ∗∗ for p<0.05 and ∗∗∗ for p<0.01.
67
68 CHAPTER 3. A WELFARE EVALUATION OF GREEN URBAN AREAS

in the Cobb-Douglas preference. The first results of the baseline model with the hyper-
bolic preference are displayed in Table 3.7, and the main results for the counterfactual
exercises are displayed in Table 3.8. On average, individuals have e15,220 as dispos-
able income and spend e8,120 for housing expenses, which is slightly more expensive
than the case of the Cobb-Douglas preference and that in reality. The average city size,
average residential space and average population are similar to the case of the Cobb-
Douglas preference. The only difference is in the computed land rent. The land rent per
square meter is 143 e/m2 /year on average for all cities. These values are higher than in
the case of Cobb-Douglas utility, which exhibits a higher income elasticity of demand
residential lands in the hyperbolic preference compared to those in the Cobb-Douglas
preference.

Table 3.7: Hyperbolic counterfactual analysis: open and closed cities

Cities Composite Housing Income Residential Green

Goods (Z) Rent (R × s) (W) Area Area (GA)
number (e1000) (e1000) (e1000 ) (km2 ) (km2 )
With heterogeneity 264 5.22 8.12 15.22 25.29 4.52
(6.86) (4.60) (5.87) (48.91) (9.87)
Without heterogeneity 264 5.01 8.32 15.22 26.55 4.52
(6.01) (3.64) (5.87) (48.91) (9.87)
Note: The standard deviation is reported in the parenthesis. Household income is taken from Eurostat
at NUTS2 level and is measured on purchasing power standard (PPS) at 1000e.

Examine again the counterfactual exercise 1, where one eliminates 50% of the green
areas in every annulus of open cities. On average, city residents increase their resi-
dential land use by approximately 10%, and the city population falls by approximately
9.5%. Those results are congruent under both hyperbolic and Cobb-Douglas prefer-
ences with a slightly higher magnitude in the hyperbolic preference. Nevertheless, the
land values decline at higher magnitude under the hyperbolic preference, which is at
9% (from 143 to 129 e per m2 per year). This situation results in a larger loss in the hous-
ing market of roughly e322 million for an average city. First, it is because of the higher
sensitivity in the bid rent in the case of the hyperbolic; we have a higher magnitude for
the change in residential land uses and larger falls in land price (9% instead of 6% in
Cobb-Douglas). Second, in the case of the hyperbolic preference, the consumption of
nonhousing goods also changed under new equilibrium, as shown in subsection 3.3.2,
leading to a smaller budget for housing consumption under the new equilibrium. In
the case of land conversion into residential land use31 , the residential land rents rise
slightly to 131 e per m2 per year because the new land is supplied at more central lo-
cations with higher land values. We compute that, compared to the baseline model,
the housing market increases its total value by approximately e22 million per year and
31 Inan average city, there is a new land supply of 2.26 km2 (half of 4.52 km2 ) on top of the baseline
residential land supply of 25.29 km2 , an increase of 8.9%.
3.4. EMPIRICAL ANALYSIS AND WELFARE VALUATION 69

city when we take into account converted areas. In contrast, Cobb-Douglas preference
results in an increase in the total value of the housing market by e54 million.
In the counterfactual exercise 2 with closed cities, the average utility shrinks from
1 to 0.84 with no conversion policy. Total residential land rents decline by a smaller
amount of e19 million per year. This reduction in utility combined with the change
in land rent requires an increase in the baseline annual net income of e970 (6.4%) per
year per household, which is the WTP measure of an average household. This measure
is slightly lower than that of the case of Cobb-Douglas at 9.4%. Suppose now that
green urban areas are converted into residential land. Then, the residential land supply
increases, land rents decrease, and residents can use more land to compensate for the
loss of green area amenities. Specifically, the land rents drop by 17% to 118 e/m2 /year,
and the total loss in the housing market rises to e236 million per year. Nevertheless, city
residents enjoy both lower land rent and larger residential land plots, which enhance
their average utility level. They reach a slightly higher average utility level (jump by
8%) and need a smaller compensating-variation wage of e14, 730 per year to maintain
their level of utility, which is equivalent to an income reduction of e490 per year, or
equivalent to 3.2% of their annual income. Those results are also in line with the Cobb-
Douglas preference, which also found that the net effect of land use planning is negative
once we allow for land conversion into residential land.
To summarize, the effects of population reduction under open cities, the measures of
compensating wage variations, and the corresponding WTP for an average household
are robust under two preference specifications. The land prices computed are costlier
under the hyperbolic preference than those under the Cobb-Douglas preference.
Table 3.9 summarizes the results across cities with different sizes and income levels,
as well as at different distances to the CBD within-cities for the hyperbolic preference.
In the case of open city systems, the decline in green urban area amenities significantly
harms residents who partly leave the city, which is robust in both utility functions. If
urban green areas are not converted to residential land, the landlords lose e12 and
e16 per m2 and year in cities belonging to the bottom- and upper-income quartiles,
respectively (see the seventh row). Furthermore, they lose e12 and e17 per m2 and
year in cities belonging to the bottom and upper population quartiles, respectively.
The pattern is similar to the case of Cobb-Douglas. However, the absolute increase is
higher as land demand is more sensitive in the hyperbolic preference. Landlords also
lose e21 per m2 and year in the central city quartile but only e9 per m2 and year in
the city periphery quartile, indicating that land rents decrease with distance from the
CBD. This pattern remains approximately the same if urban green areas are converted
to residential land (eighth row), but with much smaller magnitude.
In the case of closed cities, residents require an increase in compensating-variation
wages (negative willingness to pay), or a subsidy, of e930 per year for the bottom city
income quartile and e980 per year for the top quartile to have the same level of utility as
CHAPTER 3. A WELFARE EVALUATION OF GREEN URBAN AREAS

Table 3.8: City structure with Hyperbolic preference under closed and open scenarios g = 0.5 × g0
Total loss in
Cities s Land Rent Green value Pop. U/Ū b Comp. W housing market wage comp. population
(number) (100m2/hab) (e/m2 /y) (e/m2 /y) (mil.) (e1000) (e mil.) ( e mil.) 1, 000hab.
Baseline 0 (with heterogeneity) 264 0.96 143.14 172.33 0.31 1.00
(0.45) (200.59) (237.25) (0.68) (1.22)
g = 0.5 × g0
Counterfactual 1: Open City
Urban green conversion No 264 1.06 129.08 150.71 0.28 1.00 322.08 0 21.55
(0.53) (189.71) (219.11) (0.64) (1.22) (835.63) (47.88)
Yes 264 1.05 131.74 150.71 0.31 1.00 22.99 0 −6.60
(0.52) (193.30) (219.11) (0.70) (1.22) (479.24) (34.20)
Counterfactual 2: Closed City
Urban green conversion No 264 0.96 142.38 161.94 0.31 0.84 16.19 19.45 315.31 0
(0.45) (200.15) (230.77) (0.68) (1.22) (5.93) (72.94) (845.57)
Yes 264 1.05 118.70 136.14 0.31 1.08 14.73 236.78 −202.02 0
(0.48) (161.82) (182.87) (0.68) (1.13) (6.07) (850.99) (1130.31)
Note: Household income is taken from Eurostat at NUTS2 level and is measured on purchasing power standard (PPS) at 1000e. Population is calculated in million inhabitants. The standard deviation
is reported in the parenthesis.
Table 3.9: Quantiles analysis: between- and within-city statistics (Hyperbolic preference)
Between cities Within cities
City Income Quartiles City Size Quartiles Distance to CBD Quartiles
0-0.25 0.25-0.50 0.50-0.75 0.75-1 0-0.25 0.25-0.50 0.50-0.75 0.75-1 0-0.25 0.25-0.50 0.50-0.75 0.75-1
WTP (e1000/hab/y)
Urban green conversion No −0.93 −0.99 −0.99 −0.98 −1.05 −0.94 −0.89 −0.99 −1.06 −1.19 −0.95 −0.72
(0.53) (0.91) (0.58) (0.44) (0.96) (0.48) (0.41) (0.57) (1.04) (0.85) (0.73) (0.73)
Yes 0.35 0.84 0.47 0.31 0.60 0.28 0.47 0.61 0.71 0.41 0.23 0.48
(0.84) (1.84) (1.51) (0.97) (1.95) (0.83) (1.15) (1.27) (2.39) (1.53) (1.08) (2.20)
Landlord loss (e/m2 /y)
Counterfactual 1: Open City
Urban green conversion No 12.36 16.57 14.25 13.19 12.95 11.78 14.37 17.13 21.23 18.75 12.21 9.33
(9.19) (16.71) (16.12) (14.49) (8.09) (9.64) (14.42) (17.52) (29.09) (21.70) (16.19) (21.23)
Yes 10.60 11.50 11.09 12.50 7.09 9.45 12.83 16.22 21.23 18.75 12.21 9.33
(7.17) (19.21) (13.90) (18.36) (18.59) (8.82) (11.14) (18.92) (29.09) (21.70) (16.19) (21.23)
Counterfactual 2: Closed City
Urban green conversion No 0.67 1.28 0.73 0.38 0.80 0.61 0.89 0.75 2.99 4.10 0.42 −1.14
(1.12) (2.61) (1.56) (1.02) (2.67) (1.03) (1.51) (1.23) (26.17) (15.22) (7.37) (11.38)
Yes 17.42 31.53 28.52 20.40 19.14 16.29 26.59 35.73 30.84 29.92 21.85 19.69
(22.00) (44.78) (57.64) (50.00) (35.36) (23.67) (45.85) (65.69) (53.56) (53.71) (47.24) (56.86)
Note: Here, we report only the case with heterogeneity; for non-heterogeneity case (like in counterfactual exercises 3 and 4), the results are similar. The loss for landlord is accounted as the difference
between new land rent equilibrium and land rent in baseline model 0. The standard deviations are reported in the parenthesis.
70
3.5. CONCLUSION 71

before. There is a much smaller difference in WTP between different income and pop-
ulation quantiles in the hyperbolic preference compared to those in the Cobb-Douglas
preference. The reason is because the difference in the measures of the general equi-
librium willingness to pay in the case of Cobb-Douglas is proportional to income level,
whereas is not proportional in the case of the hyperbolic preference (for details, see Ap-
pendix B). The subsidy in the case of the hyperbolic preference is not monotonic with
distance to the city center, which is in line with the Cobb-Douglas preference; it first
increases from e1, 060 to e1, 190 per year when one moves from the first to the second
distance quartiles and then drops to e720 for the last distance quartile. This pattern
reflects the geographical distribution of the share of green urban areas. It can finally be
seen that landlords are not substantially harmed by the reduction in green urban areas
when cities are closed, and the land is not converted (ninth row), which is robust under
both preferences.

3.5 Conclusion
This paper studies the effects of land-intensive public green areas on city structures un-
der competitive land markets. We employ monocentric urban land market frameworks
with two different classes of utility functions—Cobb-Douglas and hyperbolic—to for-
malize the effects in residential land uses and households’ willingness to pay (WTP)
from a substantial change in green urban areas. The model emphasizes the importance
of incorporating the adjustments in housing markets under a nonmarginal change in
urban green areas in measuring households’ WTP.
We first compute the equilibrium in residential land uses in the theoretical model
and utilize the geographical land covers combined with actual population density to
estimate the model parameters. The level of green areas in the past is used to instru-
ment for the current level to control for the potential endogeneity in green provision.
To quantify the value of green urban areas, we present a set of counterfactual exercises,
where half of the green urban areas are removed. In case of the Cobb-Douglas prefer-
ence, we estimate that, on average, open cities lose more than 6.5% of their population
and that landlords lose e147 million in each city and year if the green urban areas are
not converted into residential land. Such a loss in population is robust for both the hy-
perbolic and Cobb-Douglas preferences, while the impact for the land market is higher
in the hyperbolic case. If the erased green areas are converted into new residential land,
the increase in total residential lands is sufficient to compensate locals with additional
residential space and to attract new city dwellers. Compared to our baseline model,
for the Cobb-Douglas preference, the housing market increases its total value by nearly
e55 million per year and city.
In closed cities where the green urban areas are not converted, the willingness to
pay to avoid such a policy is measured at e1, 430 per person and year. The estimated
72 CHAPTER 3. A WELFARE EVALUATION OF GREEN URBAN AREAS

results for the hyperbolic preference are similar at e1, 110 per person per year. These
amounts are significantly large at 6% of annual household income for the hyperbolic
preference and up to 9% in case of the Cobb-Douglas preference. However, when we
allow for the land conversion, the net costs of the green areas are up to 4.2% of annual
household income for both preferences, implying that the opportunity costs of urban
green areas are also substantial.
While accounting for the general equilibrium effects of the land market equilibrium,
our approach has obvious limitations. The first limitation is its reliance on the mono-
centric urban model. This limitation can be easily changed by including more than one
business center in theoretical model32 ; however, in the empirical parts, a polycentric
city analysis requires more intensive data on the location of job centers for every city.
As most of our city samples are monocentric cities33 , we think that monocentric urban
areas are a reasonable simplification with tractable results. The second limitation is the
heterogeneity in households’ preference for green amenity. We here consider only one
class of households; therefore, our study is valid for a representative household. The
effects of green amenity change in urban areas are calculated based on the long-run
change in the land market. In the context of a durable housing structure, the adjust-
ments of residential land use might take a considerable time. The dynamic effects in
the housing market relating to the change in land-intensive amenity change are open
for future research.
32 One approach could be as in Lucas and Rossi-Hansberg (2002).
33 The OECD study on polycentric metropolitan cities only reports 18 European cities with two urban
centers and only 6 with more than two urban cores (Barcelona, Paris, Lyon, Amsterdam, Stockholm and
London) among our sample of more than 300 cities.
APPENDIX 73

Appendix A: First stage regression and Wu-Hausmann test for

IV regressions
In this section, we report the first stage regression and the Wu-Hausman test for IV re-
gression. We use the historical level of urban green spaces (land use code 141) in Corine
Land Cover in 1990 as our instrument variable. To our knowledge, Corin Land Cover
(CLC) was the first systematized the land use over whole Europe, and its earliest ver-
sion was in 1990. However, there are two issues with CLC 1990. First, CLC 1990 did not
cover UK, Sweden and Finland as those three countries only appeared in later version
of Corine Land Cover in 2000 onward. Therefore, we need to drop the city samples
which belong to those three above countries. Second, as CLC cover not only urban area
but also the rural and all lands in Europe. Hence, its resolution is much less precise
than GMES Urban Atlas that covers only urban areas. There are also discrepancies in
these two databases. To decrease the discrepancies, we use the land cover in CLC 2006
and GMES Urban Atlas 2006 and correct for the discrepancies between these two sets.
We assume that the changes between Corine 2006 and Corine 1990 is the evolution of
urban green, while the difference between Corine 2006 and GMES 2006 are just dis-
crepancies in measurement. We adjust the Corine 1990 with these measurement errors
before using it in the first stage regression.

Cobb-Douglass preference

As showed in Table B1, the coefficient between adjusted urban green in Corine 1990 is a
very good predictors for the current level of urban green. The slope is highly significant
and is around 0.8. R2 is at 0.76 − 0.78. The Wu-Hausman test coefficient for endogeneity
(v̂error ) is not significant at 90% of confidence level, meaning that we can confirm the
alternative hypothesis of no endogeneity at 90% confidence level, which further implies
that endogeneity may not be a critical issue in this analysis.
74 APPENDIX

Table B1: Cobb-Douglas - First stage regression and Wu-Hausman test for endogeneity

First stage regression

Dependent variable:
Share of green area in GMES 2006
(1) (2) (3) (4)
Adjusted Share of green area in Corine 1990 0.802∗∗∗ 0.794∗∗∗ 0.792∗∗∗ 0.791∗∗∗
(0.027) (0.027) (0.027) (0.027)

Adjusted R2 0.760 0.782 0.784 0.785

Wu-Hausman test for endogeneity

Dependent variable: Ln Residential Space
(1) (2) (3) (4)
v̂error -0.421 0.603 0.581 0.429
(0.885) (0.472) (0.450) (0.429)

Country FE No Yes Yes Yes

City Geographical Controls No No Yes Yes
Annuli Amenity Controls No No No Yes
Sample All All All All
Observations 10,853 10,853 10,853 10,853
Note: We use the following procedure to test for endogeneity. First Stage: gijc = αgCorine90 +
ϑ0 Z + vijc . Second Stage: ln sijc = ϑZ + ϑ4 gijc + ϑerr v̂ijc + ε ijc where Z is the vector
( 1 w dist Ic X jc Aijc ) and v̂ijc is the residuals from first stage regression. For city
control, we take into account the elevation, average rain fall, average temperature in Jan 01 and
average temperature in July 01 for period 1995-2010. The observations are all annuli of all cities
covered by both GMES and Corine Land Cover 1990. Other variables are those from original
regression (GDP per capita at purchasing power standard and distance to CBD). Regression (1)
to (4) are corresponding to IV regression (5) to (8) in Table 5 in the main text respectively. The
row ”df” reports the degree of freedom. Significance levels are denoted by ∗ for p<0.1, ∗∗ for
p<0.05 and ∗∗∗ for p<0.01.

Hyperbolic preference

As showed in Table B2, the coefficient between adjusted urban green in Corine 1990
is a very good predictors for the current level of urban green, which is similar to the
Cobb-Douglas preference. The Wu-Hausman test coefficient for endogeneity (v̂error ) is
not significant at 90% of confidence level, meaning that we can confirm the alterna-
tive hypothesis of no endogeneity at 90% confidence level, which further implies that
endogeneity may not be a critical issue in this analysis.
APPENDIX 75

Table B2: Hyperbolic - First stage regression and Wu-Hausman test for endogeneity

First stage regression

Dependent variable:
Share of green area in GMES 2006
(1) (2) (3) (4)
Adjusted Share of green in 1990 0.804∗∗∗ 0.796∗∗∗ 0.793∗∗∗ 0.792∗∗∗
(0.027) (0.027) (0.027) (0.027)

Adjusted R2 0.758 0.781 0.783 0.784

Wu-Hausman test for endogeneity

Dependent variable: Inverse of Space
(1) (2) (3) (4)
v̂error 2.007 −0.626 −0.598 −0.526
(1.904) (0.844) (0.776) (0.720)

Country FE No Yes Yes Yes

City Geographical Controls No No Yes Yes
Annuli Amenity Controls No No No Yes
Sample All All All All
Observations 10,853 10,853 10,853 10,853
Note: We using the following procedure to test for endogeneity. First Stage: gijc =
αgCorine90 + ϑ0 Z + vijc ; Second Stage: s 1 = ϑZ + ϑ4 gijc + ϑerr v̂ijc + ε ijc where Z is the vec-
ijc

tor ( 1 w dist Ic X jc Aijc ) and v̂ijc is the residuals from first stage regression.
For city control, we take into account the elevation, average rain fall, average temperature
in Jan 01 and average temperature in July 01 for period 1995-2010. The observations are
all annulus from all cities covered by both GMES and Corine Land Cover 1990. Other
variables are those from original regression (GDP per capita at purchasing power stan-
dard and distance to CBD). Regression (1) to (4) are corresponding to IV regression (5) to
(8) in Table 5 in the main text respectively. The row ”df” reports the degree of freedom.
Significance levels are denoted by ∗ for p<0.1, ∗∗ for p<0.05 and ∗∗∗ for p<0.01.

Appendix B: Counterfactural analysis

B1: Willingness to pay computation

Cobb Douglas preference

One notes that within city boundary, we have R = ψ b , and replace it into the indirect
utility in equation (3.13), we obtain the general willingness to pay to keep the same
level of utility by using the indirect utility equation V (y − WTP, a1 , R1 ) = V (y, a0 , R0 )
in case of closed cities. We have:
 
 0 − 1−φ ϕ
a0 − a1 ) R
WTP = y 1 − e 1− ϕ (
ϕ
. (3.22)
R1

Using the equilibrium bid rent function in (3.14), we can recompute the general equi-
librium willingness to pay as:
 e  1−1 ϕ !
u
WTP = y 1 − (3.23)
u1
76 APPENDIX

where the new equilibrium utility u1 is determined by the population constraint condi-
tion (as we are in closed cities), and it depends on whether the new policy involved the
conversion of urban green land into residential land. Suppose, the new green amenity
level a1 required a smaller fraction of land devoted for urban green g1 < g0 , and the
unused land ( g0 − g1 )` gets converted into residential land use, we can compute the
new equilibrium utility as below:
 e φ
Rb 1− g1
   1 2πrdr 
 0 y−(1−φ− ϕ) e− ϕa1 φ 
u1conversion = ue  e  . (3.24)
 
b
1− g0
R 
 1 2πrdr
  
0
y−(1−φ− ϕ) e− ϕa
φ
0

And if there is no conversion policy, the new utility equilibrium is

 e φ
Rb 1− g0
   1 2πrdr 
 0 y−(1−φ− ϕ) e− ϕa1 φ 
u1no conversion = ue  e  . (3.25)
 
 Rb 1 − g 0 
 1 2πrdr
  
0
y−(1−φ− ϕ) e− ϕa
φ
0

Hyperbolic preference

Using similar approach as in above section, we use the equilibrium land rent R = ψ b,
and replace them into the indirect utility in equation (3.18) to obtain the willingness to
pay as below:  √ √
 √ 
WTP = a1 − a0 − 2θ R1 − R0 . (3.26)
Using the equilibrium bid rent function in (3.19), we can recompute the general equi-
librium willingness to pay as:
WTP = u1 − ue (3.27)
where the new equilibrium utility u1 is determined by the population constraint con-
dition (as we are in closed cities), and by land conversion policy. We can compute the
new equilibrium utility as below:
 be   be
Rbe

0 1 1 0 0
R R
 (1 − g )2πrdr   (1 − g )(y + a )2πrdr − (1 − g )(y + a )2πrdr 
1 e0  0 0
uconversion = u  e + .

b b e
R   R 
1
(1 − g )2πrdr 1
(1 − g )2πrdr
0 0
(3.28)
And if there is no conversion policy, the new utility equilibrium is
 be 
0 1 0
R
 (1 − g )( a − a )2πrdr 
1 e 0
uno conversion = u +  . (3.29)

b e
 R 
1
(1 − g )2πrdr
0
APPENDIX 77

B2: Parameters recovered

Cobb-Douglas preference

We recover the model parameters from the estimated coefficient of residents’ land use
using the values of ϑ0 , ϑ1 , ϑ2 , ϑ3 and ϑ4 from Column (8) in Table 5. Country utility
levels are recovered from the parameters ϑ5c and the constant term ϑ0 .34 . From the
theoretical model, we recover the residential space, composite goods and the residential
land rent as

2
ŝ(w, r, g, u, ε) = (we−τ̂1 r−τ̂2 r )ϑ̂1 u e gϑ̂4 +ϑ̂6 Xjc +ϑ̂7 Aijc +ε̂ijc
!
−τ̂ 1 r −τ̂ 2 r2 −ϑˆ1
ẑ(w, r ) = we
1 − ϑˆ1
2
we−τ̂1 r−τ̂2 r − ẑ(w, r )
R̂(w, r, g, u, ε) =
ŝ(w, r, g, u, X, e)

where we use τ̂ 1 = ϑˆ2 /ϑ̂1 , τ̂ 2 = ϑˆ3 /ϑ̂1 , and (1 − ϕ̂ − φ̂)/(1 − ϕ̂) = −ϑˆ1 /(1 − ϑˆ1 ).

Hyperbolic preference

Similarly, we have the following functions for residential space s, composite good z and
the residential land rent R for the Hyperbolic preference.

θ̂
ŝ(w, r, g, u, ε) =
w − w(τ̂ 1 r + τ̂ 2 r2 ) + β̂g − uc + ϑ̂5 X jc + ϑ̂6 Aijc + ε̂ijc
w − w(τ̂ 1 r + τ̂ 2 r2 ) − β̂g + uc − ϑ̂5 X jc − ϑ̂6 Aijc − ε̂ijc
ẑ(w, r, g, u, ε) =
2
w − w(τ̂ 1 r + τ̂ 2 r2 ) − ẑ(w, r, g, u, ε)
R̂(w, r, g, u, ε) =
ŝ(w, r, g, u, ε)

where θ̂ = 1/ϑˆ1 , τ̂ 1 = −ϑˆ2 /ϑ̂1 , τ̂ 2 = −ϑˆ3 /ϑ̂1 , and β̂ = ϑˆ1 /ϑˆ4 .

B3: Counterfactural exercises – detailed steps

We first define the baseline model and the counterfactual exercises. Both baseline
and counterfactuals use the observed distance to the city center rijc and amenities X jc
and A jc . The baseline model includes the observed city wage w jc , the share of green
34 As 1/φ 1/φ
we drop Austria in the country dummies, we have u at = ϑ0 , and all other countries as uc =
ϑ5c + ϑ0
78 APPENDIX

urban areas gijc , the estimated values of country utility ubc and the unobserved het-
erogeneity or measurement error bεijc . Formally, we set the baseline model values to
ŝ0ijc = ŝ(w jc , rijc , gijc , ubc ,bεijc ), ẑ0ijc = ẑ(w jc , rijc ) and R̂0ijc = R̂(w jc , rijc , gijc , ubc ,bεijc ).
We now investigate the impact of canceling 50% of the urban green areas in each an-
nulus. In counterfactual exercise 1, we consider open cities where utility levels and un-
observed heterogeneity are maintained at the estimated levels ubc and bεijc . This consider-
ation helps us discuss a long-term and unregulated perspective, where urban planners
do not impose restrictions on workers’ mobility within and between cities. We then
remove half of the green urban areas by setting gijc 0 = 0.5 × g . In both cases, we
ijc
1 0 1 1
set ŝijc = ŝ(w jc , rijc , gijc , ubc ,bεijc ) and ẑ = ẑ(w jc , rijc ), while R̂ = R̂(w jc , rijc , gijc 0 ,u
bc ,bεijc ).
1 0
Residents’ land use should increase (ŝijc > ŝijc ) because residents require compensa-
tion for the reduction of green area amenities. If green urban areas are left with no
use, the total available space remains constant and is given by ∑ijc 1 − gijc `ijc , where

`ijc is the land surface of annulus i in city j and country c. Since resident’s land use
increases, cities host fewer residents. If green urban areas are converted in residential
land, city populations may grow if the new supply of land, gijc 0 ` , is larger than the
ijc
0 1
increase in residents’ land demand from ŝijc to ŝijc . More formally, population grows if
 
0 < 0 1
∑ijc ∑ijc


1 − g ijc ` ijc /s ijc 1 − g ijc `ijc /sijc .
In counterfactual exercise 2, we consider closed cities with exogenous city popula-
tions. The study of closed cities can be appropriate in evaluating policy changes that
occur simultaneously in all cities, such as changes in EU policies.35 Here, our aim is to
discuss a midterm or regulated perspective, where urban planners are able to restrict
workers’ mobility between cities but allow residents’ land use to change. We again re-
0 = 0.5 × g ). We set ŝ2 = ŝ ( w , r , g0 , u2 , b
move half of the green urban areas (gijc ijc ijc jc ijc ijc jc eijc ),
2
ẑ2ijc = ẑ(w jc , rijc ) and Rˆijc = R̂(w jc , rijc , gijc
0 , u2 , b 2
jc eijc ), where u jc is the counterfactual city
utility level. In the absence of the conversion of green urban areas to residential plots,
we set the city utility level u2jc such that the city population spreads over the base-
line residential area; that is, we impose that each u2jc solves the population identity
∑i (1 − gijc )lijc /ŝ0ijc = ∑i (1 − gijc )lijc /ŝ2ijc . In the case of land conversion, we set u2jc such
that the city population spreads over the new residential land supply. Then, u2jc solves
the population identity ∑i (1 − gijc )`ijc /ŝ0ijc = ∑i (1 − gijc 0 )` / ŝ2 .36 However, although
ijc ijc
utility levels are important concepts in welfare analysis, they are difficult to interpret
quantitatively. Therefore, we also compute the compensating variation wage w2jc as the
city wage that maintains the baseline utility level when we remove green urban areas,
35 Cheshire and Shepard (2002) also use the closed city model to analyze the welfare effects of policy
changes in the UK.  
1/φ ∑i (1− gijc )`ijc /sijc
 1/φ 2
36 This approach yields the counterfactual utility levels u2 = ( u ) and
jc c
 
(∑i (1− gijc )`ijc /sijc )
0 )` s2
 1/φ ∑i (1− gijc ijc ijc
u2jc = (uc )1/φ  .
∑i (1− gijc )`ijc /sijc
APPENDIX 79

which is equivalent to setting the wage w2jc such that the above population identities
0 ,u
hold with ŝ(w2jc , rijc , gijc bjc ,bεijc ). Under the Cobb-Douglas preferences, this assumption
1
simplifies to the compensating variation wage w2jc = w jc (u2jc /u0jc ) ϑˆ1 ϕ̂ .
The above two counterfactual exercises hinge on the assumption that empirical
model residuals ε ijc reflect land heterogeneity that is unobserved to the econometricians
but observed and used by residents in their land plot size choices. Such heterogeneity
is reported in the counterfactual results. This assumption may be strong, as it imposes
strong information on behalf of residents. Therefore, we take the opposite view and
assume that the residuals ε ijc consist of measurement errors that can be observed nei-
ther by the econometricians nor the residents. In that case, residents do not base their
decisions on ε ijc , and we set ε ijc = 0 in counterfactual exercises 3 and 4.

Appendix C: Robustness check for Cobb-Douglas preference with

NUTS3 incomes
In this section, we report the results of estimation of the structural model using GDP
per capita at NUTS3 level as the proxy for city wage. The results are similar to those in
Table 3.2 in the main text.
APPENDIX

Table C1: Regression Results using GDP per capita at NUTS3 as proxy for city’s wage level
Dependent variable: Ln Residential Space
OLS IV
(1) (2) (3) (4) (5) (6) (7) (8)
Constant −0.686∗∗∗ −0.899∗∗∗ −0.834∗ −0.230 −0.682∗∗∗ −0.880∗∗∗ −0.808∗ −0.213
(0.119) (0.256) (0.467) (0.425) (0.119) (0.262) (0.473) (0.430)
Ln Household Income −0.210∗∗∗ −0.614∗∗∗ −0.608∗∗∗ −0.542∗∗∗ −0.211∗∗∗ −0.610∗∗∗ −0.604∗∗∗ −0.540∗∗∗
(0.077) (0.103) (0.098) (0.090) (0.077) (0.103) (0.098) (0.090)
Distance to CBD 1.596∗∗∗ 1.482∗∗∗ 1.514∗∗∗ 1.004∗∗∗ 1.591∗∗∗ 1.475∗∗∗ 1.509∗∗∗ 0.998∗∗∗
(0.239) (0.220) (0.212) (0.219) (0.240) (0.221) (0.213) (0.220)
Distance to CBD square −0.610∗∗∗ −0.515∗∗∗ −0.518∗∗∗ −0.322∗∗ −0.608∗∗∗ −0.512∗∗∗ −0.517∗∗∗ −0.320∗∗
(0.134) (0.146) (0.149) (0.142) (0.133) (0.146) (0.148) (0.140)
Share of Urban Green −2.045∗∗∗ −1.773∗∗∗ −1.681∗∗∗ −1.487∗∗∗ −2.079∗∗∗ −1.948∗∗∗ −1.851∗∗∗ −1.611∗∗∗
(0.386) (0.245) (0.241) (0.203) (0.413) (0.274) (0.262) (0.228)
Country FE No Yes Yes Yes No Yes Yes Yes
City Geographical Controls No No Yes Yes No No Yes Yes
Annuli Amenity Controls No No No Yes No No No Yes
Sample All All All All All All All All
Observations 10,853 10,853 10,853 10,853 10,853 10,853 10,853 10,853
Adjusted R2 0.227 0.513 0.525 0.607 0.218 0.510 0.522 0.604
Residual Std. Error 0.567 0.450 0.445 0.404 0.570 0.451 0.446 0.406
df 10,848 10,827 10,823 10,819 10,848 10,827 10,823 10,819
F Statistic 19.83∗∗∗ 32.52∗∗∗ 33.78∗∗∗ 46.95∗∗∗ 18.44∗∗∗ 31.44∗∗∗ 32.74∗∗∗ 47.94∗∗∗
Note: Here, we use the GDP per capita taken from Regional Economic Accounts from Eurostat at NUTS3 level with adjustment to purchasing power
standard (PPS) as the proxy for city income level, and it is measured in e100,000. The distance to CBD is measured in 10 kilometres. The inverse
of space is calculated by divided number of inhabitants in each annuli with the areas within the annuli minus the areas using as urban green (in
100 meters). We exclude Cyrus and Luxembourg as the Eurostat population grid database does not cover Cyrus and the household income data at
NUTS2 of Eurostat does not cover Luxembourg. United Kingdom and Finland are also excluded as they are not covered by Corine Land Cover 1990.
City boundary is chosen at 20% cut-off point. For city control, we take into account the elevation, average rain fall, average temperature in Jan 01 and
average temperature in July 01 for period 1995-2010. City amenity controls include the share of industrial, sport and leisure land use as well as the
share of forest and agriculture land within 100 meters buffer from residential area. Standard errors are clustered at city level. The row ”df” reports
the degree of freedom. Significance levels are denoted by ∗ for p<0.1, ∗∗ for p<0.05 and ∗∗∗ for p<0.01.
80
APPENDIX 81

Table C2: First stage regression and Wu-Hausman test for endogeneity

First stage regression

Dependent variable:
Share of green area in GMES 2006
(1) (2) (3) (4)
Adjusted Share of green area in Corine 1990 0.797∗∗∗ 0.793∗∗∗ 0.790∗∗∗ 0.789∗∗∗
(0.028) (0.027) (0.027) (0.027)

Adjusted R2 0.759 0.781 0.784 0.784

Wu-Hausman test for endogeneity

Dependent variable: Ln Residential Space
(1) (2) (3) (4)
v̂error −0.330 0.413 0.406 0.241
(0.911) (0.455) (0.438) (0.407)

Country FE No Yes Yes Yes

City Geographical Controls No No Yes Yes
Annuli Amenity Controls No No No Yes
Sample All All All All
Observations 10,853 10,853 10,853 10,853
Note: We using the following procedure to test for endogeneity. First Stage: gijc =
αgCorine90 + ϑ0 Z + vijc ; Second Stage: ln sijc = ϑZ + ϑ3 gijc + ϑerr v̂ijc + ε ijc where Z is the
vector ( 1 w dist Ic X jc ) and v̂ijc is the residuals from first stage regression. For city
control, we take into account the elevation, average rain fall, average temperature in Jan 01 and
average temperature in July 01 for period 1995-2010. The observations are all annulus from all
cities covered by both GMES and Corine Land Cover 1990. Other variables are those from origi-
nal regression (GDP per capita at purchasing power standard and distance to CBD). Regression
(1) to (4) are corresponding to IV regression (5) to (8) in Table A1 respectively. The row ”df”
reports the degree of freedom. Significance levels are denoted by ∗ for p<0.1, ∗∗ for p<0.05 and
∗∗∗ for p<0.01.

Table C3: Counterfactual analysis: open and closed cities

Cities Composite Housing Income Residential Green

Goods (Z) Rent (R × s) (W) Area Area (GA)
number (e1000) (e1000) (e1000 ) (km2 ) (km2 )
Spatial heterogeneity 264 6.39 11.84 26.55 25.29 4.52
(3.01) (5.58) (13.51) (48.91) (9.87)
No spatial heterogeneity 264 5.03 6.35 11.76 26.55 4.52
(2.99) (5.55) (13.51) (48.91) (9.87)
Note: Household income is taken from Eurostat at NUTS2 level and is measured on purchasing power
standard (PPS) at 1000e. More details on PPS measure, please check Eurostat technical documents.
The standard deviation is reported in the parenthesis.
APPENDIX

Table C4: City structure under closed and open scenarios g = 0.5 × g0
Total loss in
Cities s Land Rent Green value Pop. U/Ū b Comp. W housing market wage comp. population
(number) (100m2/hab) (e/m2 /y) (e/m2 /y) (mil.) (e1000) (e mil.) ( e mil.) 1, 000hab.
Baseline 0 (with heterogeneity) 264 0.95 161.88 177.25 0.31 1.00
(0.44) (127.49) (140.65) (0.69) (0.33)
g = 0.5 × g0
Counterfactual 1: Open City
Urban green conversion No 264 1.00 152.18 162.30 0.29 1.00 228.36 0 17.62
(0.46) (118.02) (124.64) (0.65) (0.33) (678.27) (45.39)
Yes 264 0.99 153.16 162.30 0.32 1.00 −131.58 0 −11.27
(0.46) (119.03) (124.64) (0.71) (0.33) (311.15) (25.71)
Counterfactual 2: Closed City
Urban green conversion No 264 0.95 161.78 172.54 0.31 0.95 29.56 4.50 1,403.63 0
(0.44) (127.61) (135.29) (0.69) (0.31) (15.28) (43.70) (5,650.12)
Yes 264 1.03 145.86 154.38 0.31 1.04 24.58 9.65 −726.84 0
(0.47) (111.58) (115.64) (0.69) (0.34) (12.62) (56.06) (2,420.57)
Baseline 0 (without heterogeneity) 264 0.89 163.66 175.17 0.28 1.00
(0.37) (122.22) (127.79) (0.53) (0.33)
g = 0.5 × g0
Counterfactual 3: Open City
Urban Green Conversion No 264 0.94 153.83 160.68 0.27 1.00 204.58 0 15.83
(0.38) (112.88) (114.27) (0.50) (0.33) (505.92) (36.13)
Yes 264 0.94 154.12 160.68 0.29 1.00 −112.14 0 −10.32
(0.38) (112.85) (114.27) (0.55) (0.33) (271.64) (22.17)
Counterfactual 4: Closed City
Urban Green Conversion No 264 0.89 163.54 170.64 0.28 0.95 16.65 6.01 1,243.83 0
(0.37) (122.30) (123.46) (0.53) (0.31) (6.47) (33.94) (4,367.02)
Yes 264 0.97 147.30 153.74 0.28 1.04 14.61 7.64 −670.88 0
(0.39) (106.00) (107.66) (0.53) (0.34) (5.70) (35.75) (2,123.28)
Note: Household income is taken from Eurostat at NUTS2 level and is measured on purchasing power standard (PPS) at 1000e. More details on PPS measure, please check Eurostat technical
1/φ̂
U jc U jc u
jc
documents. Population is calculated in million inhabitants. Utility is measured relatively with baseline average. Basically, we measure = 1 = where u jc
Ū baseline U baseline
1/φ̂
N ∑ jc jc

1 ubaseline
N ∑ jc c
is calculated as in above section. The standard deviation is reported in the parenthesis.
82
 APPENDIX

Table C5: Quantiles analysis: between- and within-city statistics

Between cities Within cities

City Income Quartiles City Size Quartiles Distance to CBD Quartiles
0-0.25 0.25-0.50 0.50-0.75 0.75-1 0-0.25 0.25-0.50 0.50-0.75 0.75-1 0-0.25 0.25-0.50 0.50-0.75 0.75-1

Baseline wage (e1000/hab/y) 13.20 20.28 27.89 44.83 19.50 25.79 26.67 34.24 30.65 30.36 30.45 30.22
(2.71) (2.16) (2.02) (12.69) (6.98) (14.34) (11.62) (15.44) (15.16) (15.00) (15.06) (15.01)
Increase in comp.var.wage(e1000/hab/y)
Counterfactual 1: Open City
Urban green conversion No 0 0 0 0 0 0 0 0 0 0 0 0

Yes 0 0 0 0 0 0 0 0 0 0 0 0

Counterfactual 2: Closed City

Urban green conversion No 1.38 2.15 3.08 5.42 2.48 2.86 2.71 3.96 3.54 4.15 3.25 2.50
(1.79) (1.32) (1.69) (2.82) (2.66) (2.34) (1.74) (2.87) (4.24) (3.97) (3.08) (2.50)
Yes −0.83 −1.59 −2.06 −3.40 −2.12 −1.96 −1.60 −2.21 −1.20 −1.66 −1.94 −2.59
(1.09) (1.43) (1.63) (2.61) (2.74) (1.99) (1.25) (1.73) (2.47) (2.03) (1.99) (2.96)

Land rent baseline (e/m2 /y) 80.16 146.52 178.45 242.40 124.73 152.69 167.74 202.36 349.81 201.97 133.21 102.17
(37.40) (95.91) (140.07) (147.18) (86.63) (110.20) (135.50) (156.72) (307.07) (157.57) (118.50) (113.42)
Landlord loss (e/m2 /y)
Counterfactual 1: Open City
Urban green conversion No 4.10 8.25 10.66 15.80 8.14 8.79 9.65 12.22 19.43 13.57 7.66 5.69
(4.68) (7.16) (11.76) (15.32) (11.90) (8.66) (12.08) (12.11) (29.44) (13.57) (11.01) (16.61)
Yes 3.58 6.89 9.99 14.44 5.88 7.79 9.12 12.11 19.43 13.57 7.66 5.69
(4.40) (6.96) (10.29) (13.97) (6.95) (7.49) (11.23) (13.52) (29.44) (13.57) (11.01) (16.61)
Counterfactual 2: Closed City
Urban green conversion No 0.12 0.14 0.10 0.05 0.00 0.03 0.15 0.23 −0.16 1.99 −0.12 −0.45
(0.17) (0.23) (0.36) (1.14) (0.40) (0.51) (0.40) (0.95) (26.40) (10.97) (4.81) (9.05)
Yes 6.28 13.39 18.19 26.24 15.44 14.53 15.09 19.03 30.99 20.90 12.81 10.08
(7.50) (11.91) (21.07) (30.11) (26.76) (15.05) (19.48) (20.69) (37.99) (23.90) (18.68) (27.27)
Note: Here, we report only the case with heterogeneity; for non-heterogeneity case (like in counterfactual exercises 3 and 4), the results are similar. The loss for landlord is accounted as the difference between
new land rent equilibrium and land rent in baseline model 0. The standard deviations are reported in the parenthesis.
83
Chapter 4

Geographical Stratification of Green

Urban Areas

”The more successfully a city mingles everyday diversity of uses and users in its everyday
streets, the more successfully, casually (and economically) its people thereby enliven and
support well-located parks that can thus give back grace and delight to their neighbour-
hoods instead of vacuity.” Jane Jacobs, The Death and Life of Great American Cities, 1961.

4.1 Introduction
The future of human development is inextricably tied to the future of our urban areas.
Throughout history, cities have been at the centre of change, civilization and attraction,
ranging from the ancient Athens to ’Renaissance’ Paris to ’financial centre’ London. The
world has changed gradually from mostly rural toward more urbanized areas, with
more than half of all the world’s citizens currently living in an urban area as of 2017
(UN DESA, 2018), and in Europe, urban dwellers account for approximately 74% of all
its population (Eurostat, 2016). In other words, when we discuss the future of urban
areas, the discussion relates to the lives of the nearly three-quarters of the European
population and one-half of all human beings who have chosen to live in urban areas.1
If this trend is irreversible, then the next question should concern how we can make
such urbanization work. How can we improve urban quality of life and make city life
sustainable? Undoubtedly, one of the main components that make cities attractive to
their residents are their public park and garden system.2
In the European context, green amenities are also a crucial factor in determining
1 By2030, this number is expected to increase to two-thirds of the world population.
2 The benefits of urban green spaces have been well documented in many fields, such as sociology,
psychology, economics, health, environment, etc. For a systematic review, see IFPRA (2013) and the EU
DG-ENVI (2012) report.

85
86 CHAPTER 4. GEOGRAPHICAL STRATIFICATION OF GREEN URBAN AREAS

the level of satisfaction with urban life. Since 2007, Eurostat has conducted a series of
surveys in selected cities3 in 28 Member States of the European Union to measure the
perceptions of urban residents about their satisfaction with city life and other aspects
such as housing, transportation, crime and green space access. Figure 1 depicts the city
satisfaction level and the satisfaction in urban green systems from the European Urban
Perception survey. The trend clearly shows that cities with well-maintained and effi-
ciently accessed parks, as indicated by a high level of satisfaction with green amenities,
are much more likely to have a higher level of life satisfaction. Well-located urban park
systems play an essential role in establishing vibrant urban life, as Jane Jacobs (1961)
noted. The value of green urban spaces in the everyday lives of urban citizens is, no
doubt, a relevant indicator to measure the quality of life in a city.

Figure 1: Perceptions of Urban Green Areas and Life Satisfaction in the Europe Union

Note: Authors’ own computation with the data from Eurostat Urban Perception Survey in 2012 for European cities. All
measures are in percentage. Urban Green Space Satisfaction is the percentage of people strongly satisfied with current
urban green space systems in their city. City Satisfaction Level is the percentage of people who strongly agreed that
their city is satisfied to live.

In this paper, we study the provision of small urban green spaces in the context of
EU cities. According to our dataset, 95% of urban green areas consists of small parks,
which accounts for 50% of the total urban green surface in EU cities. Although small
parks are valuable to residents for their close access, they are spread across the urban
area and compete for the land allocated for residential purposes. To shed light on this
3 Currently, 79 cities are used to implement the urban perception survey.
4.1. INTRODUCTION 87

question, we first develop a theoretical framework that embeds two important features
of urban green spaces: they are land intensive, and their sizes and access to them rep-
resent valuable amenities for residents. As most urban green areas are controlled by
urban planning, we derive and discuss the optimal land allocation by a benevolent ur-
ban planner. The planner balances the space allocated to residents and urban green
spaces. At the optimum, the number of urban green spaces is shown to be a non-
monotone function of distance to the city centre. The sizes of urban green spaces and
the distances to other urban green areas increase as one moves to the urban fringe. The
model also allows us to compare cities with various productivities. It is shown that
more productive cities include more numerous parks with smaller sizes and longer ac-
cess distances. These properties are used to guide the empirical investigation on the
300 largest European cities by using the GMES Urban Atlas database (European En-
vironmental Agency). The empirical analysis confirms the non-monotone relationship
between the number of urban green spaces and the distance to the city centre. It also
confirms that the distance between two parks increases as one moves toward the urban
fringe. Finally, it corroborates that richer cities are associated with a denser network of
urban green areas.
This paper contributes to the literature in public and urban economics. It adds to the
public economics literature by investigating the provision of local public goods in the
context of non-uniform population distribution and land-intensive public goods. Cre-
mers et al. (1986), Fujita (1986), and Shakashita (1987) study the location of spaceless
local public goods in cities with an exogenous uniform density distribution. Their main
conclusion is that local public goods should be equidistant. In this paper, we consider
the choices for residential space, park sizes and park locations. As a result, the residen-
tial density is not uniform, and the sizes and distances between parks change with their
location relative to the city centre. Our paper shares a close resemblance with Berliant
(2006), who considers the endogenous location of a local public good. However, as in
the above contributions, the local public good in that study is a spaceless point with no
land requirement. In contrast, this paper discusses the geographical allocation of small
local public goods that require extensive land use and provide an amenity that decays
in the distance travelled to them.
This paper also extends the numerous urban economics models that investigate the
effect of amenities on urban land use. For instance, Fujita (1989) studies the optimal
allocation of neighbourhood goods (localized and congestionable public goods) and their
decentralization through Pigouvian taxes. As such goods are spaceless, they do not use
land as an input and allow immediate access, meaning that their geographical disper-
sion cannot be studied. Lee and Fujita (1997) consider the optimal location of a green
belt and its effect on urban land rents. Lee (1998) considers the optimal size of a cen-
tral park in a city. These two works consider the land intensity of urban green areas
but focus on the provision of a single area. This contrasts with the (European) reality
88 CHAPTER 4. GEOGRAPHICAL STRATIFICATION OF GREEN URBAN AREAS

of a system of many green spaces and public parks, which is our emphasis. Our pa-
per also relates to the urban economics literature on open spaces. Open spaces differ
from green urban areas to the extent that they are generally not organized and main-
tained by planners (or other agents) for recreational purposes. Turner (2005) discusses
the residents’ choice to locate close to such open spaces in a linear city model with
discrete locations. In that work, residents have inelastic land consumption, and open
spaces are exogenous land plots. Caruso et al. (2007) study the emergence of sprawl
patterns where households balance the benefits from proximity to other humans and
to open spaces (see also Caruso et al. 2005). Walsh (2006) investigates the impacts of
open space protection and growth control on city patterns in discrete neighbourhood
settings. Wu and Plantinga (2003) introduce an exogenous open space in a continuous
two-dimensional model to discuss leap-frogging patterns. In contrast to the latter con-
tribution, the present paper studies the endogenous locations and sizes of many small
urban green spaces in a continuous two-dimensional space.
Although the literature offers theoretical guidance about the provision of green ur-
ban areas, the examination of its empirical reality appears to be scant. To the best of
our knowledge, this paper is one of the first contributions that studies the allocation of
urban green space with a high-resolution geographical database such as that provided
in the GMES Urban Atlas database (European Environmental Agency). This database
requires the assembly of a large set of land plots and the construction of the data to
be exploited in regression analyses. In a companion paper, Picard and Tran (2019) use
the same database and similar data processing to study the trade-off between residen-
tial space and green urban areas. They highlight a non-linear relationship between the
land share of green urban areas and the distance to the city centre. By contrast, this
paper disentangles the effect of the latter on the size, number and dispersion of parks.
The structure of the paper is as follows. Section 2 presents and discusses the the-
oretical model. It first highlights the general principles according to which the city
planner chooses the locations and sizes of green urban spaces. It then focuses on Cobb-
Douglas preferences to provide the closed-form solutions that are used for guidance
in the empirical research. Section 3 presents the empirical study of green urban strat-
ification across the largest European cities. The model predictions on the green urban
areas and population densities are then tested. Section 4 concludes the paper. The
appendices include additional material on the mathematical analysis and empirical ro-
bustness checks.
4.2. THE MODEL 89

4.2 The Model

4.2.1 Household preferences

We consider a monocentric circular city with a central business district surrounded first
by residences and then by agricultural areas. For simplicity, agricultural areas generate
a land rent of R A and consist of fenced private properties that provide no amenity to
urban dwellers.
Households are homogeneous and consume an amount z of a composite good, use a
land surface s for housing and enjoy a green amenity level a at their residential location.
They are endowed with a utility function U (z, s, a) that is increasing and quasi-concave
with respect to z, s and a. Households earn the same wage w but differ according to
their place of residence. Residence is measured in polar coordinates (r, φ), where r is
the length of the ray to the centre and φ is the angle with the x −axis. The city extends
to the border b, so that (r, φ) ∈ [0, b] × [0, 2π ]. Workers commute on the ray from
their residence to the business district and incur a commuting cost t(r ) with t0 > 0.4
We denote their net disposable income by y(r ) = w − t(r ). We assume a positive net
income on the whole city space: y(r ) > 0, r ∈ [0, b].
We focus on the green space amenity. Parks provide an amenity level that is increas-
ing in their size k and decreasing in their distance to one’s residence, d. Let a(k, d) be
the function representing this amenity (ak ≡ ∂a/∂k > 0 > ad ≡ ∂a/∂d). In addition to
their land cost, parks incur a maintenance cost C (k) that is an increasing and concave
function of their size (C 0 > 0 and C 00 < 0).
The analytical discussion of the allocation of residences and parks on a two-dimensional
plane is a difficult task. In this paper, we simplify the analysis by distinguishing the
commutes to work and the travels to parks on different space dimensions. Whereas
commuting takes place on the ray passing through the business district, as is stan-
dard in the urban economic literature, traveling to a park is assumed to take place
on the annulus (ring) on which a worker resides. Hence, residents residing at (r, φ)
travel to the parks on the annulus of ray r and width dr or, more precisely, on the
space [r, r + dr ] × [0, 2π ]. The amenity for a resident at (r, φ) is therefore given by
a[k(r ), d(r, φ)], where d(r, φ) is the distance travelled to the park on the annulus. Be-
cause of the symmetry of households and space, we can focus on parks that have the
same size k (r ) and same distance between one another. Furthermore, as shown in Fig-
ure 2, symmetry allows us to concentrate on the park and residents on the annulus strip
[r, r + dr ] × [0, π/n(r )]. There are 2n(r ) such annulus strips at distance from the centre
r. In this annulus strip, a resident at (r, φ) travels a distance to the closest park equal to
d(r, φ) = rφ.
4 Thetransportation cost can either be counted as a real cost or a time cost (in this case, t(r ) is a fraction
of the wage).
90 CHAPTER 4. GEOGRAPHICAL STRATIFICATION OF GREEN URBAN AREAS

Figure 2: Construction of Urban Green Areas on an Annulus

Note: The figure depicts the locations of residential and urban green areas in the theoretical model. Residents commute
to work on the ray from the coordinate (0, 0) to their residence at (r, φ). They travel to the closest park(s) on the annulus
that includes their residence at distance r from city center. Park size is k (r ) and number of parks is n(r ).
4.2. THE MODEL 91

4.2.2 Optimal green urban spaces

We consider a benevolent city planner who chooses the number n(r ), size k (r ), com-
posite good consumption z(r, φ) and residential surface s(r, φ).5 We consider an open
city where, in the spirit of Herbert and Steven (1960), the planner keeps the residents’
utility equal to the outside utility level u. The planner maximizes the surplus
Z b Z π/n(r ) Z π/n(r) 
1
max S = 2n(r ) [w − t(r ) − z(r, φ)] rdφ − R A rdφ − C [k (r )] dr
{s,z,n,k} 0 k (r )/r s(r, φ) 0

subject to
U [z(r, φ), s(r, φ), a (k (r ), rφ)] = u

The term in curly brackets represents the net resource generated in the annulus strip
[r, r + dr ] × [0, π/n(r )]. There are 2n(r ) such annulus strips. The first term in curly
brackets represents the net resources generated by a worker, namely, the value of her
production minus her commuting cost and personal consumption of composite good.
This is multiplied by the number of workers residing at (r, φ), which is given by the
land supply, rdφ, divided by the worker’s use of land, s(r, φ). Workers reside on land
from the angle k(r )/r to the angle π/n(r ). The second and third terms in curly brackets
represent the land cost and the maintenance cost.
The consumption of the composite good is given by the utility constraint. Since
U is an increasing function of z, the identity U (z, s, a) = u can be inverted to give
z̃(s, a, u). Because U is an increasing function of all its arguments, it can be verified
that ∂z̃/∂s < 0, ∂z̃/∂a < 0 and ∂z̃/∂u > 0. Maximizing point-wise the surplus S with
respect to s(r, φ) is equivalent to finding the function

w − t(r ) − z̃(s, a, u)
V (r, a, u) = max
s s
and applying it to a = a (k (r ), rφ). This function represents the planner’s value of
land generated by workers. Using the envelope theorem, it can be verified that Vr ≡
∂V/∂r < 0, Va ≡ ∂V/∂a > 0 and Vu ≡ ∂V/∂u < 0. The planner’s problem then
simplifies to
Z b Z π/n(r ) Z π/n(r) 
max S = 2n(r ) V [r, a (k (r ), rφ) , u] rdφ − R A rdφ − C [k (r )] dr
{n,k} 0 k (r )/r 0

Note that at the optimum, parks cannot occupy the full surface of each annulus, i.e.,
k (r )/r > π/n(r ). Indeed, consider each annulus at a distance r from the business
district. Otherwise, if k (r )/r ≤ π/n(r ), the first integral in curly brackets would vanish
to zero, meaning that parks would only be a net cost and that the planner would have
an incentive to eliminate the park.
5 Thecomposite good consumption and residential surface z(r, φ) and s(r, φ) can be decentralized
though a land market equilibrium.
92 CHAPTER 4. GEOGRAPHICAL STRATIFICATION OF GREEN URBAN AREAS

The optimal size of parks is obtained by differentiating point-wise this expression

by k (r ). The marginal welfare is given by
Z π/n(r)
Sk = Va [r, a (k (r ), rφ) , u] ak [k (r ), rφ] rdφ − V [r, a (k (r ), k (r )) , u] − C 0 [k (r )]
k (r )/r
(4.1)
An increase in park size has three effects on welfare. First, it gives a higher amenity
to residents and increases the value of land (first term). Second, it decreases the space
available for residences and therefore social welfare by the residential value of the par-
cel of land (second term). Finally, it increases the cost of maintenance. The optimal size
of parks requires that Sk =0.
We now turn to the optimal number of parks. To ease the analysis, we avoid the
problem of integers and treat the number of parks n(r ) as a function of real numbers
(with n(r ) ≥ 1). The optimal number of parks is obtained by differentiating point-wise
the planner’s objective by n(r ). The marginal welfare is given by
π/n(r ) Z π/n(r)
Z 
Sn = 2 V [r, a (k (r ), rφ) , u] rdφ − R A rdφ − C [k (r )]
k (r )/r 0
2πr
− {V [r, a (k(r ), rπ/n(r )) , u] − R A } (4.2)
n (r )

An increase in the number of parks has two effects. It first adds a new annulus strip
(first line). This generates a net surplus that balances the worker’s net creation of re-
sources, the cost of land and the maintenance cost. Second, it reduces the size of existing
annulus strips (second line). There are 2n(r ) annulus strips that are each reduced by a
length (d/dr ) (rπ/n(r )) = −rπ/n2 (r ). Multiplying those two figures yields the first
term on the second line. The loss on those strips is proportional the land value created
by workers V in addition to agricultural land value R A (the second term on the second
line). The optimal number of parks requires that Sn =0.
Further analysis of the planner’s problem requires characterizing the integrals in
the expression of the above first-order conditions. This requires further specification of
the functions for the utility, amenity and maintenance cost of parks.

4.2.3 Cobb-Douglas preferences

In this subsection, we assume that residents are endowed with the Cobb-Douglas utility
function U (z, s, a) = aα z1− β s β , where α and β measure the preferences for amenities
and residential space (α, β, α + β ∈ (0, 1)). The amenity function is given by a(k, d) =
a0 k ϕ d−δ , where a0 , ϕ, and δ are positive parameters. The maintenance cost is given
by C (k ) = c0 kγ , where c0 and γ are positive parameters. To make the analysis more
realistic, we restrict attention to the case in which the impact of park amenities α is small
enough compared to the impact of residential space: α < β. We also suppose that park
amenities are proportional to their size but that residents are more sensitive to distance
4.2. THE MODEL 93

to their park (δ > ϕ). We finally further suppose that the maintenance cost is a concave
function of park size, implying decreasing returns to scale in park maintenance (γ < 1).
Increasing returns may be justified in several ways. The park infrastructure must be set
up once and may not be proportional to surface (fences, paths, road, kiosk, restaurants,
and so forth). Park security also has decreasing returns to scale. The second assumption
entices the planner to reduce the distance between residents and parks, while the last
assumption entices the planner to reduce the number of parks.
For the sake of exposition, we concentrate on the annulus at distance r and formu-
late the functions n(r ), k (r ) and y(r ) with reference to r. We compute

1
− 1−β β − 1−α β
z̃(s, a, u) = s a u 1− β
1− β 1
− β1 α
V (r, a, u) = β (1 − β) β yβu aβ
− 1− β − 1−β β − αβ 1
where the optimal residential surface associated with V is ŝ = (1 − β) β y a uβ.
The first-order conditions (4.1) and (4.2) simplify to
 πr 1−δ αβ β δ

β δ

1 c0 γ
= 1+ − + − ( − ) α
−( − 1 )
(4.3)
nk ϕα ϕ ϕα ϕ k ϕ δ β γ V (r, a0 , u)
 πr 1−δ αβ β

β

1 c0
= + − 1 ( ϕ−δ) α −(γ−1) (4.4)
nk δα δα k β V (r, a0 , u)
In Appendix A, we show that if
γδ
1− ϕ α
ζ≡ > 0 and γ < 1 + ( ϕ − δ) (4.5)
δ
ϕ −1 β

the optimal park size and number of parks have the following interior and closed-form
solution:
  1
ζc0 (1−γ)+( ϕ−δ) α
∗ β
k =
V (r, a0 , u)
κπr
n∗ = ∗
k
where
   − 1 α
β β 1 1− δ
β
κ≡ + −1
δα δα ζ
Since (1 − γ) + ( ϕ − δ)α/β > 0 by (4.5), the optimal park size increases in the main-
tenance cost c0 and decreases in land value V (r, a0 , u). Since Vr < 0, the optimal park
size also increases in distance to the city centre r. Furthermore, the number of parks is
proportional to r/k∗ , which is a non-monotone function of the distance to the CBD: it first
increases for small r and then decreases for sufficiently large r. For this interior solution,
the share of park area on the annulus is given by
 − 1 α
n∗ k∗
 
β β 1 1− δ
β
= + −1 <1
πr δα δα ζ
94 CHAPTER 4. GEOGRAPHICAL STRATIFICATION OF GREEN URBAN AREAS

which is a constant independent of the distance to the city centre. The distance between
two parks is given by
  1
∗ 2πr 2 ∗ 2 ζc0 (1−γ)+(1−δ) α
β
d = ∗ = k = (4.6)
n κ κ V (r, a0 , u)

and is increasing in the distance from the city centre r. Finally, note that since V rises
with w, richer cities include more numerous parks with smaller sizes and longer access
distances.
Interior optimal solutions take place under conditions (4.5), which can be restated
as δ > ϕ and γ < min{ ϕ/δ, 1+ ( ϕ − δ) α/β}. The first condition requires that the
park amenity decreases proportionally faster with its access than its size. This entices
the planner to set up more parks. The second condition requires sufficiently increasing
returns to scale in maintenance, which entices the planner to set up fewer parks. Under
those conditions, the planner balances the two effects by implementing a finite number
and size of parks (1 < n∗ < ∞ and k∗ > 0). When those conditions are not satisfied, the
planner implements corner solutions with either infinitely many small parks (n∗ → ∞,
k∗ → 0) or a single large park (n∗ = 1).

Proposition 1 Assume Cobb-Douglas preferences and suppose that park amenities are propor-
tional to their size but that residents are more sensitive to distance to their park (δ > ϕ) and
that maintenance costs have decreasing returns to scale in park maintenance (γ < 1). Then,
under conditions (4.5) the optimal number of parks is a bell-shaped function of the distance to
the CBD, while park size and the distance between parks increases with it. More productive
cities include more numerous parks with smaller sizes and longer access distances.

To further characterize our theoretical predictions, we finally assume an exponential

commuting cost t(r ) = w(1 − e−tr ). In this case, taking the natural logarithm after
replacing the land value equation V, we obtain the following optimal number of parks:
  ϑ 
1− β
  β (1 − β )   ϑ
β
 α

− β1 β
ln n = ln κπ  + ln w + ln r − ϑtr + ϑ ln u a0 (4.7)
ζc0 β

where ϑ ≡ [(1 − γ) + ( ϕ − δ)α/β]−1 . The non-monotonic relationship between the

number of parks and distance to the city centre is reflected by the bundle of terms
ln r − ϑtr. There are two distinct channels through which the distance to the city centre
affects the number of green urban spaces. The first term in ln r indicates the positive ef-
fect of an increase in land availability on the number of parks as one moves farther from
the city core: more available land entices the planner to set a larger number of parks.
The second term (−ϑtr ) shows the effect of the commuting cost and population density
on the number of parks. As one moves away from the city centre, the commuting cost
increases, and the attractiveness of land and therefore population density decrease. The
4.3. EMPIRICAL ANALYSIS 95

city planner then has less incentive to provide many parks. Those two effects are oppo-
site in sign, and together form the non-monotone relationship between the number of
green urban spaces and the distance r. At a small r, the effect of land availability dom-
inates, leading to an increase in n, while at large r, the effect of population commuting
and density dominates and induces a decrease in n.6
Similarly, the optimal park size is written as
 
 
ζc ϑ ϑ 1 α
0  − ln w + tr − ϑ ln u a − β
ln k = ϑ ln  1− β
β
0 (4.8)
β (1 − β ) β β β

which increases with distance from the city centre and decreases with city per-capita
income. The optimal distance between parks is given by
 
   
2 ζc0 ϑ ϑ α
− β1 β
ln d = ln + ϑ ln  1− β
 − ln w + tr − ϑ ln u a0 (4.9)
κ β (1 − β ) β β β

which rises with distance from the city centre and falls with city wages. Finally, we can
discuss the spatial pattern of population density for any given configuration of green
urban space systems {k(r ), d(r )}. The logarithm of population density is written as

1−β 1−β
 
1− β α
αϕ αδ 1
ln λ = − ln ŝ = ln (1 − β) β β
a0 + ln w − tr + ln k − ln d − ln u
β β β β β
(4.10)
It increases with larger park sizes and smaller distance to parks because the these fac-
tors augment park amenities and attract more residents. Relationships (4.7) to (4.10) are
used as guidance for the empirical analysis in the next section.

4.3 Empirical analysis

4.3.1 Data

To investigate the stratification of green urban spaces across cities in Europe, we use
the Global Monitoring Environment Services (GMES) Urban Atlas from the European
Environment Agency (EEA) for the period 2006-2008.
The GMES Urban Atlas provides reliable, inter-comparable, and high-resolution
land use maps for 305 Large Urban Zones and their surroundings by collating satellite
images with 2.5 m spatial resolution data. To the best of our knowledge, GMES Urban
Atlas is the most detailed land use map for the urban areas in Europe, which provides
6 Notethat the above non-monotone relationship is not driven by the assumption of an exponential
commuting cost. Indeed, under the assumption of a linear commuting cost, we may write the net income
as y = w − tr, so that the second to fourth terms in (4.7) is written as ϑβ ln (w − tr ) + ln r, which is also a
concave function of r.
96 CHAPTER 4. GEOGRAPHICAL STRATIFICATION OF GREEN URBAN AREAS

Figure 3: Maps of most populous urban centres in Europe 2006-2008

GMES Urban Atlas 2006 urban boundary, EU Coastal and country boundary were used to made this maps. GMES 2006
database covers 305 most populous urban centres in 27 EU member states including Austria, Belgium, Bulgaria, Cyprus,
Czechia, Denmark, Estonia, Finland, France, Germany, Greece, Hungary, Ireland, Italy, Latvia, Lithuania, Luxembourg,
Malta, Netherlands, Poland, Portugal, Romania, Slovakia, Slovenia, Spain, Sweden, and United Kingdom.

an advantage over the CORINE Land Cover maps commonly used in the literature.7
The datasets cover all Functional Urban Areas (FUAs) with at least 100, 000 inhabitants8
Figure 3 displays the urban areas covered by GMES Urban Atlas 2006.
The Urban Atlas provides a classification of city zones that allows for a comparison
of the density of residential areas, commercial and industrial zones, and the extent of
green areas. In this paper, we use the data on ”green urban areas” (class 14100), which
are defined as artificial non-agricultural vegetated areas. They consist of areas with
planted vegetation that is regularly worked and/or strongly influenced by humans.
Precisely, first, green urban areas include public green areas used predominantly for
recreational use (gardens, zoos, parks, castle parks, cemeteries, etc.). Second, suburban
natural areas that have become and are managed as urban parks are included as green
urban areas. Finally, green urban areas also include forest and green areas that extend
from the surroundings into urban areas with at least two sides being bordered by ur-
ban areas and structures and containing visible traces of recreational use. Important
7 GMESmaps have a 100 times higher resolution than CORINE Land Cover.
8 See
the definitions in the Urban Audit database and European Environmental Agency, GMES Ur-
ban Atlas metadata. Link https://land.copernicus.eu/local/urban-atlas/urban-atlas-2006?tab=
metadata (accessed on June 03, 2019).
4.3. EMPIRICAL ANALYSIS 97

for our study, green urban areas do not include private gardens within housing areas,
buildings within parks, such as castles or museums, patches of natural vegetation or
agricultural areas enclosed by built-up areas without being managed as green urban
areas. It must be noted that green urban areas belong to the Urban Atlas’ class of ”ar-
tificial surfaces”, which include all non-agricultural land devoted to human activities.9
This class is distinguished from agricultural, semi-natural areas and wetlands, forest
areas and water areas devoted to non-urban activities. From this section on, we use the
terms green urban spaces, urban parks, and parks interchangeably.
We select the (oldest) town hall locations as the CBDs.10 Then, we create a set of
circles around each CBD at 100 m intervals. For each circle, we create a buffer zone of
250 m, and we define the ”annulus land area” as the intersection of the buffer zones
and the land within the urban zone area reported by the GMES. This area includes
artificial surfaces, agriculture, semi-natural areas, wetlands, and forest but does not
include water areas because those seas and oceans are not appropriate for potential
human dwellings. We then choose all the green areas located within the annulus and
record their centroids, surface areas (sizes) and the closest distances to other green ar-
eas. Figure 4 displays the annuli and the land use of green urban areas (green colour)
for Paris.
In general, urban models assume a clear separation between urban and non-urban
zones. However, urban data do not provide a neat division between residential land
use and agricultural and rural zones at each distance to the city core. Therefore, in this
paper, we follow Picard and Tran (2019) and choose the city boundaries to the annu-
lus at which the share of residential surfaces over the annulus land areas falls below
20%. Residential surfaces include urban areas with dominant residential use, inner-city
area with the central business district and private buildings. Because the residential
land ratio has high frequency variations around potential city borders, we employ the
Nadaraya-Watson Gaussian Kernel11 to smooth those variations before setting up the
city borders. Alternative thresholds do not qualitatively alter the results below.
We measure the distance between two parks as the distance between two park cen-
troids. One thing we want to avoid is the case in which a road separates a park into two
separate polygons. We apply the following procedures to identify continuous green ur-
ban spaces. First, we compute the distance between every two pairs of green urban
centroids. We further assume a circular shape for the park and derive its radius using
9 In addition to green urban areas, artificial surfaces include urban areas with dominant residential use,
inner-city areas with central business district and residential use, industrial, commercial, public, military
and private units, transport units, mines, dumps, and construction sites, and sports and leisure facilities.
10 The town hall is commonly used in the literature as the location of the CBD in European cities because

most of the cities in Europe were formed many centuries  ago.

n n
 
11 The Nadaraya-Watson estimate is given by ∑ K ( x − x i ) gi ∑ K ( x − xi ) where K ( x ) =
i =1 i =1
2 2
exp(− x /(2h ))
√ . Here, we choose bandwidth h = 2 (km). For further details, see Hansen (2009).
(2π )h
98 CHAPTER 4. GEOGRAPHICAL STRATIFICATION OF GREEN URBAN AREAS

Figure 4: Urban Green Areas in Paris

the circle area formula. Second, we obtain the distances between the two boundaries
of green urban spaces by differencing the distance between two park centroids and the
sum of two park radii. If this distance is less than ρ = 5 meters, we consider these
two parks to be adjacent parks and combine them into one park.12 The centroid of the
new combined park polygon is recalculated as the middle point of the two old cen-
troids, and the size becomes the sum of those two areas. We illustrate the criteria for
adjacent parks in Figure 5 below. Once the distinguished polygon and centroid of each
park are determined, we denote dmm0 as the distance between two parks m and m0 :
p
dmm0 ≡ ( xm − xm0 )2 + (ym − ym0 )2 , where ( xm , ym ) is the coordinate of the centroid of
park m. We compute the radius of each park and the distance to other parks as follows:

r
surface of park m
Park Radius k m =
π
Closest Distance dm = min dmm0 ∀m0 6= m
m0
!−1/ψ
1 N
N m∑
(dmm0 )−ψ
ψ
CES Distance dm = where ψ > 0
0 =1

where N is the total number of parks located within urban areas. The value of ψ can
12 The threshold for ρ here is equal to 5 meters, which approximates the width of the roads that cross
parks. We considered alternative the thresholds such as 20, −5, and −20 meters. Our results are robust to
the consideration of these thresholds.
4.3. EMPIRICAL ANALYSIS 99

Figure 5: Illustration of Merger of Adjacent Parks

change to any positive number, and note that when ψ −→ ∞, the CES distance dm
ψ

coincides with the closest distance dm = min

0
dmm0 = d∞
m.
m
In addition to GMES Urban Atlas, we use the population density from the European
population grid.13 We calculate the population mass at each distance to the city centre
and redistribute the population to the residential area14 in each annulus. As the Eu-
rostat population grid does not cover Cyprus, we drop the city of Lefkosia in Cyprus.
For the city income level, we utilize the GDP per capita measures in standard purchas-
ing power at the NUTS3 level from the Eurostat’s Regional Economic Accounts, which
give the finest detail on household income and GDP per capita for Europe.15 Other
measures of cities’ exogenous geographical characteristics are taken from the E-OBS
database from the EU-FP6 project, which covers the entire range of daily temperature,
daily precipitation, sea level pressure and the elevation across Europe at a 1 km square
raster. We finally measure the city population as the number of inhabitants (in mil-
lions) living in the City and Greater City (CGC) defined and reported in the Eurostat
database.16
13 For further information on the European population grid, please check the technical report of
the GEOSTAT 1A project from Eurostat. Link: http://inspire.jrc.ec.europa.eu/documents/Data_
Specifications/INSPIRE_Specification_GGS_v3.0.1.pdf (accessed on June 09, 2019).
14 This is the urban fabric category in GMES Urban Atlas.
15 If there exists more than one NUTS3 area for one annulus, we average the values. For the differences

between NUTS2 and NUTS3 levels, see details at Eurostat, link https://ec.europa.eu/eurostat/web/
nuts/background (accessed on July 03, 2019).
16 For further details, see the metadata file for urb esms on the Urban Audit database.
100 CHAPTER 4. GEOGRAPHICAL STRATIFICATION OF GREEN URBAN AREAS

4.3.2 Descriptive statistics

Table 4.1 provides the summary statistics of green urban spaces in European cities.
Our sample includes the 303 most populous cities in 27 countries in Europe.17 On
average, cities have approximately 196 green urban spaces, with a total area of 8.58
square kilometres per city. Catanzaro on the south coast of Italy is the city with the
smallest number of green urban spaces, and London, U.K., is the city with the highest
number of green urban spaces with more than 2000 parks. The city with the smallest
total surface of green urban spaces is the city of Konin in Poland, while that with the
highest surface is Paris, France, with nearly 300 square kilometres of green area. The
average size of a green urban is approximately 3.42 hectares per park. The city with
the smallest average park size is still Konin, and that with the largest average park
size is Madrid, Spain, with 25 hectares per park on average. Figure 6 displays the size
distribution of all green urban spaces across the 303 most populous cities in Europe.
Note that the largest 5% of parks account for more than 50% of the total surface of
green areas. This pattern is also similar within each urban area, with an even higher
concentration in larger cities such as Madrid or Paris.

Figure 6: Urban Green Size Distribution within Urban Area

City size affects the surface and distribution of green urban spaces. To illustrate
the difference in the park size distribution across cities, we calculate the Gini index
of the size of green urban spaces in each city. A low level of the Gini index indicates
homogeneity in park size, while a high coefficient indicates high heterogeneity in park
17 We
exclude Luxembourg and Cyprus as they are relatively small, and we do not have information at
the NUTS3 level on household income and population grid.
4.3. EMPIRICAL ANALYSIS 101

size. The average Gini index is 0.61, with a standard deviation of 0.11, and the city
with the highest level of the Gini index of 0.93 is Madrid, which has many large and
small parks. Table 4.2 presents correlations between the total park surface and park
distribution on the one hand and proxies for city size on the other hand (GDP per capita,
border, and population). They show that more populous and richer cities are associated
with larger park surfaces and greater park size heterogeneity. Such correlations are
smaller if we exclude the first and last percentiles of park size. In this case, total green
surfaces are almost proportional to population sizes (the coefficient is 1.098).

Table 4.1: Summary Statistics of Urban Green Spaces across European cities

A. City Statistics
Mean Std. Dev. Obs.
City Border (km) 5.68 3.98 303
Population (mil.) 0.45 0.80 303
GDP per capita (e1000/hab./y.) 26.75 12.88 303
Average Rainfall 1.91 0.59 303
Elevation (m) 212.19 211.16 303
Average Temperature in Jan 01 (◦ C) 2.28 4.61 303
Average Temperature in Jul 01 (◦ C) 19.16 3.36 303

B. Urban Green Space Statistics

Full UGS sample Excluding big UGS sample
Mean Std. Dev. Obs. Mean Std. Dev. Obs.
Average Number 196.38 295.08 303 194.41 290.44 303
Average Size (ha) 3.42 1.92 303 2.88 1.17 303
Average Radius (m) 81.19 17.07 303 78.60 14.53 303
Total Green Area (km2 ) 8.58 18.17 303 6.28 10.71 303
Gini Index 0.61 0.11 303 0.58 0.09 303
Average Closest Distance d (m) 336.85 94.59 302 336.42 94.66 302
Average 2nd Closest Distance d (m) 496.55 148.55 302 496.47 148.82 302
Average CES Distance dψ (m)
ψ=2 1726.41 1001.42 302 1722.24 996.99 302
ψ = 10 520.67 144.46 302 519.63 144.07 302
ψ = 50 368.63 101.35 302 368.11 101.37 302
Note: City border is chosen at the distance where the artificial residential area is less than 20% of
total land. The residential land share is smoothed using Gaussian Kernel before the cut-off. The
threshold for big parks is set at 99% quantile, which is at 54.8 hectares, and is computed from
whole samples of all parks within city border of all 303 cities. We excluded the city of Catanzaro,
Italy, when calculating the average minimum distance to other parks as the sample size is too
small in Catanzero.

In the theoretical model, we consider the parks located in the local area of patroniz-
ing residents but not the large parks that provide a global amenity to all residents. We
leave the theoretical investigation of the location of large parks for future research. In
this empirical analysis, we first consider the full sample of parks, but for robustness, we
dedicate special attention to the parks in the lower 99th-percentile set of park size. As
can be seen below, the results are qualitatively and quantitatively unaltered following
this change in dataset.
In this paper, we use the GDP per capita at the NUTS3 level as a baseline proxy
102 CHAPTER 4. GEOGRAPHICAL STRATIFICATION OF GREEN URBAN AREAS

Table 4.2: City Size and Urban Green Spaces across European Cities

Full Sample
Dependent variable:
Gini Index Ln Total Green Surface (km2 )
(1) (2) (3) (4) (5) (6)
Ln City GDP per capita (e1000) 0.126∗∗∗ 2.337∗∗∗
(0.024) (0.246)
Ln City Border (km) 0.125∗∗∗ 2.110∗∗∗
(0.011) (0.077)
Ln City Population (mil.) 0.064∗∗∗ 1.247∗∗∗
(0.008) (0.056)
Country FE Yes Yes Yes Yes Yes Yes
Observations 303 303 303 303 303 303
Adjusted R2 0.139 0.350 0.251 0.384 0.780 0.706
Note: City GDP per capita is measured at NUTS3 level and at e1000 per inhabitant. The table presents
the results of the regression of the log of total surfaces in each city and the Gini index of urban green
sizes within each city. Data: GMES, European cities. Standard errors are reported in parenthesis and
significance levels are denoted by ∗ p<0.1, ∗∗ p<0.05, ∗∗∗ p<0.01.

for city income w. GDP per capita varies significantly across EU cities, from e6,000 to
e83,700 per inhabitant, with an average of e26,750 measuring in terms of purchasing
power standard.18 GDP per capita represents more than 1.5 times the residents’ net in-
come, which reflects the tax wedge between production value and income. City eleva-
tion also varies vastly, from two meters below sea level in Amsterdam, the Netherlands,
to 1,614 meters in the mountainous city of Innsbruck, Austria. The European continent
belongs to a mild climate zone with temperature varying between −8 and +28 degrees
Celsius for the lowest and highest day-time temperatures in winter and summer on
average (measured on January 1 and July 1, respectively, for the period 1995 − 2006).19
The controls for climate capture the possible effects of residents’ preference for nice
weather (e.g., not too cold in winter and not too hot in summer; see Rappoport 2007).

4.3.3 Number of green urban spaces

In this section, we examine the patterns in the numbers of green urban spaces across
locations within the city area. The theoretical model predicts that the number of green
urban spaces is a non-monotone function of the distance to the city centre: it first in-
creases at distances close to CBD and then decreases as one moves closer to the urban
fringe. Using equation (4.7) above, we can test the following reduced-form regression
18 The purchasing power standard unit is robust to the difference in the prices of goods
and services across city and country boundaries. For further details, see Eurostat Statis-
tics Explained Glossary, link https://ec.europa.eu/eurostat/statistics-explained/index.php/
Glossary:Purchasing_power_standard_(PPS) (accessed on June 20, 2019).
19 The samples in our data do not contain some Northern European cities in Iceland and Norway.
4.3. EMPIRICAL ANALYSIS 103

ln nijc = ρ0 + ρ1 ln w jc + ρ2 ln distijc + ρ3 distijc + ∑ Xijc + ∑ X jc + ∑ Ic + eijc

(1) (2)

ijc jc c

where (i, j, c) is an observation of annulus i in city j in region c, nijc is the number of

parks for each annulus i, distijc the distance from the city centre and w jc is city income
(1)
per capita. Xijc are control variables for annuli (proximity of agricultural and forest
(2)
land) and X jc for city observable characteristics (elevation, temperature, rain fall, lat-
itude, etc.). Ic is a region dummy variable controlling for unobservable regional char-
acteristics such as environmental policies or local regulation. Our theoretical model
predicts that the number of parks first increases and then decreases as one moves far-
ther from city centre; hence, the sign of ρ2 is expected to be positive, while ρ3 is expected
to be negative. We also expect that the number of parks increases with city income w;
hence, ρ1 is expected to be positive.
The regression results are reported in Table 4.3. The standard errors are clustered
and robust at the city level, which controls for the possible correlation between obser-
vations within the same urban area. Column (1) includes no controls, while Columns
(2) and (3) successively introduce the city geographical controls and the annulus con-
trols for nearby agriculture and forestry. Columns (4) and (5) add fixed effects at the
country and regional level. Columns (6), (7), (9) and (10) restrict the data to monocen-
tric cities. Columns (8), (9) and (10) exclude large parks by using the parks in the lower
99th percentiles of park sizes. Table 3 shows a positive and significant coefficient for ρ̂2
and a negative and significant coefficient for ρ̂3 across all kinds of controls and datasets.
This confirms the non-monotonic relationship between the number of parks and the distance to
the city centre and the presence of the two above different channels of land availability and pop-
ulation density. Moreover, as expected from the theoretical model, higher city income
increases the number of parks. Note that the presence of agriculture and forest land
controls diminishes the number of parks. This is to be expected, as those substitute for
park amenities.
The magnitudes of ρ̂2 and ρ̂3 reflect the quantitative effects of the channels of land
availability and population density. According to Column (7), a 1% increase in the
distance to the city centre first leads to a 0.776% increase in the number of parks through
the land availability channel, while a one-kilometre increase in distance is associated
with a 7.9% decrease in the number of parks through the population density channel.
To fix ideas, when one moves from 1 kilometre to 2 kilometres from the city centre, the
first channel generates a 79% increase in the number of parks, while the second channel
yields a 7.9% decrease, so the overall number of parks increases by 74%. Therefore,
land availability matters more in locations close to the city centre. However, when one
moves from 20 kilometres to 21 kilometres from city centre, the first channel generates
a 3.8% increase in the number of parks, and the second channel yields a 7.9% decrease,
104 CHAPTER 4. GEOGRAPHICAL STRATIFICATION OF GREEN URBAN AREAS

Table 4.3: Number of Urban Green Spaces across Locations within Urban Area (OLS)
Dependent variable: Ln Number of Urban Green Spaces
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10)
Ln GDP per capita 0.561∗∗∗ 0.478∗∗∗ 0.235∗∗∗ 0.229∗∗∗ 0.147 0.272∗∗ −0.236 0.150 0.269∗ −0.236
(0.065) (0.079) (0.073) (0.076) (0.122) (0.138) (0.148) (0.121) (0.138) (0.147)
Ln Distance to CBD 0.564∗∗∗ 0.564∗∗∗ 0.771∗∗∗ 0.774∗∗∗ 0.794∗∗∗ 0.783∗∗∗ 0.776∗∗∗ 0.788∗∗∗ 0.778∗∗∗ 0.771∗∗∗
(0.048) (0.047) (0.037) (0.033) (0.026) (0.030) (0.028) (0.026) (0.030) (0.028)
Distance to CBD (km) −0.044∗ −0.044∗∗ −0.044∗∗∗ −0.036∗∗∗ −0.046∗∗∗ −0.051∗∗∗ −0.079∗∗∗ −0.046∗∗∗ −0.050∗∗∗ −0.078∗∗∗
(0.023) (0.022) (0.017) (0.014) (0.013) (0.017) (0.014) (0.013) (0.017) (0.015)
Annuli controls
Agriculture Land (%) −5.228∗∗∗ −5.295∗∗∗ −6.119∗∗∗ −6.034∗∗∗ −5.401∗∗∗ −6.095∗∗∗ −6.021∗∗∗ −5.394∗∗∗
(0.512) (0.546) (0.430) (0.457) (0.420) (0.430) (0.457) (0.420)
Forest Land (%) −7.236∗∗∗ −7.114∗∗∗ −4.975∗∗∗ −4.647∗∗∗ −2.840∗∗∗ −4.958∗∗∗ −4.642∗∗∗ −2.848∗∗∗
(0.891) (0.842) (0.910) (0.989) (0.769) (0.911) (0.991) (0.774)
City Samples Full Full Full Full Full Monocentric Monocentric Full Monocentric Monocentric
samples samples samples samples samples cities cities samples cities cities
UGS Samples Full Full Full Full Full Full Full Excld. UGS Excld. UGS Excld. UGS
samples samples samples samples samples samples samples abv. 99th per. abv. 99th per. abv. 99th per.
City Geo. Controls No Yes Yes Yes Yes Yes Yes Yes Yes Yes
FE Controls No No No Country NUTS2 NUTS2 NUTS2 × NUTS2 NUTS2 NUTS2 ×
City Size City Size
Observations 16,782 16,782 16,782 16,782 16,782 14,368 14,368 16,780 14,366 14,366
Adjusted R2 0.350 0.374 0.486 0.547 0.675 0.645 0.679 0.671 0.642 0.676
Note: Here, we use the GDP per capita taken from Regional Economic Accounts from Eurostat at NUTS3 level with adjustment to purchasing power standard
(PPS) as the proxy for city income level, and it is measured in e1,000 per inhabitant. The distance to CBD is measured in kilometre. City geographical controls
includes the average Elevation, Latitude, Longitude and the average and square of the average of the rain fall, the temperature in summer and winter time. The
average rain fall, and average temperature in Jan 01 (winter time) and average temperature in July 01 (summer time) are measured for the period 1995 − 2010.
City boundary is chosen at 20% cut-off point of residential land use share (See details in Picard and Tran (2019)). The thresholds for urban green size at 99th
percentile is 54.82 hectares. Columns (6), (7), (9) and (10) isolate monocentric cities by excluding the polycentric cities reported by OECD. Columns (1) to (7)
include all urban green spaces within urban boundary, while Columns (8) to (10) exclude all big parks above 99th percentile. The Agriculture and Forest Land
from the last two annuli control variables are measured within 100-meter buffer around the residential area. Standard errors are clustered at city level and
reported in parenthesis. The significance levels are denoted by ∗ p<0.1, ∗∗ p<0.05, ∗∗∗ p<0.01.
4.3. EMPIRICAL ANALYSIS 105

so the overall number of parks falls by 4.1%. Thus, the non-monotone relationship is
particularly marked in larger cities.
The regression analysis of the non-linear relationship ln r − ϑtr is driven by the the-
oretical model but is not commonplace in the empirical literature. A standard way to
test for the concavity of the number of parks with distance to the CBD is to assume a
quadratic function of distance. Appendix B shows that the quadratic regression also
yields a concave relationship between the number of parks and the distance to the cen-
tre.
Finally, Table 4.3 also shows a positive and significant sign for the coefficient for the
city income level of ρ̂1 in most cases. Using the regression results from Column (6), a
1% increase in income level is associated with a 0.27% increase in the number of parks
after controlling for city observable characteristics and NUTS2 fixed effects. We also
observe a significant and negative effect for the share of agricultural and forest land
surrounding residences on the number of green urban spaces. Nearby agriculture and
forest land can be seen as a substitute for the amenities of urban parks. A 1% increase
in agriculture (resp. forest) land leads to a reduction of 6.0% (resp. 4.6%) in the number
of parks.

4.3.4 Size of green urban spaces

In the theoretical model, the optimal park sizes rise with distance to the city centre and
fall with city income. Using (4.8), we can run the following regression:

ln k ijc = η0 + η1 ln w jc + η2 distijc + ∑ Xijc + ∑ X jc + ∑ Ic + eijc

(1) (2)
(4.11)
ijc jc c

where k ijc is the average size of the parks located at annulus i of city j of region c, distijc
(1) (2)
is the distance of the annulus to the CBD and Xijc , X jc and Ic are the same control
variables as in the previous subsection. From the theoretical model, we predict that
green urban systems increase in size with distance to the city centre η2 > 0 and decrease
in size with city income η1 < 0.
Table 4.4 reports the regression results with the same column structure as in Table
4.3. First, the effect of distance is positive and significant in all regressions whatever
the controls and data set. This suggests that the size of green urban spaces increases with
distance to the city centre. Notably, for a 1-kilometre increase in the distance to the CBD,
the park radius increases by 2.3% (i.e., by approximately 5% of park area). When we
restrict attention to monocentric cities and omit the large green urban spaces, the mag-
nitude declines slightly to 2.0%. Second, the effect of city income w on park size is not
significant, except when one restricts the dataset to monocentric cities. In that case, the
magnitude of the effect is small; Column (6) suggests that a 1% increase in city income
leads to a 0.15% increase in park radius (i.e., an approximately 0.3% increase in park
106 CHAPTER 4. GEOGRAPHICAL STRATIFICATION OF GREEN URBAN AREAS

area). Finally, it is surprising to observe that the presence of nearby forestry does not
affect park size. Nearby agricultural land also has no effect on monocentric cities.

4.3.5 Distance to other green urban spaces

According to our theoretical prediction, the distance between parks increases with the
distance from the city centre and falls with city income. We run a similar regression
as in the previous sections with the distance between two parks as the explanatory
variable. Precisely, we run the following reduced-form regression:

ln dijc = µ0 + µ1 ln w jc + µ2 distijc + ∑ Xijc + ∑ X jc + ∑ Ic + eijc

(1) (2)
(4.12)
ijc jc c

where dijc is the average of the closest distance to other parks in annulus i of city j in
region c. As above, we expect the sign of the coefficient for the distance to the CBD µ2
to be negative.
Table 4.5 reports the regression results and has the same structure as Tables 4.3 and
4.4. On the one hand, the distance to the closest park also increases as one moves
toward the urban fringe. Notably, a one-kilometre increase in distance to the CBD
leads to a 2.2% increase in the distance to the closest park, as indicated in Column (5)
with NUTS2 region fixed effects. This figure increases slightly when we restrict data to
monocentric cities and further introduce the city size fixed effects. On the other hand,
richer cities are associated with a denser network of green urban spaces, as shown in
Columns (1) to (3). However, the effect becomes smaller and non-significant when we
control for regional fixed effects and city size. Note that the distance between parks
increases in the shares of nearby agricultural and forest land, as the latter is likely to
provide amenities that compensate for longer trips to parks. For instance, a 1% increase
in the land shares of nearby agriculture and forest leads to 1.8% and 1% increases, re-
spectively, of the distance to the closest park.
It is interesting to note the close similarity between the coefficients on distance to the
CBD in Tables 4.4 and 4.5. This suggests a strong consistency with the Cobb-Douglas
model, where by (4.6) the effect on the distance to the closest park is proportional to that
on park size: d ln d/d lnr = d ln k/d lnr. However, the consistency is weaker when we
can compare the coefficients in Tables 4.3 and 4.5. By (4.6), the theoretical model gives
the relationship d ln d/d lnr = 1 − d ln n/d lnr. Empirically, this means to compare
the value 0.024 in Table 4.5 to the value 1 − 0.794 obtained from Table 4.3, Column (5).
Hence, the Cobb-Douglas model fits well the relationship between the distance to the
closest park and park size but fits less well the relationship between the former and the
number of parks.
How do those relationships hold when one considers not only the closest park but
the parks surrounding a residence? To answer this question, we again use the CES
distance dψ with various values of the parameter ψ. As described above, this index
 Table 4.4: Sizes of Urban Green Spaces (OLS)

Dependent variable: Ln Average Urban Green Space Radius

(1) (2) (3) (4) (5) (6) (7) (8) (9) (10)
Ln GDP per capita 0.031 −0.001 −0.019 0.002 0.047 0.151∗∗∗ 0.145∗∗ 0.049 0.129∗∗∗ 0.141∗∗∗
(0.024) (0.025) (0.024) (0.029) (0.056) (0.044) (0.057) (0.047) (0.040) (0.051)
Distance to CBD (km) 0.022∗∗∗ 0.022∗∗∗ 0.026∗∗∗ 0.025∗∗∗ 0.023∗∗∗ 0.023∗∗∗ 0.021∗∗∗ 0.021∗∗∗ 0.020∗∗∗ 0.020∗∗∗
(0.002) (0.002) (0.002) (0.003) (0.003) (0.004) (0.004) (0.003) (0.003) (0.004)
Annuli controls
4.3. EMPIRICAL ANALYSIS

Agriculture Land (%) −0.460∗∗∗ −0.513∗∗∗ −0.397∗∗ −0.176 −0.163 −0.371∗∗∗ −0.188 −0.193
(0.155) (0.166) (0.165) (0.153) (0.159) (0.139) (0.129) (0.134)
Forest Land (%) −0.370 −0.388 0.283 0.224 0.313 0.284 0.180 0.232
(0.338) (0.314) (0.323) (0.318) (0.303) (0.296) (0.287) (0.263)
City Samples Full Full Full Full Full Monocentric Monocentric Full Monocentric Monocentric
samples samples samples samples samples cities cities samples cities cities
UGS Samples Full Full Full Full Full Full Full Excld. UGS Excld. UGS Excld. UGS
samples samples samples samples samples samples samples abv. 99th per. abv. 99th per. abv. 99th per.
City Geo. Controls No Yes Yes Yes Yes Yes Yes Yes Yes Yes
FE Controls No No No Country NUTS2 NUTS2 NUTS2 × NUTS2 NUTS2 NUTS2 ×
City Size City Size
Observations 16,782 16,782 16,782 16,782 16,782 14,368 14,368 16,780 14,366 14,366
Adjusted R2 0.070 0.158 0.163 0.202 0.411 0.430 0.435 0.413 0.429 0.435
Note: Here, we use the GDP per capita taken from Regional Economic Accounts from Eurostat at NUTS3 level with adjustment to purchasing power
standard (PPS) as the proxy for city income level, and it is measured in e1,000 per inhabitant. The distance to CBD is measured in kilometre. City
geographical controls includes the average Elevation, Latitude, Longitude and the average and square of the average of the rain fall, the temperature
in summer and winter time. The average rain fall, and average temperature in Jan 01 (winter time) and average temperature in July 01 (summer time)
are measured for the period 1995 − 2010. City boundary is chosen at 20% cut-off point of residential land use share (See details in Picard and Tran
(2019)). The thresholds for urban green size at 99th percentile is 54.82 hectares. Columns (6), (7), (9) and (10) isolate monocentric cities by excluding
the polycentric cities reported by OECD. Columns (1) to (7) include all urban green spaces within urban boundary, while Columns (8) to (10) exclude
all big parks above 99th percentile. The Agriculture and Forest Land from the last two annuli control variables are measured within 100-meter buffer
around the residential area. Standard errors are clustered at city level and reported in parenthesis. The significance levels are denoted by ∗ p<0.1,
∗∗ p<0.05, ∗∗∗ p<0.01.
107
108 CHAPTER 4. GEOGRAPHICAL STRATIFICATION OF GREEN URBAN AREAS

Table 4.5: Closest Distance to Other Urban Green Spaces (OLS)

Dependent variable: Ln Average Closest Distance to other parks dmin
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10)
Ln GDP per capita −0.124∗∗∗ −0.130∗∗∗ −0.068∗∗ −0.049 −0.057 −0.036 0.073 −0.055 −0.036 0.072
(0.027) (0.031) (0.031) (0.034) (0.056) (0.068) (0.068) (0.056) (0.068) (0.068)
Distance to CBD (km) 0.040∗∗∗ 0.040∗∗∗ 0.028∗∗∗ 0.023∗∗∗ 0.022∗∗∗ 0.023∗∗∗ 0.031∗∗∗ 0.021∗∗∗ 0.023∗∗∗ 0.031∗∗∗
(0.005) (0.005) (0.005) (0.004) (0.005) (0.006) (0.006) (0.005) (0.006) (0.006)
Annuli controls
Agriculture Land (%) 1.549∗∗∗ 1.553∗∗∗ 1.785∗∗∗ 1.862∗∗∗ 1.687∗∗∗ 1.814∗∗∗ 1.890∗∗∗ 1.713∗∗∗
(0.218) (0.220) (0.209) (0.228) (0.223) (0.210) (0.228) (0.223)
Forest Land (%) 1.446∗∗∗ 1.384∗∗∗ 0.991∗∗ 1.056∗∗∗ 0.483 0.963∗∗ 1.022∗∗ 0.450
(0.423) (0.406) (0.388) (0.406) (0.342) (0.388) (0.406) (0.341)
City Samples Full Full Full Full Full Monocentric Monocentric Full Monocentric Monocentric
samples samples samples samples samples cities cities samples cities cities
UGS Samples Full Full Full Full Full Full Full Excld. UGS Excld. UGS Excld. UGS
samples samples samples samples samples samples samples abv. 99th per. abv. 99th per. abv. 99th per.
City Geo. Controls No Yes Yes Yes Yes Yes Yes Yes Yes Yes
FE Controls No No No Country NUTS2 NUTS2 NUTS2 × NUTS2 NUTS2 NUTS2 ×
City Size City Size
Observations 16,782 16,782 16,782 16,782 16,782 14,368 14,368 16,780 14,366 14,366
Adjusted R2 0.164 0.184 0.236 0.309 0.452 0.454 0.478 0.453 0.455 0.479
Note: Here, we use the GDP per capita taken from Regional Economic Accounts from Eurostat at NUTS3 level with adjustment to purchasing power
standard (PPS) as the proxy for city income level, and it is measured in e1,000 per inhabitant. The distance to CBD is measured in kilometre. City
geographical controls includes the average Elevation, Latitude, Longitude and the average and square of the average of the rain fall, the temperature
in summer and winter time. The average rain fall, and average temperature in Jan 01 (winter time) and average temperature in July 01 (summer time)
are measured for the period 1995 − 2010. City boundary is chosen at 20% cut-off point of residential land use share (See details in Picard and Tran
(2019)). The thresholds for urban green size at 99th percentile is 54.82 hectares. Columns (6), (7), (9) and (10) isolate monocentric cities by excluding
the polycentric cities reported by OECD. Columns (1) to (7) include all urban green spaces within urban boundary, while Columns (8) to (10) exclude
all big parks above 99th percentile. The Agriculture and Forest Land from the last two annuli control variables are measured within 100-meter buffer
around the residential area. Standard errors are clustered at city level and reported in parenthesis. The significance levels are denoted by ∗ p<0.1,
∗∗ p<0.05, ∗∗∗ p<0.01.
4.3. EMPIRICAL ANALYSIS 109

Table 4.6: Closest Distance dmin and CES Distance dψ

Dependent variable:
Ln Closest Dist. Ln CES Distance dψ
dmin ψ=2 ψ=5 ψ = 10 ψ = 15 ψ = 20 ψ = 50
(1) (2) (3) (4) (5) (6) (7)
Ln GDP per capita −0.057 0.483∗∗∗ 0.205 ∗∗∗ 0.077 0.033 0.010 −0.030
(0.056) (0.086) (0.062) (0.055) (0.055) (0.055) (0.055)
Distance to CBD (km) 0.022∗∗∗ 0.066∗∗∗ 0.041∗∗∗ 0.032∗∗∗ 0.028∗∗∗ 0.027∗∗∗ 0.024∗∗∗
(0.005) (0.005) (0.004) (0.004) (0.004) (0.004) (0.004)
Annuli controls
Agriculture Land (%) 1.785∗∗∗ 0.717∗∗∗ 1.236∗∗∗ 1.514∗∗∗ 1.607∗∗∗ 1.653∗∗∗ 1.733∗∗∗
(0.209) (0.211) (0.207) (0.206) (0.207) (0.207) (0.208)
Forest Land (%) 0.991∗∗ −0.568 0.160 0.546 0.687∗ 0.760∗∗ 0.895∗∗
(0.388) (0.392) (0.328) (0.346) (0.358) (0.364) (0.378)
City Samples Full Full Full Full Full Full Full
UGS Samples Full Full Full Full Full Full Full
City Geo. Controls Yes Yes Yes Yes Yes Yes Yes
FE Controls NUTS2 NUTS2 NUTS2 NUTS2 NUTS2 NUTS2 NUTS2
Observations 16,782 16,782 16,782 16,782 16,782 16,782 16,782
Adjusted R2 0.445 0.843 0.611 0.499 0.472 0.461 0.448
Note: Here, we use the GDP per capita taken from Regional Economic Accounts from Eurostat at NUTS3
level with adjustment to purchasing power standard (PPS) as the proxy for city income level, and it is
measured in e1,000 per inhabitant. The distance to CBD is measured in kilometre. City geographical
controls includes the average Elevation, Latitude and Longitude. The average rain fall, and average
temperature in Jan 01 (winter time) and average temperature in July 01 (summer time) for period 1995 −
2010. City boundary is chosen at 20% cut-off point of residential land use share (See details in Picard and
Tran (2019)). The table shows results from the regression of number of urban green spaces at each level
of distance to city centre across 303 most populous cities in Europe. The Agriculture and Forest Land
from the last two annuli control variables are measured within 100-meter buffer around the residential
area. Standard errors are clustered at city level and reported in parenthesis. The significance levels are
denoted by ∗ p<0.1, ∗∗ p<0.05, ∗∗∗ p<0.01.

includes all parks but assigns higher weights to closer parks as ψ rises. As ψ approaches
∞, the index tends to the value of the distance to closest park dmin. The first column
in Table 6 reports the coefficient estimates for the measure dmin of the distance to the
closest park measure dmin, while the next six columns report the estimates for the CES
distance at various values of ψ. The second line of this table shows the coefficient
estimates of the effect of the distance to the city centre. As one can see, the effect of
distance to the CBD is positive and significant for all measures of distance between
green urban spaces. The distance between parks increases with the distance from the
city centre. This confirms and extends our theoretical prediction in the context where
residents visit more than one park. Finally, it can be noted that the coefficient estimate
on the distance to the city centre decreases in ψ. This stems from the fact that δψ is
an increasing function of distance to the city centre. To understand this, consider the
situation in which parks are uniformly spread across the city space. Then, residents
located at the city centre would have a lower average distance to parks than those
dwelling at the city border. Such a difference diminishes with a higher value of ψ as
more weight is placed on the (equally distant) closest parks.
110 CHAPTER 4. GEOGRAPHICAL STRATIFICATION OF GREEN URBAN AREAS

4.3.6 Population density

In this subsection, we test the relationship between population density and the system
of green urban spaces. Theory tells us that population density increases with larger
park sizes and smaller distance to parks, as the latter augment the park amenities and
attract more residents. From (4.10), we obtain the following reduced-form regression
model:

ln λijc = θ0 + θ1 ln w jc + θ2 distijc + θ3 ln k ijc + θ4 ln dijc + ∑ Xijc + ∑ X jc + ∑ Ic + eijc

(1) (2)

ijc jc c
(4.13)
where k ijc and dijc are the average of park radius (or park size) and the distance to the
(1) (2)
closest park in annulus i of city j and region c. Xijc , X jc and Ic are the same control
variables as before. We expect that θ3 > 0 and θ4 < 0. In the theoretical model, we
assume that park amenities rise with park size, but residents are more sensitive to the
distance to their park (δ > ϕ). If that statement is correct, we should expect the magni-
tude of the distance effect |θ4 | to be greater than that of park size |θ3 |. The model also
predicts that population density increases with income w and decreases with distance
to the city centre; hence, we expect a negative sign for θ2 and a positive sign for θ1 .
The regression results are reported in Table 4.7 with standard errors clustered at the
city level. The effect of the distance to the city centre on population density is robust
and significant with all control variables included. Notably, the results in Column (5),
which include the whole sample of green urban spaces and all necessary controls and
region fixed effects, show that a 1-kilometre increase in distance r to the CBD lessens
the population density by 5.4%. This figure rises to 6.2% and 7.3% when we consider
only the sub-sample of all monocentric cities20 and add city size fixed effects as shown
in Columns (6) and (7). Those numbers are robust when we exclude all large green
urban spaces above the 99th percentile. The coefficients for city income level w are also
positive across all regressions with varying levels of controls. Specifically, a 1% increase
in income boosts the population density by 0.15-0.35%, ceteris paribus. Overall, the
signs for θ̂ 1 and θ̂ 2 match the prediction from the theoretical model and are robust to
different levels of controls.
Table 4.7 also presents an interesting result about the effect of green urban systems
on population density. Precisely, the effect of the shortest distance to a green urban
space d is significant and economically relevant, whereas the effect of average park
20 We exclude all polycentric cities reported in the OECD database. The list comprises Praha (CZ), Brno
(CZ), Ostrava (CZ), Frankfurt am Main (DE), Stuttgart (DE), Madrid (ES), Barcelona (ES), Bilbao (ES), Paris
(FR), Lyon (FR), Lille (FR), Rennes (FR), Napoli (IT), Amsterdam (NL), Eindhoven (NL), Lisboa (PT), Stock-
holm (SE), Malmo (SE), London (UK), Birmingham (UK), Glasgow (UK), Newcastle upon Tyne (UK) and
Portsmouth (UK). Further details are available at https://stats.oecd.org/Index.aspx?Datasetcode=
CITIES.
 Table 4.7: Population Density and Urban Green Stratification (OLS)

Dependent variable: Ln Population Density

Base Regressions Robustness
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10)
Ln GDP per capita 0.150∗∗ 0.330∗∗∗ 0.192∗∗∗ 0.334∗∗∗ 0.278∗∗∗ 0.349∗∗∗ 0.148 0.278∗∗∗ 0.348∗∗∗ 0.148
(0.059) (0.052) (0.048) (0.059) (0.104) (0.085) (0.106) (0.103) (0.085) (0.105)
Dist. to CBD (km) −0.058∗∗∗ −0.059∗∗∗ −0.030∗∗∗ −0.024∗∗∗ −0.054∗∗∗ −0.062∗∗∗ −0.073∗∗∗ −0.054∗∗∗ −0.062∗∗∗ −0.073∗∗∗
(0.009) (0.009) (0.008) (0.006) (0.007) (0.009) (0.010) (0.007) (0.009) (0.010)
Ln UGS Radius k −0.007 0.103∗∗ −0.001 −0.037 −0.029 −0.036 −0.038 −0.032 −0.036 −0.034
(0.044) (0.043) (0.031) (0.026) (0.026) (0.027) (0.026) (0.029) (0.029) (0.027)
4.3. EMPIRICAL ANALYSIS

Ln Closest Dist. dmin −0.432∗∗∗ −0.411∗∗∗ −0.217∗∗∗ −0.143∗∗∗ −0.136∗∗∗ −0.140∗∗∗ −0.103∗∗∗ −0.138∗∗∗ −0.143∗∗∗ −0.106∗∗∗
(0.048) (0.046) (0.035) (0.028) (0.024) (0.027) (0.027) (0.024) (0.027) (0.027)
Annuli controls
Agriculture Land (%) −3.908∗∗∗ −4.315∗∗∗ −2.948∗∗∗ −2.833∗∗∗ −2.637∗∗∗ −2.943∗∗∗ −2.826∗∗∗ −2.631∗∗∗
(0.328) (0.305) (0.342) (0.366) (0.372) (0.342) (0.366) (0.371)
Forest Land (%) −5.592∗∗∗ −4.640∗∗∗ −4.398∗∗∗ −3.935∗∗∗ −3.551∗∗∗ −4.400∗∗∗ −3.939∗∗∗ −3.556∗∗∗
(0.748) (0.659) (0.558) (0.573) (0.666) (0.557) (0.573) (0.665)
City Samples Full Full Full Full Full Monocentric Monocentric Full Monocentric Monocentric
samples samples samples samples samples cities cities samples cities cities
UGS Samples Full Full Full Full Full Full Full Excld. UGS Excld. UGS Excld. UGS
samples samples samples samples samples samples samples abv. 99th per. abv. 99th per. abv. 99th per.
City Geo. Controls No Yes Yes Yes Yes Yes Yes Yes Yes Yes
FE Controls No No No Country NUTS2 NUTS2 NUTS2 × NUTS2 NUTS2 NUTS2 ×
City Size City Size
Observations 16,782 16,782 16,782 16,782 16,782 14,368 14,368 16,780 14,366 14,366
Adjusted R2 0.290 0.404 0.552 0.679 0.787 0.783 0.794 0.787 0.783 0.794
Note: Here, we use the GDP per capita taken from Regional Economic Accounts from Eurostat at NUTS3 level with adjustment to purchasing power
standard (PPS) as the proxy for city income level, and it is measured in e1,000 per inhabitant. The distance to CBD is measured in kilometre. City
geographical controls includes the average Elevation, Latitude, Longitude and the average and square of the average of the rain fall, the temperature
in summer and winter time. The average rain fall, and average temperature in Jan 01 (winter time) and average temperature in July 01 (summer time)
are measured for the period 1995 − 2010. City boundary is chosen at 20% cut-off point of residential land use share (See details in Picard and Tran
(2019)). The thresholds for urban green size at 99th percentile is 54.82 hectares. Columns (6), (7), (9) and (10) isolate monocentric cities by excluding
the polycentric cities reported by OECD. Columns (1) to (7) include all urban green spaces within urban boundary, while Columns (8) to (10) exclude
all big parks above 99th percentile. The Agriculture and Forest Land from the last two annuli control variables are measured within 100-meter buffer
around the residential area. Standard errors are clustered at city level and reported in parenthesis. The significance levels are denoted by ∗ p<0.1,
∗∗ p<0.05, ∗∗∗ p<0.01.
111
112 CHAPTER 4. GEOGRAPHICAL STRATIFICATION OF GREEN URBAN AREAS

size k is much smaller and, in addition, non-significant (see, for instance, Column (5)).
Such figures maintain a similar magnitude and significance when we restrict the data
to monocentric cities and add an additional control for city size. This confirms our
theoretical assumption that residents are more sensitive to park distance than park size. It
even suggests the stronger argument that residents are not sensitive to park size (i.e., ϕ =
0). Only park access seems to matter to them.
This analysis raises the question of whether residents are more sensitive to the clos-
est park rather than to the set of parks surrounding their dwellings. To explore this
issue, we re-run the regressions with the distance to the closest park dmin and the CES
distances dψ with ψ ∈ {2, 5, 10, 15, 20, 50}. Intuitively, for small ψ, the CES distance dψ
averages the distances to many far surrounding parks. For higher ψ, it places more
weight on closer parks and therefore averages the distances to a subset of closer parks.
For ψ = ∞, it returns the distance to the closest park. If the effect of distance becomes
more significant when ψ increases, this means that individuals are more concerned with
the parks closer to their residences. The results are displayed in Table 4.8. First, note
that the effect on income and distance to the city centre are unaffected by the measure
of distance to parks. The effect of city income w and distance to the CBD r remain sig-
nificant with different measures of distance dmin and dψ . The coefficient on park size
θ̂ 3 remains non-significant and has a smaller magnitude than the coefficient on park
distance θ̂ 4 . Finally, the significance level of θ̂ 4 increases as ψ increases. Furthermore,
the R2 value is slightly higher for higher ψ. Such results suggest that households care more
about the nearest park to their homes than the distance to other nearby parks.
Finally, the regression model (4.13) might be subject to endogeneity issues, as there
could exist confounding omitted variables that affect both the dependent variable (pop-
ulation density) and the independent variables (park size and distance to closest park).
Therefore, Table 4.7 may present biased estimates of the coefficients θ̂ 3 and θ̂ 4 . To check
for this omitted variable issue, we first extract the residuals of the density regression
model (4.13) and then run the park size and distance-to-closest-park regression models
(4.11) and (4.12) augmented with a linear term including the above residuals and hav-
ing coefficients η3 and µ3 . In the case of an omitted variable, the estimated coefficients
η 3 and µ
b b3 should be significantly different from zero (see Appendix C for further de-
tails). We also apply this procedure to the regression on the number of parks to ensure
that the regression model (4.13) is also robust to the inclusion of this variable (see the
estimated coefficient b ρ4 ).
The estimation results are reported in Table 4.9. The controls are the same as in
previous tables. We report results for the samples with all cities and only monocentric
cities. It can be observed that the estimated coefficients b η3, µ
b3 and b
ρ4 are no longer
significant once we introduce the city controls for geography and annulus controls
(proximity of agriculture and forest land). In the presence of those controls, there is
no suspicion of omitted variable problems. Moreover, the magnitude of those coeffi-
4.3. EMPIRICAL ANALYSIS 113

Table 4.8: Population Density and Urban Green Stratification with Closest Distance and
CES Distance (OLS)

Dependent variable: Ln Population Density

(1) (2) (3) (4) (5) (6) (7)
Ln GDP per capita 0.278∗∗∗ 0.268∗∗∗ 0.302∗∗∗ 0.294∗∗∗ 0.290∗∗∗ 0.287∗∗∗ 0.282∗∗∗
(0.104) (0.103) (0.106) (0.105) (0.105) (0.104) (0.104)
Dist. to CBD (km) −0.054∗∗∗ −0.059∗∗∗ −0.054∗∗∗ −0.054∗∗∗ −0.054∗∗∗ −0.054∗∗∗ −0.054∗∗∗
(0.007) (0.008) (0.008) (0.007) (0.007) (0.007) (0.007)
Ln UGS Radius k −0.029 −0.056∗∗ −0.039 −0.034 −0.032 −0.031 −0.030
(0.026) (0.027) (0.026) (0.026) (0.026) (0.026) (0.026)
Ln Closest Dist. dmin −0.136∗∗∗
(0.024)
Ln CES Dist. dψ
ψ=2 0.040
(0.038)
ψ=5 −0.078∗∗∗
(0.027)
ψ = 10 −0.109∗∗∗
(0.025)
ψ = 15 −0.118∗∗∗
(0.025)
ψ = 20 −0.123∗∗∗
(0.025)
ψ = 50 −0.131∗∗∗
(0.024)
City Samples Full Full Full Full Full Full Full
UGS Samples Full Full Full Full Full Full Full
City Geo. Controls Yes Yes Yes Yes Yes Yes Yes
Annuli controls Yes Yes Yes Yes Yes Yes Yes
FE Controls NUTS2 NUTS2 NUTS2 NUTS2 NUTS2 NUTS2 NUTS2
Observations 16,782 16,782 16,782 16,782 16,782 16,782 16,782
Adjusted R2 0.787 0.783 0.784 0.785 0.786 0.786 0.786
Note: Here, we use the GDP per capita taken from Regional Economic Accounts from Eurostat at NUTS3
level with adjustment to purchasing power standard (PPS) as the proxy for city income level, and it is mea-
sured in e1,000 per inhabitant. The distance to CBD is measured in kilometre. City geographical controls
includes the average Elevation, Latitude and Longitude. The average rain fall, and average temperature
in Jan 01 (winter time) and average temperature in July 01 (summer time) for period 1995 − 2010. City
boundary is chosen at 20% cut-off point of residential land use share (See details in Picard and Tran (2019)).
The Agriculture and Forest Land from the last two annuli control variables are measured within 100-meter
buffer around the residential area. Standard errors are clustered at city level and reported in parenthesis.
The significance levels are denoted by ∗ p<0.1, ∗∗ p<0.05, ∗∗∗ p<0.01.

Table 4.9: Check for Endogeneity: estimation of ρ4 , η3 , and µ3

(1) (2) (3) (4) (5) (6) (7)

ρ̂4 0.162∗∗∗ 0.259∗∗∗ 0.095 0.048 0.078 0.075 −0.003
(0.055) (0.057) (0.059) (0.052) (0.052) (0.061) (0.057)
η̂ 3 −0.000 −0.000 −0.000 −0.000 −0.000 −0.000 −0.000
(0.019) (0.019) (0.018) (0.019) (0.022) (0.023) (0.023)
µ̂3 0.000 0.000 0.000 0.000 0.000 0.000 0.000
(0.023) (0.025) (0.024) (0.023) (0.025) (0.030) (0.030)
City Samples Full Full Full Full Full Monocentric Monocentric
samples samples samples samples samples cities cities
UGS Samples Full Full Full Full Full Full Full
samples samples samples samples samples samples samples
City Geo. Controls No Yes Yes Yes Yes Yes Yes
Annuli controls No No Yes Yes Yes Yes Yes
FE Controls No No No Country NUTS2 NUTS2 NUTS2 ×
City Size
Note: Standard errors are clustered at city level and reported in parenthesis. The significance levels
are denoted by ∗ p<0.1, ∗∗ p<0.05, ∗∗∗ p<0.01.
114 CHAPTER 4. GEOGRAPHICAL STRATIFICATION OF GREEN URBAN AREAS

cients diminishes with the introduction of various fixed effects and the restriction of
the data to monocentric cities. This point strengthens our belief in the absence of omit-
ted variables.

4.4 Conclusion
Urban land is a scarce resource that is highly contested for different kinds of uses. In
particular, green urban spaces compete with residential properties to which they bring
positive amenities. Therefore, the choice of the location and size of green urban spaces
is a challenging task for city governments. This paper sheds light on this choice from a
theoretical and empirical perspective.
The paper shows that the optimal number of green urban spaces is not uniform
across city locations. The number of parks first increases at small distances from the
city centre and then decreases at large distances from the centre, forming a bell-shaped
pattern in large cities. Two channels affect the number of green urban areas: on the one
hand, at short distances from the city centre, the limited land availability constrains the
creation of green urban areas; on the other hand, in the city suburbs, the low population
density diminishes the social benefit and therefore the desirability of creating parks. By
contrast, the optimal park sizes and distances to other parks are shown to be monotone
increasing functions of the distance to the city centre. The main contribution of this pa-
per is to empirically confirm those relationships with the high-resolution geographical
database of GMES Urban Atlas. From this dataset, we find that the number of green ur-
ban areas is bell-shaped or concave in city size. Our empirical analysis also reveals that
a 1-kilometre increase in distance to the city centre is associated with increases of 2.0%
and 2.7% in average park size and average distance to the closest park, respectively.
While the empirical analysis fits well the properties of the distribution of green ur-
ban spaces with respect to their distance to the city centre, it presents a weaker match
with respect to the productivity of cities. In theory, more productive cities are expected
to include more numerous parks with smaller sizes and longer access distances. The
empirical analysis finds support for the effect of higher productivity cities on the num-
ber of parks but no support for its effect on park sizes or the distance between parks.
In addition, population density is theoretically expected to rise with the presence of
larger and closer parks because residents are expected to prefer larger and closer parks.
The data suggest a strong effect of park access but no effect of park size on population
density. This suggests that residents are much more sensitive to the proximity than the
size of green urban areas.
APPENDIX 115

Appendix A: Cobb-Douglas utility

In this appendix we prove (4.3) and (4.4). We recall that U (z, s, a) = aα z1− β s β , a(k, d) =
a0 k ϕ d−δ and C (k ) = c0 kγ .

Under this setting we have ak = k a and V̂a (r, a, u) = αβ 1a V̂ (r, a, u) while V̂ (r, a(k, d), u) =
ϕ

−δ α α
V̂ r, βk ϕ d−δ , u = V̂ r, βk ϕ (rφ)−δ , u = (φ) β (k ϕ r −δ ) β V̂ (r, a0 , u) where φ is the an-

gle of the polar coordinate of the individual at (r, φ). So, we also compute
Z π/n   Z π/n
−δ αβ
V̂ [r, a (k, rφ) , u] rdφ = V̂ r, a0 k ϕ r −δ , u r (φ) dφ
k/r k/r
1−δ αβ
− (k/r )1−δ β
α
ϕ −δ
α (π/n)
= r (k r ) V̂ (r, a0 , u)
β
1 − δ αβ

Z π/n Z π/n
α ϕ
V̂a [r, a (k, rφ) , u] ak [k, rφ] rdφ = V̂ [r, a (k, rφ) , u] rdφ
k/r k/r β k
1− δ α 1−δ αβ
α ϕr ϕ −δ αβ (π/n) β − (k/r )
= (k r ) V̂ (r, a0 , u)
β k 1 − δ αβ

First order conditions

We can check conditions (4.1) and (4.2). We successively get

1− δ α 1− δ α
α ϕr ϕ −δ αβ (π/n) β − (k/r ) β
Sk =0 ⇐⇒ (k r ) V̂ (r, a0 , u) − V̂ [r, a (k, k) , u] − c0 γkγ−1 = 0
β k 1 − δ αβ
1− δ α 1− δ α   α
α (π/n) β − (k/r ) β k δ β −1 ϕ−δ αβ α
⇐⇒ ϕ (k ) V̂ (r, a0 , u) − (k ϕ−δ ) β V̂ (r, a0 , u) − c0 γkγ−1 = 0
β 1 − δβ α
r
"
1− δ β
α #
α πr −1 c0 γ
− 1 k( ϕ−δ) β −(γ−1) =
α
nk
⇐⇒ ϕ
β 1 − δβ α
V̂ (r, a0 , u)
 πr 1−δ αβ 1 − δβ α 
c0 γ 1

⇐⇒ = +1 +1
V̂ (r, a0 , u) k( ϕ−δ) β −(γ−1)
α
nk ϕ αβ
 πr 1−δ αβ 1 − δ αβ 1 − δ αβ c0 γ 1
⇐⇒ = 1+ + − β −( γ −1)
α
nk α
ϕβ α
ϕ β V̂ (r, a0 , u) k ( ϕ δ )
 πr 1−δ αβ 
1 β
 
1 β

c0 γ 1
⇐⇒ = 1+ −δ + −δ − β −( γ −1)
α
nk ϕ α ϕ α V̂ (r, a0 , u) k ( ϕ δ )

and
116 APPENDIX

π/n(r ) Z π/n(r)
Z 
Sn = 0 ⇐⇒ 2 V̂ [r, a (k (r ), rφ) , u] rdφ − R A rdφ − C [k (r )]
k (r )/r 0
2πr 
− V̂ [r, a (k (r ), rπ/n(r )) , u] − R A = 0
n (r )
1−δ αβ 1−δ αβ
( )
α (π/n) − ( k/r ) πr
⇐⇒ 2 r (k ϕ r −δ ) V̂ (r, a0 , u)
β − RA − c0 k γ −
1 − δ αβ n
2πr
   −δ  
ϕ πr
V̂ r, a0 k , u − RA = 0
n (r ) n
1− δ α 1− δ α  −δ 
(π/n) β − (k/r ) β

ϕ −δ β
α πr ϕ πr
⇐⇒ r (k r ) V̂ (r, a0 , u) − c 0 k γ
− V̂ r, a 0 k ,u = 0
1 − δ αβ n (r ) n
 δ αβ 1− δ α 1− δ α
k ( ϕ−δ) αβ (π/n) β − (k/r ) β
⇐⇒ r k V̂ (r, a0 , u) − c0 k γ
r 1 − δ αβ
  α
πr nk δ β ( ϕ−δ) αβ
− k V̂ (r, a0 , u) = 0
n πr
1− δ α 1− δ α
"   α #
k δ β −1 (π/n) β − (k/r ) β  πr 1−δ αβ
( ϕ−δ) αβ
⇐⇒ k − k k V̂ (r, a0 , u) − c0 kγ = 0
r 1 − δβ α
nk

1−δ αβ  
δ αβ πrnk − 1 ( ϕ−δ) α
⇐⇒ k  k β V̂ (r, a , u ) − c k γ −1  = 0
0 0
1 − δ αβ

1−δ αβ 
δ αβ πrnk − 1
 k( ϕ−δ) β V̂ (r, a0 , u) − c0 kγ−1 = 0 if k > 0
α
⇐⇒ 
1 − δβ α


1−δ αβ 
δ αβ πrnk − 1 c0
 k( ϕ−δ) β −(γ−1) =
α
⇐⇒  if k > 0
1 − δβ α
V̂ (r, a0 , u)
α  πr 1−δ αβ
 
α c0 1
⇐⇒ δ = 1+ 1−δ −
if k > 0
β −( γ −1)
α
β nk β V̂ (r, a0 , u) k ( ϕ δ )
 
 πr 1−δ αβ 1 1 − δ α
β c0 1
⇐⇒ = α + −
if k > 0
β −( γ −1)
α
nk δβ δβα
V̂ (r, a0 , u) k ( ϕ δ )
 πr 1−δ αβ β

β

c0 1
⇐⇒ = + −1 −
if k > 0
β −( γ −1)
α
nk δα δα V̂ (r, a0 , u) k ( ϕ δ )

and for k > 0. The last equality gives an inverse relationship between n and k.

Suppose for moment park amenities are proportional to park sizes and distances
and maintenance cost is linear (ϕ = δ = γ = 1) . Then, the two conditions become
co-linear and give

   − β−β α
nk β β c0
= + −1
πr α α V̂ (r, a0 , u)
APPENDIX 117

which is smaller than 1 as β > α. It’s true as

 
β β c0
β > α =⇒ + −1 >1
α α V̂ (r, a0 , u)
β
β > α =⇒ − <0
β−α
   − β−β α
β β c0
Then, + −1 <1
α α V̂ (r, a0 , u)

Now, suppose that δ > ϕ and γ < 1, plugging condition (1) into condition (2), we
have

     
1 β 1 β c0 γ 1 β β c0 1
1+ −δ + −δ = + − 1
V̂ (r, a0 , u) k( ϕ−δ) β −(γ−1) V̂ (r, a0 , u) k( ϕ−δ) β −(γ−1)
α α
ϕ α ϕ α δα δα
        
1 β β β β γ c0 1
⇐⇒ 1+ −δ − − −1 − −δ −
=0
β −( γ −1)
α
ϕ α δα δα α ϕ V̂ (r, a0 , u) k ( ϕ δ )
        
1 β β β β γ c0 1
⇐⇒ 1+ −δ − = −1 − −δ − β −( γ −1)
α
ϕ α δα δα α ϕ V̂ (r, a0 , u) k ( ϕ δ )
   
1 + ϕ1 α − δ − δα
β β
c0 1
⇐⇒ =   
V̂ (r, a0 , u) k( ϕ−δ) β −(γ−1)
α
β β γ
δα − 1 − α − δ ϕ
   
β β γ
( ϕ−δ) αβ −(γ−1) δα − 1 − α − δ ϕ c0
⇐⇒ k =   
1 + ϕ1 α − δ − δα V̂ (r, a0 , u)
β β

  
β γδ
− 1 1 − c0
( ϕ−γ)+(1−δ) αβ δα ϕ
⇐⇒ k =   
δα
β
−1 δ
ϕ− 1 V̂ (r, a0 , u)
  1
γδ (1−γ)+( ϕ−δ) α
1− ϕ c0 β 
α

⇐⇒ k =   if 1−δ 6= 0
δ
ϕ − 1 V̂ (r, a0 , u) β

and

 πr 1−δ αβ β

β

c0 1
= + −1 − β −( γ −1)
α
nk δα δα V̂ (r, a0 , u) k ( ϕ δ )

 πr 1−δ αβ   δ −1
β β ϕ
⇐⇒ = + −1 1
nk δα δα ϕ − γδ
!− 1−1δ α
ϕ −1
  δ
πr β β β
⇐⇒ n = + −1 1
k δα δα ϕ − γδ
 
Therefore, when 1 − δ αβ 6= 0 we have the interior solution for n and k as
118 APPENDIX

! 1
(1−γ)+(1−δ) α
∗ ζc0 β
k =
Vb (r, a0 , u)
κπr
n∗ = ∗
k
where
γδ
1− ϕ
ζ≡ δ
ϕ −1
   − 1 α
β β 1 1− δ
β
κ≡ + −1
δα δα ζ
 
When 1 − δ αβ = 0 we can rewrite the condition (1) and (2) as follows:

 
π k β β c0 γ 1
Sk = 0 ⇐⇒ ln − ln = +
α V̂ (r, a0 , u) k ϕ β −γ
α
n r α
 
π k c0 1
Sn = 0 ⇐⇒ ln − ln = 1+
V̂ (r, a0 , u) k β −γ
α
n r ϕ

if k > 0. Replacing condition (1) into condition (2), we have the solution for n and k

  ϕ1
∗ ζc0 −γ
k = δ

V̂ (r, a0 , u)
1 πr
n∗ =
1 k∗
 
exp 1 + ζ

Interior conditions

In order to have interior solutions, the two follow conditions need to satisfy

γδ
1− ϕ
(i) ζ ≡ >0
−1 δ
ϕ
 
β β 1
(ii) + −1 >0
δα δα ζ
APPENDIX 119

γδ
1− ϕ ϕ
Condition (i) ⇐⇒ > 0 ⇐⇒ γ < as δ > ϕ
−1δ
ϕ
δ
 
β β 1
Condition (ii) ⇐⇒ + −1 >0
δα δα ζ
  1 − γδ
β β ϕ
⇐⇒ + −1 δ >0
ϕ −1
δα δα
    
β γδ β δ γδ
⇐⇒ 1− + −1 − 1 > 0 as 1 − >0
δα ϕ δα ϕ ϕ
   
β β β δ β
⇐⇒ −γ + − − +1 > 0
δα ϕα ϕα ϕ δα
β δ β
⇐⇒ − +1 > γ
ϕα ϕ ϕα
δα ϕα
⇐⇒ 1 − + >γ
β β
α
⇐⇒ 1 + ( ϕ − δ) > γ
β
ϕ
It can be combined as γ < min{ δ ; 1+ ( ϕ − δ) αβ }.
Now we need to prove that under interior condition above, the share of urban green
in each annulus is smaller than one. More specifically, we need to prove that

 − 1 α
n∗ k∗
 
β β 1 1− δ
β
=κ≡ + −1 <1
πr δα δα ζ
We consider two case:  
Case 1: δ < α which is automatically satisfy condition (ii) as δα − 1 1ζ > 0
β β

We have
β α 1
δ< =⇒ 1 − δ > 0 =⇒ − <0
α β 1 − δ αβ
 
β β β 1
δ < =⇒ > 1 and −1 > 0 as ζ > 0
α δα δα ζ
 
β β 1 1
=⇒ + −1 > 1 and − <0
δα δα ζ 1 − δ αβ
   − 1 α
β β 1 1− δ
β
=⇒ κ ≡ + −1 <1
δα δα ζ
β
Case 2: δ > α
120 APPENDIX

We have
β α 1
=⇒ 1 − δ < 0 =⇒ −
δ> >0
α β 1 − δ αβ
 
β β β 1
δ > =⇒ < 1 and −1 < 0 as ζ > 0
α δα δα ζ
 
β β 1
=⇒ + −1 <1
δα δα ζ
 
β β 1
Condition (ii) =⇒ + −1 >0
δα δα ζ
 
β β 1 1
=⇒ 0 < + −1 < 1 and − >0
δα δα ζ 1 − δ αβ
   − 1 α
β β 1 1− δ
β
=⇒ κ ≡ + −1 <1
δα δα ζ
β
Case 3: δ = α we are back to the special case where:

  ϕ1
∗ ζc0 −γ
k = δ

V̂ (r, a0 , u)
1 πr
n∗ =  
exp 1 + 1 k
ζ

nk 1
then the share of green in annulus at distance r is πr =
exp(1+ 1ζ )
< 1.

Comparative statics

Under interior condition (i) and (ii), we have

  1
ζc0 (1−γ)+( ϕ−δ) α
∗ β
k =
V̂ (r, a0 , u)
κπr
n∗ = ∗
k

We have
1
−1
−ζcV̂r
 
1 ζc0 (1−γ)+( ϕ−δ) α
k∗r
β
=
2
(1 − γ) + ( ϕ − δ) αβ V̂ (r, a0 , u) V̂ (r, a0 , u)
1 −k∗ V̂r
= >0
(1 − γ) + ( ϕ − δ) β V̂ (r, a0 , u)
α

α
As (1 − γ) + ( ϕ − δ) > 0 and V̂r < 0
β
APPENDIX 121

Note that:
α α
(1 − γ ) + ( ϕ − δ ) > 0 ⇐⇒ (1 − γ) > −( ϕ − δ)
β β
α
⇐⇒ γ < 1 + ( ϕ − δ)
β
α α
⇐⇒ γ < 1 + ϕ − δ
β β

which is always true under Interior Condition (ii)

122 APPENDIX

Appendix B: Quadratic regression robustness check

In this section, we use quadratic form of distance to city centre to test for the hypothesis
that the number of urban green spaces is first increasing with small value of distance
and then decreases as one moves away from the center. Specifically, we test the follow-
ing reduced form regression

nijc = α1 distijc + α2 dist2ijc + ∑ Xijc + ∑ X jc + ∑ Ic + eijc

(1) (2)

ijc jc c

(1) (2)
where (i, j, c) is an observation of annuli at annulus i of city j of region c; Xijc and X jc
are control variables for annuli and city observable characteristics; Ic is region dummy
variable controlled for unobservable characteristics like environmental policies or local
regulation. Theoretical model predicts the number of parks first increases and then
decreases as one move farther from city centre; hence, the sign of α1 is expected to be
positive while α2 is expected to be negative.
The regression results are reported in Table B1. The standard errors are clustered at
city level, which controls for the possible correlation between observations within one
urban area. The regression shows a negative and significant coefficient for the square
of the distance to CBD, suggesting the non-monotonic relationship between numbers
of urban green spaces and the distance to CBD. It also shows a positive and significant
coefficient for αˆ1 . It means that the number of parks first increases and then decreases
with distance to city centers. Column (1) indicates the results without controlling for
city geographic and observable characteristics and without city fixed effect. As it can
be seen, the magnitude increases when we introduce more controls, but the sign ad
significant level are not affected. Column (3) to (10) further introduce the control for
other type of open spaces like Agriculture and Forest Land available within 100-meter
buffer from residential areas in each annuli. The magnitude for αˆ1 and αˆ2 increase
nearly two folds from −0.083 to −0.153 for the square of the distance to city center α̂2 ,
and from 2.8 to 4.5 for distance to CBD α̂1 with same level of significance. The higher
presence of agriculture and forest land associate with lower number of urban green
spaces. Specifically, a 1% increase in agriculture land associates with a reduction of 60-
74 parks, and the magnitude for forest land is similar at 39-104 parks reduction for 1%
increase in forest land within 100-meter buffer from residential area. Column (8) to (10)
further exclude big parks at 99th percentile threshold, and the sign and magnitude are
still similar in those cases.
 Table B1: Number of Urban Green Spaces (Quadratic OLS Regression)
APPENDIX

Dependent variable: Ln Number of Urban Green Spaces

(1) (2) (3) (4) (5) (6) (7) (8) (9) (10)
Dist. to CBD (km) 3.190∗∗∗ 2.856∗∗∗ 4.551∗∗∗ 4.753∗∗∗ 4.800∗∗∗ 5.138∗∗∗ 4.581∗∗∗ 4.701∗∗∗ 5.043∗∗∗ 4.491∗∗∗
(0.437) (0.400) (0.509) (0.470) (0.412) (0.434) (0.471) (0.411) (0.431) (0.467)
Dist. to CBD square −0.098∗∗∗ −0.083∗∗ −0.152∗∗∗ −0.154∗∗∗ −0.188∗∗∗ −0.216∗∗∗ −0.210∗∗∗ −0.183∗∗∗ −0.212∗∗∗ −0.206∗∗∗
(0.036) (0.034) (0.039) (0.035) (0.024) (0.030) (0.030) (0.024) (0.030) (0.029)
GDP per capita 0.212∗∗∗ 0.112∗∗∗ 0.124∗∗∗ −0.071 0.095 −0.101 −0.060 0.096 −0.101
(0.039) (0.040) (0.041) (0.053) (0.087) (0.108) (0.051) (0.086) (0.106)
Annuli controls
Agri. Land (%) −74.079∗∗∗ −74.436∗∗∗ −75.091∗∗∗ −71.006∗∗∗ −60.495∗∗∗ −73.499∗∗∗ −69.756∗∗∗ −59.356∗∗∗
(9.486) (9.528) (7.766) (8.035) (7.541) (7.659) (7.924) (7.431)
Forest Land (%) −104.275∗∗∗ −95.416∗∗∗ −61.227∗∗∗ −71.406∗∗∗ −39.950∗∗∗ −60.126∗∗∗ −70.414∗∗∗ −39.194∗∗∗
(15.831) (14.642) (16.677) (17.293) (12.261) (16.583) (17.202) (12.257)
City Samples Full Full Full Full Full Monocentric Monocentric Full Monocentric Monocentric
samples samples samples samples samples cities cities samples cities cities
UGS Samples Full Full Full Full Full Full Full Excld. UGS Excld. UGS Excld. UGS
samples samples samples samples samples samples samples abv. 99th per. abv. 99th per. abv. 99th per.
City Geo. Controls No Yes Yes Yes Yes Yes Yes Yes Yes Yes
FE Controls No No No Country NUTS2 NUTS2 NUTS2 × NUTS2 NUTS2 NUTS2 ×
City Size City Size
Observations 16,782 16,782 16,782 16,782 16,782 14,368 14,368 16,780 14,366 14,366
Adjusted R2 0.239 0.325 0.421 0.486 0.671 0.615 0.656 0.665 0.611 0.653
Note: Here, we use the GDP per capita taken from Regional Economic Accounts from Eurostat at NUTS3 level with adjustment to purchasing power
standard (PPS) as the proxy for city income level, and it is measured in e1,000 per inhabitant. The distance to CBD is measured in kilometre. City
geographical controls includes the average Elevation, Latitude, Longitude and the average and square of the average of the rain fall, the temperature
in summer and winter time. The average rain fall, and average temperature in Jan 01 (winter time) and average temperature in July 01 (summer time)
are measured for the period 1995 − 2010. City boundary is chosen at 20% cut-off point of residential land use share (See details in Picard and Tran
(2019)). The thresholds for urban green size at 99th percentile is 54.82 hectares. Columns (6), (7), (9) and (10) isolate monocentric cities by excluding
the polycentric cities reported by OECD. Columns (1) to (7) include all urban green spaces within urban boundary, while Columns (8) to (10) exclude
all big parks above 99th percentile. The Agriculture and Forest Land from the last two annuli control variables are measured within 100-meter buffer
around the residential area. Standard errors are clustered at city level and reported in parenthesis. The significance levels are denoted by ∗ p<0.1,
∗∗ p<0.05, ∗∗∗ p<0.01.
123
124 APPENDIX

Appendix C: Robustness Check - Test for endogeneity

Note that in the theoretical model, the urban planner makes decisions in two steps.
First, given the configuration of urban green spaces, she assigns the consumption of
composite goods, housing slot sizes, and at the same time, the population density. Sec-
ond, when we know the change in population density and housing slot size at each
configuration of urban green, she decides the optimal locations. We assume there is
a heterogeneous variable that we could not observe from the data, but it affects the
household decision. Specifically, we assume

U (z, s, a) = ξaα z1− β s β

where ξ follow some distribution with mean one. Using the same procedure as in the
Cobb-Douglas Utility section, we have
1
1 ξβ
λ= = 1− β
s − − 1− β − α 1
(1 − β ) β y β a β u β
Here, we observe and infer λ, w, k, and d from our database, and from each observation
of λ, we can infer ξ as the residuals (which we affects the population density but are
not explained by urban green space systems and other observable controls). Basically,
we have the log-linear equation and the reduced-form equation as follows:

1−β 1−β
 
1− β α
α α 1 1
ln λ = ln (1 − β) β β
a0 + ln w − tr + ln k − δ ln d − ln u + ln ξ
β β β β β β

ln λr f c = θ0 + θ1 ln w f c + θ2 distr f c + θ3 ln kr f c + θ4 ln dr f c + ∑ Xr f c + ∑ X f c + ∑ I f c + er f c
(1) (2)

rfc fc fc

Then, we have

1
êr f c ≡ − ln ξ r f c =⇒ ξ̂ r f c = e− βêr f c
β
If the distribution of error terms approximate the normal distribution with mean zero,
then we also have the distribution of ξ with a mean close to 1.
Now, we also have

  1
ζc0 (1−γ)+(1−δ) α
β
k= ∗
V (r, a0 , u, ξ )

   
 
1 ζc0  − 1 ln w + 1 tr − ln u− β a
1 α
1
ln k = ln  1− β 0
β
− ln ξ 
(1 − γ ) + (1 − δ ) β
α
β (1 − β ) β β β β
APPENDIX 125

Using the value of ξ̂ r f c above, we can add them into the regression for the size of ur-
ban green space and the distance dmin and dφ . By that way, we can control for other
types of controls that drive the population density but not from the green systems and
controlled variables. As the decision of urban planner in our model channels through
the land value and population density, if we can examine here, how other amenities or
location characteristics affect the choice of sizes and locations of urban green that are
not purely geographical (like distance to CBD or elevation).
Similarly, with the total number of parks
1
V ∗ (r, a0 , u, ξ )
 
(1−γ)+(1−δ) α
∗ β
n = κπr
ζc0

  ϑ 
1− β
β (1 − β ) β
 α

ϑ − β1 β ϑ
ln n = ln κπ  + ln r + ln w − + ln u a + ln ξ
   ϑtr ϑ 0
ζc0 β β

Hence, we run the following regressions

ln nijc = ρ0 + ρ1 ln w jc + ρ2 ln distijc + ρ3 distijc + ρ4 êijc + ∑ Xijc + ∑ X jc + ∑ Ic + ε ijc

(1) (2)

ijc jc c

ln k ijc = η0 + η1 ln w jc + η2 distijc + η3 êijc + ∑ Xijc + ∑ X jc + ∑ Ic + ε ijc

(1) (2)

ijc jc c

ln dijc = µ0 + µ1 ln w jc + µ2 distijc + µ3 êijc + ∑ +∑ + ∑ Ic + ε ijc

(1) (2)
Xijc X jc
ijc jc c

to estimate the effect of the unobserved factor in household decisions which affect both
the population density and the choices for urban green spaces numbers, sizes and dis-
tances. If the estimated value of ρ̂4 , η̂ 3 , and µ̂3 is not significant and close to zero, we can
exclude a systematic issues in omitted unobserved variables that affect both population
density and the location and size of urban green space systems at the same time.
The estimation results are indicated in Table 4.9 in the main text. We use both full
samples of all annulus in all cities as well as annulus from only monocentric cities with
different level of controls and fixed effects. Specifically, in Column (1) and (2), we use
no control and use only control for city geographical variables, and we can see the es-
timated value for ρ̂4 is positive and significant suggesting that the annuli controls for
agriculture and forest presence and the regional fixed effects are important factors in
deciding the total number of parks in each annulus across locations and affect the pop-
ulation density as well. Once, we control for annuli factors like agriculture and forest
shares, all variable ρ̂4 , η̂ 3 , and µ̂3 are close to zero in magnitude and are not signifi-
cant. The estimation remains the same with sub-samples of annulus within monocen-
tric cities only.
Chapter 5

Urban Floods, Amenity Land and

City Structure

”Nature no longer exists apart from humanity. Henceforth, the world we will inhabit is
the one we have made.” - Jedediah Purdy (2015), After Nature: A Politics for the
Anthropocene

5.1 Introduction
Cities or urban areas are the places that host half of the world population; in addition,
this proportion will soon be two-thirds in the next thirty years (UN DESA, 2014). In
some countries, urban areas are also estimated to generate approximately 70 to 80%
of the gross domestic product (GDP) (Weiss, 2001, Dobbs et al., 2011). Meanwhile,
the total build-up areas account for only 1-2% of the total land on earth. Such a high
level of concentration of population and properties in urban areas is essential for in-
novation and growth. However, it also exposes cities to spatial vulnerabilities, which
could make them susceptible to the higher risk of natural hazards than rural areas. If
the urbanization process, together with its downside vulnerabilities, is unchangeable
and inevitable, the next question should be how we manage to make the urban system
work. The question regarding how to manage urban areas has been one of the most
critical issues on the UN post-2015 development agenda (UN DESA, 2014).
From our point of view, flood hazard is one of the most recurring and costliest natu-
ral disasters. It is estimated that from 1985-2014, floods have caused a death toll of half
a million people and approximately 500 billion in US dollars lost in property world-
wide (Dartmouth Flood Observatory, 2014). In Europe from 1985 to 2008, there were
285 flood events in total, which on average lasted 8,5 days, killed approximately eight
people, displaced approximately 11,000 people and damaged approximately 1 billion
worth in US dollars per a flood event as indicated in Table 5.1. Despite this fact, flood-
prone areas are more likely to be overpopulated due to the cost of constructing defen-

127
128 CHAPTER 5. URBAN FLOOD, AMENITY LAND AND CITY STRUCTURE

Table 5.1: General flood statistics in EU27 for 1985-2008

obs. mean sd median min max

Duration (days) 285 8.51 11.61 4.00 1.00 102.00
No. of people death (hab.) 285 8.05 15.67 2.00 0.00 148.00
No. of people displaced (thousand hab.) 185 11.09 35.38 1.00 0.00 294.10
Damages (millions USD) 100 1039.27 4223.08 32.79 0.00 36466.01
Affected Area (thousand km2 ) 285 49.65 110.11 19.59 0.01 1164.00
Note: The data are taken from Dartmouth Flood Observatory (DFO) flood map for the period
1985-2008, and only the flood disasters happened within the administrative boundaries of EU27
are taken into account. The data for displaced population and damages for some flood events
were not available.

sive infrastructure and because the costs of reconstruction are partially borne by the
residents in safer places or nonresidents (Kydland and Prescott,1977; Turner, 2012). If
this mechanism is precisely the case, it will create a considerable misallocation of the
resources in the economy. A land-intensive amenity such as green urban areas typically
involves a high opportunity cost as it requires the land in areas where it is most valu-
able. Therefore, a lower land price at a central area due to higher flood risk results in
lower cost for building any land-intensive green area. This situation, in turn, attracts
more residents and induces a higher level of residents susceptible to flood risks. The
same mechanism also applies to the building of flood defense systems. Hence, an un-
derstanding of such trade-offs when planning for the provision level of public amenity
and flood defense infrastructures is essential for urban planners.
Another important question is also why we should care about flood risk in urban
areas and whether it it better to relocate the whole city to a safe area. In the Euro-
pean context, most of its old and most populous cities such as London, Paris, Dublin,
Brussels or Berlin are close or next to networks of rivers, as they provided easy access
to water and defense in the past. Even though the waterways have become increas-
ingly less important in contemporary society, the substantial cost of moving and the
long duration of housing infrastructure make the relocation of a whole city rigid and
hard. Many studies also report that households positively value the proximity to wa-
terways such as rivers or canals and the coasts (Orford, 2002; Luttik, 2000; Cho et al.,
2006). Therefore, it is vital to incorporate the flood risk into urban land use planning,
especially in cities that contain any areas susceptible to this type of risk.

In this paper, we focus chiefly on the impact of flood hazards on optimal urban
amenity provision, flood defense systems and city structures under both first-best and
second-best scenarios. Toward this aim, we first create a model of a two-dimensional
city with an exogenous level of flood hazards. Households can decide on how much
consumption of composite goods, the sizes of their housing floor service, and where to
locate within an urban area. Under the first-best scenario, a city planner decides the
residential area and amenity land, as well as the density of households at each location,
5.1. INTRODUCTION 129

and how much to invest in the flood defense systems. Hence, in this case, the urban
planner internalizes all cost borne by the flood hazard. Under this scenario, we find a
nonmonotonic relationship between the level of flood risks and the optimal provision
of public flood defenses and amenity land. In the second-best case, the city government
collects taxes to cover the expenses of flood defense systems, as well as the provision
of amenity land. It faces budget balance constraints. We find that a locational tax is
efficient while an income tax system would result in over-concentration of population
and housing construction in a flood-prone area.
In the second part of the paper, we examine the current state of land uses in more
than 200 of the most populated European cities that are close to at least a river network
and susceptible to some flood hazards. Those cities are a subset of the city samples
of the Urban Atlas in Picard and Tran (2019).1 We combine that information with the
European 500-year and 100-year flood hazard maps provided by the European Envi-
ronment Agency (EEA) and Eurostat population grid database. Both of these datasets
cover the entirety of Europe. We first find that the population density and the fraction
of dense residential construction are more likely to establish in a safer area than in a
hazardous area, which confirms the prediction from our theoretical model. Second, the
provision of amenity lands is nonmonotonic with the level of flood risks and depends
on the relative location within the urban area. Specifically, the fraction of amenity land
is more likely to be built in risky and dense area than in a risky but sparse area near the
urban fringe.
The contribution of this paper to the literature is twofold. First, on the theoretical
side, it contributes to the understanding of the optimal allocation of amenity and urban
land use planning under an adverse and exogenous flood risk. Frame (1998) is the first
paper that incorporates the flood risk and insurance into a residential development
model. He finds a negative impact of flood risk on residential development under a
competitive land market. Bakkensen and Barrage (2018) use a dynamic asset pricing
model with heterogeneous beliefs of households to explain the ambiguous effect of
flood risk on housing prices. However, neither the papers addresses the effects of flood
risk on the choices of amenity land and flood defense systems. Picard and Tran (2019)
are the closest to this paper; nevertheless, they do not incorporate the flood risk inside
their model. To the best of our knowledge, this is the first paper that considers the effect
of river flood risks on optimal amenity land provision and flood defense systems under
both first-best and second-best frameworks.
The second contribution is to empirically measure the effects of river flood risk on
city structures and amenity provision across cities in Europe. Most of the empirical pa-
1 GMES Urban Atlas uses Earth observation satellite images with 2.5-m spatial resolution multispectral
data, which is at a much higher resolution than that of the traditional CORINE land cover resolution.
According to the European Environmental Agency, the current resolution of the GMES Urban Atlas is 100
times higher than that of the CORINE land cover.
130 CHAPTER 5. URBAN FLOOD, AMENITY LAND AND CITY STRUCTURE

pers in the literature use hedonic pricing to estimate the effects of flood hazards. Many
have reported negative effects of floods following hurricanes and other catastrophic
events (Hallstrom and Smith, 2005; Atreya et al., 2013; Bin and Landry, 2013; Zhang,
2016) as well as increases in flood insurance take-up rates (Gallagher, 2014). However,
those effects are often short-lived and mostly vanish after five years. Kocornik-Mina
et al. (2018) use the night light data to investigate the short-term effect of significant
flooding events worldwide. They also find that the economic activities in urban ar-
eas were shortly reduced immediately after the flood but soon reverted to a reasonable
level. One exception is Ortega and Sulleyman (2018), which find a persistent decrease
in prices of flood-prone properties in New York City after Hurricane Sandy due to a
more extensive and more detailed dataset at a precise postal-code. In this paper, we
utilize a different dataset of geographical land covers to track different types of land
uses at precise locations within an urban area. We also additionally combine those data
with the Eurostat population grid and the detailed flood hazard maps with multiple
return periods in Europe. These data at a high spatial resolution help us to examine
the prediction from the theoretical model precisely. We confirm a persistent pattern of
lower population density and lower residential construction in flood-prone areas. We
also find a nonmonotone relationship between amenity and flood risk level.

The organization of this paper is as follows: Section 2 introduces the theoretical

model and discusses the effects of flood risk on the city structures, as well as the optimal
provision of flood prevention facilities and amenity areas. The second part of Section
2 provides the framework of second-best policies with two different types of taxes:
locational and income tax. Section 3 is devoted to our empirical approach and results.
We first provide evidence on the negative impact of flood risk on the population density
and the residential construction cover. We, then estimate the amenity land choices with
flood risk levels across locations within the city area. The last section concludes the
paper.

5.2 Theoretical model

5.2.1 Model settings

5.2.1.1 Households and flood risk

We construct an urban spatial model to incorporate flood risk, commuting, housing

cost and local public amenity provision, as well as flood defense allocation. To achieve
this, we first follow the framework of Fujita (1989) and add an exogenous flood risk
level inside the model together with flood prevention efforts. We consider a monocen-
tric city where all households commute to a central business district (CBD) to work.
We assume that households necessitate commuting cost t and receive an income of w.
5.2. THEORETICAL MODEL 131

They are homogeneous and have the same preference U (z, h, a) represented over the
nonhousing composite good z, housing floor service h and amenity level a. We assume
U is increasing in each of the components, indicating the desirability of the additional
consumption of housing and nonhousing goods, and a higher level of amenity such that
Uz , Uh , Ua > 0. All goods are assumed to be normal goods with decreasing marginal
utility level Uzz , Uhh , Uaa < 0. We also further assume global concavity for our utility
function U.
We introduce the flood risk in our model in the following way. Households that
inhabit a flood risk area face a financial loss once a flood event occurs. The severity
of the loss is uncertain ex ante for households and is measured by a random variable
x ∈ [0, 1]. Once a flood event impacts households, households realize a financial loss,
which increases with the severity level x. We denote l ( x ) as the financial loss function
per each square meter of housing floor for each household. We assume that the financial
loss is increasing with the severity level x, such that lx > 0. Households also decide to
invest in some flood prevention effort e, which combined with the public prevention
effort p can reduce the financial loss from a flood event. Denote m(e, p) : R2 → [0, 1] as
the reduced fraction of financial loss stemming from the prevention efforts. We further
assume an increasing and concave function of the prevention fraction with respect to e
and p as well as

lim m(e, p) = 0 and lim m(e, p) = 1. (5.1)

e→0;p→0 e→+∞;p→+∞

Specifically, given a level of flood protection e > 0 and p > 0, the realized financial
loss from a flood event is l ( x )(1 − m(e, p)) per square meter of housing floor service.
The cost of providing the flood prevention from households is C H (e) per square meter
of housing floors. C H (e) is assumed to be an increasing and convex function CeH >
H > 0. We can then formalize the budget constraint for households as w − t =
0; Cee
z + Rh + h l ( x )(1 − m(e, p)) + C H (e) . The price of numeraire composite good z is

normalized to one, and the cost of renting a housing floor unit is R.

Household location is indicated by a pair (r, φ) where r is the Euclidean distance
from the CBD, and φ is the radian distance from the East axis. The space of households’
locations is then given by the product of R+ × [0, 2π ], and is Lebesgue measurable. For
each location (r, φ), the cumulative distribution of flood severity is F ( x; r, φ). We also
further assume that the flood risk distribution has the property of first-order stochastic
dominance, meaning that for every pair of locations (r, φ) and (r 0 , φ0 ), we can always
compare the risk level between them. A location (r 0 , φ0 ) is said to be riskier than lo-
cation (r, φ) if and only if F ( x; r 0 , φ0 ) ≤ F ( x; r, φ), ∀ x ∈ [0, 1]. A location is safe when
F (0; r, φ) = 1.
132 CHAPTER 5. URBAN FLOOD, AMENITY LAND AND CITY STRUCTURE

5.2.1.2 Housing sector

Following Owens et al. (2017), we assume that the housing sector is developed by a
large number of small developers with the free entry condition. Denote q(r, φ) as the
total housing floor developed per unit of land at location (r, φ). Land developers have
a convex variable cost f (q) = κqα with κ > 0, α > 1, in addition to a land rent Ψ from
absentee landlords.
Residential developers then choose the quantity housing floor q to maximize the
total profit
max Π = Rq − κqα − Ψ. (5.2)
q

The free entry condition states that

Π = 0 ⇐⇒ Ψ = Rq − κqα (5.3)

5.2.2 First-best world

Consider a benevolent urban planner whose objective is to maximize social welfare at

the minimal cost. To avoid the unequal treatment of the equals as indicated in Mirrlees
(1972), we assume that each consumer reaches a level of utility u, and the objective of
the urban planner is to minimize the costs to attain that level of utility. There are two
types of land uses that we consider in this model: residential and amenity land. We
denote g(r, φ) as the fraction of land used for amenity at each location, and the rest
1 − g(r, φ) is for the housing sector. For simplicity, it is further assumed that linear
amenity service a(r, φ) = θg(r, φ) with θ > 0, and a linear maintenance cost C M (r, φ) =
cg(r, φ). Additionally, the urban planner incurs a cost CU ( p) per unit of land for a
public flood prevention effort p. We assume that CU ( p) is an increasing and convex
function of p such that CU U
p > 0 and C pp > 0.
Due to the uncertain severity of flood disaster events, the urban planner then chooses
z(r, φ), h(r, φ), n(r, φ), e(r, φ), q(r, φ), p(r, φ), and g(r, φ) to minimize the following ex-
pected costs:

Zb Z2πZ1   
EC = z + t + h l ( x )(1 − m(e, p)) + C H (e) n`dF ( x; φ, r )dφdr
0 0 0
Zb Z2πZ1  h i 
+ R A + cg + CU ( p) + κqα (1 − g) `dF ( x; φ, r )dφdr (5.4)
0 0 0

where n(r, φ), `(r, φ) are the number of inhabitants and available land at location (r, φ)
respectively, and b is the city boundary. The optimization is subject to the housing floor
constraint n(r, φ)s(r, φ) = (1 − g(r, φ))q(r, φ) ∀(r, φ) ∈ [0, b] × [0, 2π ], the population
5.2. THEORETICAL MODEL 133

constraint
Zb Z2π
n(r, φ)dφdr = N (5.5)
0 0
and free mobility condition
U (z, h, a) ≥ u (5.6)
Applying the housing floor constraint condition, and since wages w are exogenous
in the urban areas, we can transform the minimization of the expected social expense in
(5.4) to the maximization of expected social surplus ES = wN − EC. After substitution
of the population and land use constraints, this provides
Zb Z2πZ1 
w−t−z

H
ES = max − l ( x )(1 − m(e, p)) − C (e) q(1 − g)`dF ( x; φ, r )dφdr
z,h,n,e,p,g h
0 0 0
Zb Z2π  
− κqα + CU ( p) (1 − g) + R A + cg `dφdr (5.7)
0 0

The social planner then chooses a bundle (z, h, n, e, p, g) to maximize the social surplus
such that U (z, h, a) ≥ u.

5.2.2.1 Housing demand and housing provision

We first denote the average financial loss from a flood risk event at a location (r, φ) with
flood severity accumulative distribution F ( x; r, φ) as
Z1
l (r, φ) ≡ l ( x )dF ( x; φ, r ). (5.8)
0
Since the urban planner is risk-neutral, the uncertainty can be eliminated in (5.7).
The optimization problem of the urban planner can be solved in three stages. First,
the urban planner chooses the optimal consumption bundle (z, h) given by the point-
wise first-order condition of the residential housing floor value.
w−t−z
 
H
V (h, u, a) ≡ max − l (r, φ)(1 − m(e, p)) − C (e) s.t U (z, h, a) ≥ u (5.9)
s,z h
We denote the solution of U (z, h, a) = u as ẑ(h, u, a). As U (z, h, a) is an increasing
function of z, it is easy to see that ẑ(h, u, a) exists and is unique. Replacing it in equation
(5.9), we can now take the first-order condition with respect to s, and equalize to zero,
the optimal housing slot size needed to satisfy the following condition;
ẑ − hẑh = y (5.10)
where y = w − t(r ) is the disposable income after subtracting the commuting cost.
Given the concavity of the utility function yielding zhh > 0, the second-order conditions
are satisfied at any location2 , which implies a unique solution to the above maximiza-
2 We have the second order equal to −hzhh < 0.
134 CHAPTER 5. URBAN FLOOD, AMENITY LAND AND CITY STRUCTURE

tion problem. Denote the solution from equation (5.10) as h̃(y, u, a), and we obtain the
optimal composite good consumption as z̃(y, u, a) and the land value Ṽ (y, u, a, e, p, l )

w − t − z̃(y, u, a)
Ṽ (y, u, a, e, p, l ) = − l (1 − m(e, p)) − C H (e)
h̃(y, u, a)

where l is the average financial loss of flood risks at location (r, φ). One notices that
the demands for composite good z̃ and housing floor size h̃ do not depend on the flood
risk and protection effort, as the financial loss is linear with housing floor size. Using
the envelop theorem, we get the change in land value with respect to the risk is

Ṽl = −(1 − m(e, p)) < 0 (5.11)

We, then, have that the housing value declines with the risk level, which is intuitive as
the potential damage from flood risk makes the land less attractive to the city residents.

Proposition 1 An increase in risk levels has a negative impact on housing value Ṽ, while there
is no change in housing demand h̃ and demand for composite good z̃.

The objective of the urban planner, now, can be rewritten as follows.

Zb Z2π 
ES = max [Ṽ (y, a, u, e, p, l )q − κqα − CU ( p)](1 − g) − R A − cg `dφdr (5.12)
e,p,q,g
0 0

The first-order condition w.r.t the total housing construction q(r, φ) requires
  α−1 1
α −1 V
Ṽ = ακq ⇐⇒ q̃ = (5.13)
ακ

Taking derivative w.r.t. l on both sides of equation (5.13), we obtain

  α−1 1 −1
1 Ṽ Vl
q̃l = <0 (5.14)
α−1 ακ ακ

as Vl < 0 and α > 1. It states that more housing floors are constructed at locations
with higher housing value V and with lower flood risk. It is interesting to note that the
housing construction decreases with flood risk regardless of the form of the financial
loss function. From here, we obtain as well the population density at each developed
location
q̃(1 − g) 1−g
ñ = =⇒ ñl = q̃l < 0 (5.15)
h̃ h̃
which is negative as q̃l < 0, suggesting that the population density decreases with the
level of flood severity.
One notes that we have a lower bound of zero surplus at any undeveloped area
since the urban planner can choose not to lease the land from absentee landlords and
5.2. THEORETICAL MODEL 135

set n(r, φ) = 0. Therefore, a location will be developed if and only if it incurs non-
negative point-wise social surplus. In other words, the condition for a location gets
developed is

[Ṽ q̃ − κ q̃α − CU ( p)](1 − g) − R A − cg ≥ 0 (5.16)

  α−α 1
α U 1
⇐⇒ Ṽ ≥ V ( R A , p) ≡ ( R A + cg + C ( p)(1 − g)) (ακ ) α (5.17)
α−1

From equation (5.11), we can see that the flood risk level reduces the bid of households,
leading to lower housing floor value Ṽl < 0. If the effect of the flood risk is strong
enough to reduce the bid rent V at some location to be lower than the threshold V,
then this location will be left undeveloped. This corner equilibrium can lead to the
discontinuity in land use within the urban area. Specifically,

y − z̃(y, a, u)
Ṽ ≥ V ( R A , p) ⇐⇒ L( p, l ) ≡ l (1 − m(e, p)) + C H (e) ≤ − V ( R A , p)
| {z } h̃(y, a, u)
Loss due to flood hazard
(5.18)
y−z̃
h̃
− V ( R A , p) − C H (e)
⇐⇒ l ≤ l ( p, R A ) where l ( p, R A ) ≡ (5.19)
(1 − m(e, p))

Any location with average financial loss of flood risk l higher than l is not developed.
One notices that l is decreasing with distance to the city center, meaning the restriction
is stricter near urban fringe and with higher outside opportunity cost of land, all else
being equal.

Proposition 2 All locations with average flood damage lower than l ( p, R A ) threshold are de-
veloped, and the housing construction level q̃ as well as population density in those areas are
lower than those in safe areas.

5.2.2.2 Optimal prevention efforts

First-order conditions with respect to the private e(r, φ) and public protection effort
p(r, φ) from the social surplus function give us the following conditions:

Ve = 0 (5.20)
CU
p
Vp = (5.21)
q

The equation (5.20) states that the optimal level of private protection effort e is chosen
such that its marginal benefit on housing value is equal to zero. Using the equation (5.9)
in the previous section, we can rewrite the condition for optimal private prevention
effort e as
l × me (e, p) = CeH (e) (5.22)
136 CHAPTER 5. URBAN FLOOD, AMENITY LAND AND CITY STRUCTURE

It means at the optimal e, the marginal cost of an additional prevention effort is equal
to the marginal reduction in financial loss once the location is hit by a flood event.
The optimal level of public prevention effort p is such that the marginal benefit for
each unit of housing value multiplied by all housing floors at one location is equal to
the marginal cost CU p . To see how the decision differs compared to that of private pre-
vention effort e, we utilize again equation (5.9) and rewrite condition (5.21) as follows:

l × m p (e, p) × q = CU
p ( p) (5.23)

Equation (5.23) states that the marginal cost of an additional public prevention effort p
is equal to the total additional benefits (in the form of financial loss reduction) for all
households located at the same location. Taking the derivative w.r.t to average flood
loss l on both sides of equations (5.22) and (5.23), we have the following comparative
statics characteristics of the optimal private and public prevention effort ẽ, p̃ with re-
spect to the average financial loss from flood risk l.
me
ẽl = h
>0 (5.24)
−l × mee + Cee
CUp
mp + q̃
q̃2 l CU
p
p̃l = > 0 ⇐⇒ m p + q̃l > 0 (5.25)
−l × m pp + CU
pp q̃2

Interestingly, the private protection effort is an increasing function with respect to

the flood risk ẽl > 0, while the flood risk has an ambiguous effect on the optimal public
protection effort p̃. There are two channels here that affect the provision of public pro-
tection p: the financial loss in the floodplain area and the household density (in terms
of housing construction). First, a higher level of flood risk leads to a higher level of
financial loss, which induces the urban planner to provide a higher level of p (the first
term m p > 0). Second, a higher risk results in a lower level of housing construction,
leading to a lower housing floor available and lower population density (the second
term q̃l < 0); hence, the urban planner provides lower protection p.3 One also notices a
corner solution ẽ = p̃ = 0 at safe areas l = 0.

Proposition 3 The optimal private protection e is an increasing function with the level of flood
risk, while the effect of flood risk on the optimal public prevention choice p is ambiguous.

ẽl > 0
CU
p
p̃l ≶ 0 ⇐⇒ m p + q̃l ≶ 0
q̃2

In the secured area where l = 0, the optimal protection efforts are equal to zero e = p = 0.
3 As seen in condition (5.23), the marginal benefit of a public prevention p is increasing in total housing
floor constructed q or population density n.
5.2. THEORETICAL MODEL 137

5.2.2.3 Optimal provision of amenity land

The pointwise first-order condition of the social surplus function in (5.12) with respect
to g gives us the following condition for the optimal prevention efforts and amenity
provision:
θqVa (1 − g) = [qV − κqα − CU ( p)] + c (5.26)

Equation (5.26) shows the principle to provide urban amenity, which balances the
total benefits of an increase in the fraction of land used for amenity and the forgone cost
of the parcel of land used for providing amenity instead of housing and the marginal
maintenance cost c. Here, the urban planner also needs to take into account the ex-
pected loss in land value where she chooses to construct the amenity spaces. To see
how the change in the risk level influences the amenity level, we take the total differen-
tiation of equation (5.26) with respect to the risk level. We have

q̃l Ṽa (1 − g) − q̃Ṽl

gl = (5.27)
q̃(Ṽa − θ Ṽaa (1 − g))

Applying the concavity condition of the utility function with respect to amenity service
a, we have
sign( gl ) = sign (χql − Vl ) (5.28)
V (1− g )
where χ ≡ a q > 0. The sign of gl is ambiguous as both ql and Vl are negative.
There are two forces at play here: first, the sign of gl is in the opposite direction of
Vl , showing that the hazardous area with less valuable land value yields an advantage
in cost reduction. That is the reason why, ceteris paribus, when the risk level increases
leading to a lower land price as Vl < 0, it is optimal for the planner to put more amenity
land there as the cost of the land is lower. The second part of the equation relates to the
population density or the housing construction in the hazardous area. As we noted
from the previous section, ql < 0, indicating a lower level of housing floors constructed
in the risky area and therefore a lower number of inhabitants, which in turns induces
the planner to provide less amenity. These two forces work in opposite directions, and
the sign of gl is dependent on which channel is dominating at which location.

Proposition 4 Under the uncertainty of severity when a flood event hits in risky locations, the
optimal provision of land-consuming amenity facilities g needs to satisfy

θqVa (1 − g) = [qV − κqα − CU ( p)] + c

The change in g with respect to the risk level is also ambiguous and depends on the change in
land value and population density with risks:

gl ≶ 0 ⇐⇒ χql − Vl ≶ 0
138 CHAPTER 5. URBAN FLOOD, AMENITY LAND AND CITY STRUCTURE

To obtain a closed-form solution of the land use under the uncertainty of flood risk,
we need the explicit form of the utility function. In the appendices, we consider some
popular and common forms of the utility functions and derive how the change in risk
level influences urban structure and the provision of endogenous amenity. In the next
section, we consider how the urban structure and the provision of protection efforts and
local amenity change when the urban planner faces the budget constraints and needs
to work with a decentralized equilibrium market.

5.2.3 Second-best world

Suppose a competitive land market with absentee landlords and a budget-balanced

city government. We investigate the outcome of the land market when the tax systems,
amenity levels, distribution of flood severity and damage are known for all agents. We
consider two types of tax systems: a locational tax and an income tax. We show that a
locational tax4 can decentralize the first-best outcomes; hence, it inherits all the proper-
ties of the first best and is analyzed in Appendix A. However, a locational tax might not
be easy to apply since it requires the city government to collect all information on flood
damage and flood risk at different locations across cities over a long period of times.
An income tax system is much easier to collect, but as in our analysis, it deviates from
the first-best outcomes. To see how different the city structure under the income tax,
we construct our second-best city with an income tax system as follows.
In this section, we consider a tax rate γ on household income to finance the city
government’s public spending. The housing bid rent changes as households now face
a lower level of disposable income after tax (1 − γ)w. We first study housing demand
from the household side, then move to the land developers’ problem once the bid rent
function is determined, and the decisions of the city government over the provision of
public flood mitigation and amenity land.

Housing demand

The household chooses z, h, e. During a unique stage, the household finds z, s, e that
Z1  
max U (z, h, a) s.t. z + t + Rh + h l ( x )(1 − m(e, p)) + C H (e) dF ( x; r, φ) = (1 − γ)w.
z,h,e
0
(5.29)
One can interpret this as the consumer considering her permanent income (she can
borrow to smooth her consumption).
R1
Denote again l = l (r, φ) ≡ l ( x )dF ( x; r, φ). We will drop the location indicator
0
(r, φ) where there is no confusion in the notation. Each household will bid for the
4 We consider the case where a locational tax is levied on land developers for easy notation and calcu-
lation. The locational tax on households results in similar outcomes.
5.2. THEORETICAL MODEL 139

location of their preference, and the slot will be given to the highest bidder. Hence, we
can write the bid rent function as follows The household program is given by equation
(5.29). The bid rent function is given by

(1 − γ ) w − t − z
 
Υ = max − l (1 − m(e, p)) − C H (e) s.t. U (z, h, a) = u (5.30)
z,h,e h

We have utility function U is strictly increasing in z; therefore, we can write ẑ(u, h, a)

as the unique solution of equation U (z, h, a) = u, at each level of housing consump-
tion and amenity level of indifferent curve. For general form of utility, we have zh =
−Uh /Uz < 0, z a = −Ua /Uz < 0 as utility is increasing function in all z, h, a; therefore,
the consumption of composite good along the indifferent curve is decreasing with the
housing floor service h and amenity level a. Then, the bid rent function can be defined
as follows:
(1 − γ)w − t − ẑ(u, h, a)
 
Υ = max H
− l (1 − m(e, p)) − C (e) (5.31)
h,e h

We have the following first-order conditions:

ẑ − hẑh = y − v (5.32)
l × me (e, p) = CeH (e) (5.33)

where v = γw is the total amount of tax that each household contributes to the city
government’s budget. We get the solutions h(y − v, a, u), z(y − v, a, u), and e( p, l ),
which gives the equilibrium housing floor area value

y−v−z
R = V (y − v, a, u, p, l ) ≡ − l (1 − m(e, p)) − C H (e). (5.34)
h
As the left-hand side (LHS) of the equation (5.32) is a decreasing function5 for housing
floor area h, the lower disposable income of the right-hand side (RHS) results in higher
housing floor area consumption h. The new equilibrium housing bid rent R in (5.34)
is lower for each location (r, φ) compared to the case of first-best city and of locational
taxation, keeping the same level of p and l.

Housing supply

On the supply side, we assume the land is owned by absentee landlords, who have
two options to either lend the piece of land for land developers to construct residential
houses or to the agriculture sectors at R A . At each location, a land developer pays a
land price of Ψ to the landlords to build q units of housing floors that can be rented out
for households at R.6 The construction cost is κqα with α > 1 as in the previous section.
5 This is due to the concavity of the utility function in h.
6 As it is clear that the tax and land price are at each location (r, φ), so we drop the notation.
140 CHAPTER 5. URBAN FLOOD, AMENITY LAND AND CITY STRUCTURE

The residential developers, then, choose the quantity housing floor q to maximize the
total profit
max Π = Rq − κqα − Ψ
q

Taking the first-order condition w.r.t q, we have

! α−1 1
R
q= . (5.35)
ακ

Additionally, the free entry condition dictates

! α−1 1
α−1 R
Ψ= R . (5.36)
α ακ

The rent is at least equal to or higher than the outside opportunity cost of land which is
the agricultural land rent R A . Therefore, we can denote the actual land rent at location
(r, φ) is as follows:
R(Lr,φ) = max{ R A ; Ψ(r,φ) }

Therefore, a location will be developed if and only if

  α−α 1
α 1
Ψ(r,φ) ≥ R A ⇐⇒ R ≥ R( R A ) ≡ RA (ακ ) α (5.37)
α−1

From the condition (5.17), one sees that the threshold level R( R A ) is lower than that for
first-best scenario and for locational taxation (see Appendix A). This means that some
locations that are left undeveloped by land developers in the case of locational tax are
now eligible for development. The corresponding flood damage threshold is also lower.
Specifically, we have

y−v −z
− R( R A ) − C H (e)
h
l ≤ l ( p, R A ) where l ( p, R A ) ≡ (5.38)
(1 − m(e, p))

City government

The city government’s objective changes to

Zb Z2π
α 
ES = max [V (y − γw, a, u, p, l )q − κ q − CU ( p)](1 − g) − R A − cg `dφdr
p,g
0 0
(5.39)
such that U (z, h, a) ≥ u, and

Rb 2π
CU ( p)(1 − g) + cg `dφdr
R

0 0
γ= . (5.40)
w×N
5.2. THEORETICAL MODEL 141

This equation (5.40) specifies the tax rate that balances the city government’s budget.
The higher the provision for flood prevention p, the higher the tax rate γ, and since the
cost of flood protection is convex, we also have an increasing marginal tax rate with p.
In another world, we have γ p > 0 and γ pp > 0. An increase in amenity land induces
two opposite effects on γ. First, it increases the maintenance cost, hence resulting in a
higher tax rate. Second, an increase in g in risky areas means a smaller fraction of land
is used for residential purpose, leading to a smaller area for flood protection; therefore,
it induces lower cost for flood mitigation facilities and a lower tax rate. As the latter
effect is shut down in safe areas, an increase in amenity land in safe zones will lead to
an increase in the tax rate. However, the effect is unclear for an increase of amenity land
in risky zones as these two forces work in opposite directions.
Taking the first order conditions for the city government’s maximization problem
in (5.39) with respect to p, we have

q × (Vp + Vγ γ p ) = CU
p (5.41)

where Vγ is the marginal change in land bid rent with income tax γ. It is apparent
that the higher level of income tax, the lower the land bid rent. Therefore, we have
Vγ < 0, which implies lower marginal benefits on the LHS of equation (5.41) compared
to the condition with locational tax in (5.58). An increase in public protection effort p
has two effects in the case of income tax. First, it lowers the average financial loss from
flood hazard, resulting in higher bid rent from households (the first term of the LHS in
equation (5.41)). At the same time, an increase in flood protection induces a higher tax
rate, leading to less disposable income and lower bid rent (the second term). Therefore,
this leads to a lower level of public flood protection in risky areas compared to the case
of locational tax. The flood protection level in safe areas does not change and is equal
to zero.7
The implication from equation (5.41) is surprising and not straightforward. The
common insight is that income tax would lead to over-provision of flood defense sys-
tems since residents located in safer areas would partially finance the cost. Here, the
public flood protection effort is lower in risky areas under an income tax mechanism.
Additionally, it is apparent from condition (5.37) that the threshold for land develop-
ment in the case of income tax is lower than in the case of locational tax. This condition
further implies that more land in flood-prone areas could be converted into residential
land in the case of an income tax system. Then, an income tax system would lead to
an outcome in which more locations in floodplain areas are converted into residential
areas while each unit of residential land in a risky zone has less investment in terms of
public flood mitigation infrastructure.
7 Note that we have l = 0 in safe areas; hence, we have Vp = l × m p = 0. If it is replaced in equation
(1− g)`
(5.41) and rearranged, we get (1 − λVγ ) CU
p = 0 where λ = w× N > 0. Hence, CU
p = 0 ⇒ p = 0.
142 CHAPTER 5. URBAN FLOOD, AMENITY LAND AND CITY STRUCTURE

The first-order condition with respect to g gives us the following equation:

α
q(1 − g) θVa + Vγ γg = Vq − κq − CU ( p) + c.


(5.42)

Similar to the choice of p above, we also have a secondary effect of an increment in

amenity land on the tax rate and bid rent V (the second term on the LHS of equa-
tion (5.42)). As discussed above, in areas with no flood risk, we have γg > 0. Hence,
an increase in amenity land in safe zones would lead only to a higher tax rate, and a
higher tax rate imposes lower land bid rent as Vγ < 0. The total marginal benefits of
an increase in g on land bid rent are reduced by the second term, leading to a lower
level of amenities in safe areas compared to the level of g in the same area in the case
of locational tax. In risky areas, the intuition is ambiguous as there are two effects
of an increase in g in these areas on tax, as discussed above. If amenity maintenance
costs dominate, an increase in g in risky areas has the same effect as in safe areas. A
compelling case is when the decrease in the flooded protection zone dominates, which
leads to an overall reduction in tax, hence making the second term in the LHS of equa-
tion (5.42) positive. This increases the marginal benefit of an increase in g in risky areas,
resulting in a higher level of g in equilibrium compared to the case of locational tax.

Proposition 5 Under locational taxation rule, the outcomes of the first-best optimization are
decentralized through market equilibrium through locational taxation.

Proof. See Appendix A

Proposition 6 Under the income tax system, more land in risky areas is converted into resi-
dential use, while population density and housing construction are lower in both safe and risky
zones. The city government provides less flood mitigation efforts p per unit of land in risky areas
compared to the first best and the locational tax and is more likely to provide more amenity land
there compared to the optimal level in the first best case.

The analysis of the two types of tax systems confirms Henry George’s theorem,
which states that locational tax is sufficient to finance local public goods and does the
least harm as it does not distort the market. However, this type of locational tax would
not be simple to implement in reality as it requires all necessary information about the
flood risk of a property to be reflected in household decisions. However, it may be valid
with large-scale events8 as these are more likely to increase the perceived risk among
residents. With smaller-scale flood events, the literature regularly reports only short-
live effects (Atreya et al., 2013; Hallstrom and Smith, 2005; Kocornik-Mina, 2019; Zhang,
2016). The income tax system is more convenient to implement; however, it leads to
more construction in frequently flooded zones since people in safer areas partially bear
the costs. Combining income tax with a partial zoning policy for frequently flooded
areas could be a good alternative to a locational tax.
8 likethe flooding after Hurricane Sandy, which affected property values in New York in the long-run
(Ortega and Taspinar, 2018)
5.3. EMPIRICAL ANALYSIS 143

5.3 Empirical analysis

In this section, we test some predictions from the theoretical models to determine if they
reflect the reality. We test the prediction from Proposition 2 that predicts a lower level
of housing construction and a lower population density in risky areas, respectively. Re-
garding the level of flood mitigation facilities, p and the level of amenity land provision
g, the theoretical model gives ambiguous predictions. We check both flood prevention
facilities and the level of amenity land in both safe and risky zones to determine which
channel dominates in reality.

5.3.1 Data

5.3.1.1 Main sample

Flood risk is typically characterized by two main parameters: the return period and the
damage associated with a flood. The return period, or recurrence interval, indicates the
frequency that a flood event will occur. For example, a 100-year return flood hazard
describes the potential damage of a flood event that occurs once every 100 years, which
is equivalent to a 1% probability of occurrence. Similarly, a 10-year flood hazard map
shows the locations flooded once every 10 years. The second parameter is the potential
damages or the magnitude of a flood (if it occurs) at a specific location. There are two
main types of flood mapping: flood hazard maps and flood risk maps. According to
Directive 2007/60/EC of the European Commission, flood hazard maps show the ex-
tent and inundation depth of a flooded area with different return periods.9 Meanwhile,
a flood risk map indicates the potential damage to the population, economic activities,
and the environment (e.g., the pollution aftermath of a flood event) with different flood
return periods. While the return period or the frequency of flooding is similar for both
types of flood maps, the magnitudes are different. For example, the substantial inunda-
tion depth of a flood hazard in a rural or deserted area could cause almost no damage
to economic activities and result in no human casualties.
In this paper, we utilize flood hazard maps from the Joint Research Centre (JRC) of
the European Commission. Figure 1 is a 100-year return-period flood hazard map for
Europe with different levels of the inundation depth. We employ a flood hazard map
instead of a flood risk map because the flood hazard map is exogenous and relatively
independent of human economic activities. In contrast, the flood risk map takes into
account population density and economic activities when mapping and hence could
cause endogeneity in our regression. Furthermore, the flood hazard map produced by
the JRC covers all of Europe and is consistent across all European countries.10 An ad-
9 See more detail at https://ec.europa.eu/environment/water/flood_risk/flood_atlas/ (ac-
cessed on Dec 03, 2019).
10 For more detail about the methodology and quality controls, see Alfieri et al. (2014) and Alfieri et al.

(2015).
144 CHAPTER 5. URBAN FLOOD, AMENITY LAND AND CITY STRUCTURE

Figure 1: EU - Flood Hazard Map (100-year return period)

Figure 2: Europe Main River Networks and City Locations

Note: Main river networks combined of large rivers and other main rivers from WISE database. According to EEA,
large rivers are rivers that have a catchment area large than 50,000 km2 (in dark color) and other main rivers are those
that have a catchment of at least 5,000 km2 (in brighter color). Location of urban areas are taken as the location of the
main communal hall in each city. The list of cities are contained most populous 305 cities in Europe from GMES Urban
Atlas 2006 database.
5.3. EMPIRICAL ANALYSIS 145

ditional advantage is that the JRC flood hazard maps are high resolution, which allows
us to compare the potential risk at different locations across urban areas. Flood maps
like those provided by Dartmouth Flood Observatory (DFO, 2014)11 are also useful as
they outline the actual flooded areas in significant flood events; however, the cover-
age is broad and at low resolution. Further, it is impossible to derive the frequency of
flooding from the DFO maps. Hence, this type of map cannot be used to investigate the
effect of flood risk for small, local zones within urban areas such as those discussed in
the theoretical part.
The JRC flood hazard maps include six return periods: 500-, 200-, 100-, 50-, 20-,
and 10-year return periods. For each return period, the flood hazard maps indicate
the inundation depth at each location using gridded cells. To implement our research
design, we determine the flood exposure level as the average inundation depth at each
1km × 1km grid cell for all return periods. This exposure level is used as a proxy for
the magnitude of the potential damage that households face if they choose to locate
there. We also calculate the flood frequency as the inverse of the return period (i.e., 1%
probability for a once-in-100-years event) for each gridded cell in our database. Any
cells within the safe area have zero probability of getting hit by a flood event, and no
flood exposure is measured. We calculate this for all 1km × 1km gridded cells in the
305 most populous cities in Europe. Among these cities, 221 have at least one grid
measuring 1 km2 that contains a positive flood risk level, which accounts for 72% of the
cities in our sample. As shown in Figure 2, these cities are located close to main river
networks. In this section, we restrict our samples to cities in which a significant part of
the city faces flood risk. Therefore, we choose only cities with at least 20% of the city
area under flood hazard, which results in a sample of 162 cities.
To investigate the land uses within urban areas, we follow Picard and Tran (2019)
and use the GMES Urban Atlas. This dataset offers a high-resolution land use map
of urban areas, with information derived from Earth observation (EO) and backed by
other reference data such as commercial off-the-shelf (COTS) navigation data and to-
pographic maps. The GMES Urban Atlas uses EO satellite images with 2.5-m spatial
resolution and multi-spectral data, which offers much higher resolution compared to
traditional CORINE land cover resolution 12 . It also covers 305 of the most populous
towns and cities in Europe (EU 27); the data include all EU capitals plus a large sample
of medium-sized cities. We focus on the ‘urban fabric’ classes, which specify differ-
ent levels of continuity of residential areas13 , as well as different types of amenity land
uses such as green urban areas (14100) and sports and leisure facilities (14200). To calcu-
late population density, we combined these two datasets with the Eurostat population
11 Kocornik-Mina et al. (2019) uses the DFO flood database.
12 According to the European Environmental Agency, the current resolution of the GMES Urban Atlas is
100 times higher than that of CORINE land cover.
13 the class codes are 11100, 11210, 11220, 11230, and 11240.
146 CHAPTER 5. URBAN FLOOD, AMENITY LAND AND CITY STRUCTURE

Figure 3: Paris Urban Boundaries and River Buffer Zones

grids to measure the density in locations across cities. We chose the city center (CBD)
for each city at the location of the old town hall. The distance of each gridded cell to
the city center is the Euclidean distance from the centroid of the grid to the CBD.
Most of the flood risk cities are close to river networks and experience moderate
to high rainfall. As outlined in the previous section, the effects of location in a flood-
plain are difficult to measure due to the dual effect of water proximity. Many studies
have reported a positive effect of being close to water (Luttik, 2000; Or, 2017, chap.
5). Therefore, in this section, we first restrict our sample to grids that are within a 10-
kilometer buffer of main river networks14 . Figure 3 illustrates the buffer zone from the
main river network in an urban area for the city of Paris. By this restriction, we can
control for river proximity and check, for the same distance from the CBD, and the
same distance from the river, whether, how, and at which magnitude the flood hazard
changes household location decisions and public provision of amenity lands. We also
run a robustness check for the whole city sample and for different buffer zones around
the river networks, the results of which are provided in the Appendix.
The summary statistics are indicated in Table 5.2. We define a dummy variable
Flood Exposure that has a value of 1 if a grid has a positive probability of flooding and
0 otherwise. The average distance to the city center for the whole sample and for the
flooded sample is not significantly different. A grid in a risky zone is, on average, 11.41
km away from the CBD, while the average for a whole city is 11.53 km. However,
14 For maps and a definition of main river networks in Europe, please check the WISE database and
Figure 4. We can adjust the buffer zone to 5, 7.5, and 12.5 kilometers as a robustness check. However, the
results do not change significantly.
5.3. EMPIRICAL ANALYSIS 147

Table 5.2: Summary Statistics

Panel A: Urban grids

Full River Sample Flood Exposure = 1
Mean Std. Dev. Obs. Mean Std. Dev. Obs.
Population (thous. hab./km2 ) 2.29 2.71 22239 2.28 2.47 9319
Distance to CBD (km) 11.53 8.60 22239 11.41 9.14 9319
Avg. Elevation (m) 118.81 110.99 22239 101.58 108.57 9319
Relative Avg. Elevation (m) 17.55 42.87 22239 -2.77 22.69 9319
Ruggedness 7.76 10.07 22239 7.15 10.92 9319
Distance to River (km) 3.80 2.74 22239 2.48 2.43 9319
Amenity Land (ha) 8.10 11.71 22239 8.50 11.40 9319
Commercial and Public Bld. (ha) 13.43 14.67 22239 14.44 14.94 9319
Flood Frequencies (per annum) 0.07 0.09 22239 0.16 0.06 9319
Flood Exposure 0.42 0.49 22239
Avg. Inundation Depth Level (IDL) (m) 2.02 1.61 9319
Panel B: Within 20% cut-off boundaries
Full River Sample Flood Exposure = 1
Mean Std. Dev. Obs. Mean Std. Dev. Obs.
Population (thous. hab.) 3.17 3.36 11269 3.20 3.04 4516
Distance to CBD (km) 7.07 4.76 11269 6.05 4.54 4516
Avg. Elevation (m) 112.38 100.63 11269 90.90 97.59 4516
Relative Avg. Elevation (m) 19.22 45.67 11269 -3.82 20.14 4516
Ruggedness 8.01 10.76 11269 6.88 10.98 4516
Distance to River (km) 3.61 2.66 11269 2.21 2.24 4516
Amenity Land (ha) 10.08 13.37 11269 10.60 13.04 4516
Commercial and Public Bld. (ha) 14.75 15.22 11269 15.94 15.29 4516
Flood Frequencies (per annum) 0.07 0.09 11269 0.16 0.06 4516
Flood Exposure 0.40 0.49 11269
Avg. Inundation Depth Level (m) 2.16 1.55 4516
Note: This table indicates the summary statistics from our samples of 1 km2 grid cells. All grids are
restricted to those within 10 kilometres from the main river networks. Urban cells are those that have
at least 20% of land as artificial land use. Urban boundaries with 20% cut-off points are the distance
to city center of the rings that have at least 20% of land are urban fabric or residential land use. See
Paris map for the illustration. Relative Elevation is the difference between the grid’s elevation and
the elevation of the CBD of the same city. Ruggedness are measured as the standard deviation of the
elevation within the grids. Amenity lands include green urban areas and the land used for sport and
leisure activities.

flood zones are more likely to be at a lower elevation, less rugged, and closer to river
networks. More precisely, a grid located in a flood-prone area is, on average, 1.3 km
closer to a main river network. Flood zones are also more likely to be located at a lower
elevation; the average elevation difference between a grid located in a safe area and
one in a risky zone is 18 m. On average, a location in a flood zone has fewer residents
and a higher amount of amenity land. Ruggedness index values are also lower in flood
zones. Therefore, if the hypothesis is that it is easier to build a home in a less rugged
area, this means that it is less expensive to build a home close to a river. Therefore, we
would expect negative ruggedness values for areas with a high population density.
If we consider only grids that are at least 20% inside the cut-off boundary of an
urban zone, the sample is reduced to nearly half of the whole sample. The average
distance to the CBD is reduced to around 7 km instead of 11 km as before, and it is
148 CHAPTER 5. URBAN FLOOD, AMENITY LAND AND CITY STRUCTURE

6 km for grids in flood zones. Flood frequencies are calculated as the inverse of the
return period. On average, a grid in a floodplain area has approximately 0.16 floods
per annum. The frequency for a grid in the whole sample is approximately half, at
0.07. In addition, risk zones suffer higher inundation once a flood occurs. The average
inundation depth for risk zones is 2.16 m while the average for the whole sample is 0.86
m. We later use the inundation level as a proxy for the potential damage once a location
is hit by a flood event.

5.3.1.2 Supplemental data

Building heights

We combine our main datasets with building height data from the European Environ-
mental Agency (EEA) under the framework of the Copernicus program. This dataset
contains information on building block height at a spatial resolution of 10 m for the ref-
erence year of 2012.15 The information on building height allows us to test the housing
supply channel, as stated in Proposition 2. In particular, we use the data to obtain infor-
mation about the vertical density of housing and to investigate if there are differences
in housing structures in floodplain areas compared to those that are in safety zones.
However, the data provided by EEA covers only the capital cities across EU28. To com-
plement studies on housing with flood hazard, we utilize data on different categories
of residential surfaces in the GMES Urban Atlas.

Table 5.3: Building Height Summary Statistics

Panel A: Urban grids

Full River Sample Flood Exposure = 1
Mean Std. Dev. Obs. Mean Std. Dev. Obs.
Avg. Building Height (m) 2.53 2.42 4391 2.47 2.41 1417
Avg. Height of Highest Building (m) 32.74 24.38 4391 34.10 27.52 1417

Panel B: Within 20% cut-off boundaries

Full River Sample Flood Exposure = 1
Avg. Building Height (m) 2.85 2.56 3507 2.94 2.54 1074
Avg. Height of Highest Building (m) 35.54 26.24 3507 39.06 30.82 1074
Note: This table indicates the summary statistics of building heights from our samples of 1 km2 grid
cells. All grids are restricted to those within 10 kilometres from the main river networks. Urban cells
are those that have at least 20% of land are artificial land use. Urban boundaries with 20% cut-off
points are the distance to city center of the rings that have at least 20% of land are urban fabric or
residential land use.

Table 5.3 reports the basic statistics for our building height database for all capital
cities with at least 20% of land under flood risk. The average building height for all
grids in the sample is 2.53 m, which is equivalent to a one-story building. The height
15 Full details are at https://land.copernicus.eu/local/urban-atlas/building-height-2012?tab=

metadata (accessed on Dec 06, 2019).

5.3. EMPIRICAL ANALYSIS 149

for flood exposed grids only is slightly lower at 2.47 m. The average height of the
tallest building for all urban grids is 32.74 m, which approximately equal to a 10-story
building. The height for floodplain grids is surprisingly a bit more at 34.10 m. When
including only the grids within a circular urban border (which is defined as the border
at which at least 20% of the land is residential), the number is higher on average. This
indicates that the sample within the circular border is closer to the dense core area of
the city. In our sample, the lowest point of a building block is 0 m, indicating a location
at which there is no building, and the highest building block is 368 m, which is located
in Berlin’s city center.

Public flood prevention data

The second set of supplemental data that we use in this paper is for flood preven-
tion infrastructure in France. We obtained the data from the French government on
the Georisques16 . The dataset contains information about the length and location of the
flood prevention infrastructure currently in place in France’s territories. It mostly con-
tains information on dykes and pumping stations. In this paper, we consider only the
infrastructure classified as “Ouvrage de protection”. An illustration of the flood pre-
vention infrastructure for the city of Paris is included in Appendix B. Currently, there
is no systematic data on flood mitigation facilities for all cities in Europe.

Table 5.4: Flood Prevention Summary Statistics

Full River Sample Flood Exposure = 1

Mean Std. Dev. Obs. Mean Std. Dev. Obs.
Flood Prevention Infrastructures (m)
A: All urban grids 50.83 267.71 6104 130.65 418.27 2279
B: Within 20% cut-off borders 72.91 325.59 3662 215.10 530.57 1232
Note: This table indicates the summary statistics of flood prevention infrastructures in France.
All grids are restricted to those within 10 kilometres from the main river networks. Urban cells
are those that have at least 20% of land are artificial land use. Urban boundaries with 20% cut-
off points are the distance to city center of the rings that have at least 20% of land are urban
fabric or residential land use.

Table 5.4 describes the summary statistics for flood prevention facilities measured
as the length of the “Ouvrage de protection” for all the grids in our sample. The city
sample includes all French cities in both the GMES Urban Atlas data in which at least
20% of the land is under flood hazard. The length for the grids within the floodplain
areas is at least two to three times the average for the full samples. Another important
detail is that the average flood prevention within the urban core (the 20% cut-off bor-
der) is higher than for the sample with all urban grids, indicating a possible different
treatment in terms of the provision of flood facilities between the urban core and urban
fringes.
16 Link: https://www.georisques.gouv.fr.
150 CHAPTER 5. URBAN FLOOD, AMENITY LAND AND CITY STRUCTURE

In the next part, we study the effect of flood hazard on city structure. Specifically,
we examine its effects on population density, level of housing construction, public flood
prevention facilities, and local amenity land. We also compare the predictions from the
theoretical part and the outcomes in reality.

5.3.2 Population density in floodplain areas

From the theoretical model, we predict that the population density decreases along with
the level of flood risk. However, in reality, most risky areas are in close proximity to a
water body such as a main river. For a simple check, we can see that grids that are close
to a main river network exhibit a higher risk of flooding. It is well-documented in the
literature that households gain some amenities by being close to a river or canal (e.g.,
ease of walking, view). Historically, being close to a river contributed to better defense
and a better transportation network. Therefore, we follow the strategy of Berstein et al.
(2019) and create buffer zones around river networks, restricting our grid samples to
this zone to further control for the effect of river networks.
As discussed above, there are two main characteristics of flooding, namely fre-
quency and magnitude. In this section, we examine the effects of both on population
density. The reduced-form regressions of the effect of flood exposure on household lo-
cation choice or population are specified below. Consider an observation (i, r, c) where
i refers to 1 km × 1 km cells, r is the distance to the city center, and c is the city to which
the grids belong. Our estimation model is as follows:

ln Popirc = βFloodDirc + θDistrc + ∑ αc Ic + ∑ Xirc + eirc (5.43)

c

where FloodD is our explanatory variable of interest. We consider three different mea-
sures for the FloodD variable. First, we utilize the indicator Flood Exposure, which is
equal to 1 if the grid has a positive risk of inundation and 0 otherwise. The second
possible measure is flood frequency, as stated in a previous section. The third measure
is the magnitude of flooding for each location, which is proxied by different inundation
depth levels (IDL). We also use Xirc as a grid control. It absorbs the variation in popu-
lation density due to the interaction between relative location and land characteristics
such as ruggedness (lower cost of building) and elevation. Specifically, we control for
distance to the river network17 , a grid’s elevation above sea level, and its ruggedness.
City fixed effects are also used to control for unobserved city-wide policies.
The baseline regression results reported in Table 5.5 provide initial evidence of the
effect of flood exposure on population density. In Column (1), we simply run the re-
gression of the natural logarithm of population density with only the controls for city
fixed effect and flood exposure. The significantly positive coefficients for Flood Exposure
17 We can also follow Bernstein et al.
(2019) to introduce distance to river categories instead of continuous
distance. In this way, we can avoid the assumption of linear effects.
5.3. EMPIRICAL ANALYSIS 151

indicate that flood-prone grids are associated with higher population density, which is
consistent with Kocornik-Mina et al. (2019). The same regression with Flood Frequen-
cies also gives similar results, as shown in Column (1) of the second panel of Table 5.5.
These results are not surprising, given that, on average, the grids in floodplain areas are
more proximate to river and canal networks. In Column 2, we further control for the ge-
ographical characteristics of each grid, such as the ruggedness and elevation level. The
coefficients for Flood Exposure are still positive, but no longer significant after clustering
the standard errors at the city level. The result in Column (3) shows that the correlation
between flood exposure and the proximity to the water drives the positive relationship
between flood exposure and population density. After controlling for the distance to
river networks, and 2-km from the coast fixed effects, the relationship between popu-
lation density and flood exposure becomes significantly negative. Any further controls
for possible variables with possible influence on population density do not alter the
sign and significance of the coefficients for flood exposure on population density.

Figure 4: Inundation Depth Level and population density

This figure denotes the relationship between the % change in population density of the grids within the zones with
flood exposure (in comparison with unexposed grids), partitioned by different levels of inundation depth level. We
categorize the flood hazard according to the exposure level from Very Low to Very High, which are defined as the
average Inundation Depth Level (IDL) in all return periods. We set the IDL between 0 and 1 meter as very Low level
and above 6 meters in Very High level.

The estimated coefficients for Flood Exposure of −0.085 and −0.082 in Columns (5)
and (8), respectively, with the strictest controls, suggest that a grid in a flood-prone
areas has 8.2% to 8.5% lower population relative to an unexposed grid in the same lo-
cation in the city center and with similar characteristics. Similarly, the regression for
 Table 5.5: Population density and flood hazard - Main regression results
CHAPTER 5. URBAN FLOOD, AMENITY LAND AND CITY STRUCTURE

Dependent variable: Ln(Pop)

Panel A: (1) (2) (3) (4) (5) (6) (7) (8)
Flood Exposure 0.168∗∗∗ 0.009 −0.083∗∗ −0.090∗∗∗ −0.085∗∗∗ −0.256∗∗∗ −0.091∗∗∗ −0.082∗∗∗
(0.033) (0.035) (0.037) (0.028) (0.028) (0.048) (0.030) (0.029)
Dist. to CBD Y Y Y Y Y Y Y Y
City FE Y Y Y Y Y Y Y Y
Geographical controls N Y Y Y Y Y Y Y
Dist. to river N N Y Y Y Y Y Y
Natural Amenity Controls N N N Y Y N Y Y
Artificial Amenity Controls N N N N Y N N Y
River Buffer Zones < 10 km < 10 km < 10 km < 10 km < 10 km < 10 km < 10 km < 10 km
Grid Sample Urban grids Urban grids Urban grids Urban grids Urban grids Within 20% Within 20% Within 20%
cut-off border cut-off border cut-off border
Adjusted R2 0.374 0.398 0.405 0.598 0.623 0.427 0.691 0.711
N 22,121 22,121 22,121 22,121 22,121 11,172 11,172 11,172
Panel B: (1) (2) (3) (4) (5) (6) (7) (8)
Flood Frequencies 0.751∗∗∗ −0.102 −0.633∗∗∗ −0.636∗∗∗ −0.569∗∗∗ −1.964∗∗∗ −0.887∗∗∗ −0.790∗∗∗
(0.186) (0.195) (0.211) (0.163) (0.162) (0.323) (0.182) (0.180)
Dist. to CBD Y Y Y Y Y Y Y Y
City FE Y Y Y Y Y Y Y Y
Geographical controls N Y Y Y Y Y Y Y
Dist. to river N N Y Y Y Y Y Y
Natural Amenity Controls N N N Y Y N Y Y
Artificial Amenity Controls N N N N Y N N Y
River Buffer Zones < 10 km < 10 km < 10 km < 10 km < 10 km < 10 km < 10 km < 10 km
Grid Sample Urban grids Urban grids Urban grids Urban grids Urban grids Within 20% Within 20% Within 20%
cut-off border cut-off border cut-off border
Adjusted R2 0.373 0.398 0.406 0.599 0.624 0.432 0.693 0.712
N 22,121 22,121 22,121 22,121 22,121 11,172 11,172 11,172
Note: This table reports the OLS estimates where the dependent variable is the natural log of population count in 1k × 1km grid. The explanatory variable of interest
is the Flood Exposure which equal to 1 if the cell is in floodplain area and 0 otherwise. We, then, further use the flood annual frequencies (which is the inverse of
flood return periods). Column (1) to (5) includes all the grids with urban status equal to 1. The cells that have urban status equal to 1 are those with at least 20% of
artificial land use within the grid. Column (6) to (8), we consider all cells within a circular urban boundary. Urban boundaries are defined as the circular cut-off at
which the total urban fabric (or residential land use) are of at least 20% of total land within its rings. For more details, see Picard and Tran (2019). Other controls
152

include the coastal grid indicator. Significance levels are denoted by ∗ for p<0.1, ∗∗ for p<0.05 and ∗∗∗ for p<0.01. Standard errors are clustered at city level and
reported in parentheses.
5.3. EMPIRICAL ANALYSIS 153

Flood Frequencies in Columns (5) and (8) in Panel B gives the coefficients of −0.569 and
−0.790, respectively. This suggests that an increase of flood frequency of 0.01 (equiva-
lent to one additional flood event in 100 years, or an increase of 1% in flood probability)
leads to a reduction of 0.57% to 0.79% in population density. In other words, an increase
of one additional flood event every 10 years is associated with a decrease population
density of approximately 5.6% to 7.9%.
Figure 4 illustrates the effect of Flood Exposure on population density using a more
comprehensive measure of flood exposure. In the figure, we estimate the regression
(5.43) again, but we replace FloodD with a series of indicators of the IDL. This specifi-
cation allows us to look at the non-linear relationship between IDL or potential flood
damage and the population density. Across all interactions, we see a statistically signif-
icant and negative exposure level above 2 m in IDL. This means that all exposed IDLs
above 1 m are related to a population decrease relative to those in safe areas. When
we consider the sample with all urban grids, the population decrease begins even at a
deficient IDLs. The greatest decline is at locations with an IDL above 2 m, with a mag-
nitude between 10% and 30%. It is surprising to see a lower magnitude for the highest
IDL (above 6 m), which could be explained by the lower initial settlements or zoning
policies in highly inundated areas.

5.3.3 Housing construction density

Proposition 2 in our theoretical model predicts that the total floor area of housing con-
structed in floodplain areas is lower compared to the area in safety zones if everything
else is equal. This is also the main channel in our model, which results in lower popu-
lation density in flooded zones as the choice of housing floor area for each household
does not change with flood hazard. In other words, lower population density is mainly
due to lower housing supply in risky zones. Therefore, it is necessary to check if this
mechanism matches what happens in reality. We thus run the following regression

HCLirc = β h FloodDirc + θh Distrc + ∑ αc Ic + ∑ Xirc + eirc (5.44)

c

where HCLirc is the housing construction level of grid i at distance r from CBD of city
c.
As discussed above, we use two types of measures for the HCL, namely average
building height and residential land classification. An advantage of the building height
database is that it precisely measures the vertical density of housing construction. How-
ever, the building height data does not cover all of Europe, only the capital cities. In
contrast, the residential land classification data cover the most populous cities across
EU27; however, it measures the horizontal density of sealed surfaces. Hence, we first
run a simple regression with building height as the dependent variable and different
classes of residential urban fabric as the explanatory variables. This regression reveals
154 CHAPTER 5. URBAN FLOOD, AMENITY LAND AND CITY STRUCTURE

the relationship between building height and different classes of residential land. In
particular, we use all available classes for residential land, including continuous fabric
(H11) with at least 80% of residential land as sealed surfaces (S.L.). The other classes for
residential land range from dense, discontinuous urban fabric with S.L. between 50%
and 80% (H21) to the very low density with S.L. below 10% (H24).

Table 5.6: Correlation between building height and different residential land classifica-
tions

Dependent variable: Avg. Building Height

(1) (2) (3) (4) (5) (6)
Share of residential land categories
H11 Continuous Fabric 10.638∗∗∗ 10.767∗∗∗
(Sealing Surface (S.L.) > 80%) (1.146) (1.266)
H21 Discontinuous Dense Fabric −0.044 1.306
(S.L. 50% - 80%) (1.295) (0.819)
H22 Discontinuous Medium −5.051∗∗∗ 0.325
Density Fabric (S.L. 30% - 50%) (1.301) (0.521)
H23 Discontinuous Low Density −7.785∗∗∗ −1.374
Fabric (S.L. 10% - 30%) (1.742) (1.065)
H24 Discontinuous Very Low −13.433∗∗∗ −5.307∗∗∗
Density Fabric (S.L. < 10%) (0.704) (0.862)
City FE Y Y Y Y Y Y
Adjusted R2 0.597 0.070 0.104 0.090 0.075 0.606
N 4,391 4,391 4,391 4,391 4,391 4,391
Note: Significance levels are denoted by ∗ for p<0.1, ∗∗ for p<0.05 and ∗∗∗ for p<0.01. Standard errors are
clustered at city level and reported in parentheses.

Table 5.6 describes the first regression results. The standard errors are all robust and
clustered at the city level. Columns (1) through (5) show the results of the regressions
of average building height for each class of residential land use, and Column (6) reports
the regression results for all classifications together. Continuous residential urban fab-
ric (H11) with S.L. above 80% has the highest explanation power with a significantly
positive coefficient and adjusted R2 equal to 0.597. The other classes of residential land
have low explanation power with adjusted R2 ranging from 0.070 to 0.104, which are
just a sixth of the adjusted R2 for the H11 class. The coefficient for the second residential
fabric class (H21) is not significant, and the coefficients for the lower residential fabric
classes (H22, H23, and H24) are all significantly negative. The results in Column (6) are
similar when we rerun the regression with all residential urban fabric classes. These
results suggest that building height is highly associated with dense, continuous urban
areas. This is unsurprising as we mostly observe high-rise buildings in very dense core
and central areas. Hence, we utilize both building height and the type of continuous
urban fabric (H11) as proxies for housing construction density level (HCL).
Table 5.7 reports the OLS estimates for Flood Exposure in regression (5.44) with build-
ing height and share of continuous residential fabric as proxies for HCL. Panel A in Ta-
ble 5.7 shows the first results for the regression with building height levels and different
 Table 5.7: Housing construction and flood hazard - regression results

Panel A: Dependent variable: Building Height

(1) (2) (3) (4) (5) (6) (7) (8)
Flood Exposure −0.012 −0.298∗ −0.377∗∗ −0.322∗∗ −0.262∗ −0.342∗∗ −0.328∗∗ −0.249∗
(0.146) (0.181) (0.183) (0.144) (0.155) (0.153) (0.136) (0.134)
Adjusted R2 0.415 0.452 0.460 0.526 0.601 0.474 0.541 0.624
N 4,391 4,391 4,391 4,391 4,391 3,507 3,507 3,507

Panel B: Dependent variable: Share of Continuous Residential Categories (S.L > 80%)
5.3. EMPIRICAL ANALYSIS

(1) (2) (3) (4) (5) (6) (7) (8)

Flood Exposure 0.002 −0.008∗ −0.018∗∗∗ −0.019∗∗∗ −0.018∗∗∗ −0.036∗∗∗ −0.028∗∗∗ −0.024∗∗∗
(0.005) (0.005) (0.005) (0.004) (0.004) (0.007) (0.005) (0.005)
Adjusted R2 0.212 0.223 0.233 0.348 0.420 0.363 0.411 0.500
N 22,121 22,121 22,121 22,121 22,121 11,172 11,172 11,172

Dist. to CBD Y Y Y Y Y Y Y Y
City FE Y Y Y Y Y Y Y Y
Geographical controls N Y Y Y Y Y Y Y
Dist. to river N N Y Y Y Y Y Y
Natural Amenity Controls N N N Y Y N Y Y
Artificial Amenity Controls N N N N Y N N Y
River Buffer Zones < 10 km < 10 km < 10 km < 10 km < 10 km < 10 km < 10 km < 10 km
Grid Sample Urban grids Urban grids Urban grids Urban grids Urban grids Within 20% Within 20% Within 20%
cut-off border cut-off border cut-off border
Note: This table reports the OLS estimates where the dependent variable is the average building height and the share of continuous residential land (with sealing
surface (S.L.) > 80%) as proxies for housing construction level (HCL) in 1km × 1km grid. The explanatory variable of interest is the Flood Exposure which equal to
1 if the cell is in floodplain area and 0 otherwise. We, then, further use the flood annual frequencies (which is the inverse of flood return periods). Column (1) to
(5) includes all the grids with urban status equal to 1. The cells that have urban status equal to 1 are those with at least 20% of artificial land use within the grid.
Column (6) to (8), we consider all cells within a circular urban boundary. Urban boundaries are defined as the circular cut-off at which the total urban fabric (or
residential land use) are of at least 20% of total land within its rings. For more details, see Picard and Tran (2019). Other controls include the coastal grid indicator.
Significance levels are denoted by ∗ for p<0.1, ∗∗ for p<0.05 and ∗∗∗ for p<0.01. Standard errors are clustered at city level and reported in parentheses.
155
156 CHAPTER 5. URBAN FLOOD, AMENITY LAND AND CITY STRUCTURE

levels of control as in the previous section. The grid samples are restricted to capital
cities with at least 20% of the city area under flood hazard. The regression results are
similar to those of the population density regression. Column (1), with a simple control
for the city fixed effects and distance to the CBD, suggests a negative but insignificant
effect on building height. Once we further control for geography and proximity to wa-
ter, the coefficient is significant and negative. Using the coefficients in Columns (5)
and (8), we find a 0.25 to 0.26 m decrease in building height in grids exposed to flood
risk relative to unexposed grids. This measure accounts for approximately 9% of the
average building height (2.5 m, as reported in Table 5.3).
Panel B of Table 5.7 reports the regression (5.44) with the share of continuous and
dense urban fabric as a proxy for HCL. The number of observations is considerably
higher than in Panel A, which is reasonable as the building height database covers
only part of the datasets in Panel B. The results are robust even with the change in
measurement for HCL. A grid situated within a floodplain has a 1.8% to 2.4% lower
share of the continuous and dense residential urban fabric, as indicated in Column (5)
and Column (8), respectively. These effects are robust with different control levels and
urban cut-offs. The results in this section provide initial evidence for our theoretical
prediction in Proposition 2.

5.3.4 Flood prevention infrastructure systems

In this section, we investigate the effects of flood hazard on the provision of public
flood prevention facilities. As stated in Proposition 4, the change in flood mitigation
facilities is not linear to the level of potential flood damage. In fact, the first derivative
of public flood protection effort p with respect to average financial loss from flooding is
ambiguous and depends on whether the increased marginal cost or the reduced benefit
coverage dominates. To test the effect of flood hazard on the public protection level,
we utilize data from France, which provide the length and location of flood protection
facilities as shown in the supplemental data section above. In particular, we run the
following regressions:

FloodPrevirc = β f FloodDirc + ∑ αc Ic + ∑ Xirc + eirc (5.45)

c
4
FloodPrevirc = FloodDi × ∑ ( β f k DistQuartilekc ) + ∑ αc Ic + ∑ Xirc + eirc (5.46)
k =1 c

where FloodPrevirc is the length of flood prevention facilities within a grid of 1 km ×

1 km. The regression (5.45) initially examines the overall effect of flood exposure and
flood frequency on the provision of flood mitigation facilities. The second regression in
(5.46) further introduces the interaction between flood exposure and different quartiles
of the distance to CBD. This allows us to look at the non-linear relationship between
flood exposure at different locations in urban areas.
5.3. EMPIRICAL ANALYSIS 157

Table 5.8: Regression results: Flood prevention infrastructure and flood hazards

Dependent variable: Flood Prevention Infrastructures Level (km)

(1) (2) (3) (4) (5) (6)
Flood Exposure 0.100∗∗∗ 0.161∗∗∗
(0.019) (0.033)
Flood Frequency 0.610∗∗∗ 1.014∗∗∗
(0.120) (0.198)
Flood Exp. × Dist. to CBD 0.177∗∗∗ 0.290∗∗
(1st quartile) (0.039) (0.126)
Flood Exp. × Dist. to CBD 0.039∗∗ 0.184∗∗∗
(2nd quartile) (0.019) (0.054)
Flood Exp. × Dist. to CBD 0.034 0.152∗∗∗
(3rd quartile) (0.023) (0.047)
Flood Exp. × Dist. to CBD 0.014 0.099∗∗
(4th quartile) (0.012) (0.039)
Dist. to CBD Y Y Y Y Y Y
City FE Y Y Y Y Y Y
Geographical controls Y Y Y Y Y Y
Dist. to river Y Y Y Y Y Y
Natural Amenity Controls Y Y Y Y Y Y
Artificial Amenity Controls Y Y Y Y Y Y
River Buffer Zones < 10 km < 10 km < 10 km < 10 km < 10 km < 10 km
Grid Sample Urban grids Within 20% cut-off border
Adjusted R2 0.142 0.147 0.160 0.163 0.174 0.170
N 6,104 6,104 6,104 3,662 3,662 3,662
Note: This table reports the OLS estimates where the dependent variable is the flood prevention infras-
tructure within 1km × 1km grids. The explanatory variable of interest is the Flood Exposure which equal
to 1 if the cell is in floodplain area and 0 otherwise. We, then, further use the flood annual frequencies
(which is the inverse of flood return periods). Column (1) to (3) includes all the grids with urban sta-
tus equal to 1. The cells that have urban status equal to 1 are those with at least 20% of artificial land
use within the grid. Column (4) to (6), we consider all cells within a circular urban boundary. Urban
boundaries are defined as the circular cut-off at which the total urban fabric (or residential land use) are
of at least 20% of total land within its rings. For more details, see Picard and Tran (2019). Other controls
include the coastal grid indicator. Significance levels are denoted by ∗ for p<0.1, ∗∗ for p<0.05 and ∗∗∗
for p<0.01. Standard errors are clustered at city level and reported in parentheses.

Table 5.8 reports the regression results. We consider all the possible controls that
we apply in the previous section. Columns (1) and (4) indicate the effect of Flood Expo-
sure on the level of flood infrastructure. Specifically, for a grid located in a hazardous
area, there is an approximately 0.1 to 0.16 km increase in the length of flood prevention
infrastructure. The results in Columns (2) and (5) also show a significant and positive
coefficient for flood frequency, suggesting that an increase in flood frequency leads to a
higher level of flood protection, on average. In particular, an increase of one additional
flood event every 10 years (equivalent to a 0.1 increase in annual flood frequency) leads
to an increase of 0.06 to 0.1 km in the length of flood prevention facilities. Columns (3)
and (4) show the estimation results of the regression (5.46). The regression results show
positive coefficients for all interactions between Flood Exposure and different quartiles
of distance to the CBD. Interestingly, the magnitude of the estimators decreases with
higher distance quartiles. Specifically, using the results in Column (6), a grid located in
158 CHAPTER 5. URBAN FLOOD, AMENITY LAND AND CITY STRUCTURE

a floodplain zone in the city core (within the first distance quartile) has 0.29 km more
flood prevention facilities. Meanwhile, a grid in a floodplain zone in the urban fringe
(the last quartile) has only 0.099 km more of such facilities, which is just one-third the
magnitude for the city core.

Figure 5: Flood prevention infrastructures and IDL exposure

This figure demonstrates the relationship between the change in flood prevention infrastructure (measured as kilome-
ters of “Ouvrage de protection”) relative to the unexposed location and the IDL exposure categories. These coefficients
are based on the regression of the stock of the flood prevention infrastructures and five different categories of IDL from
Very Low level (0-1m) and Very High level (above 6m) will all controls as in Column (5) of Table 5.8.

Figure 5 illustrates the effect of IDL exposure on the amount of flood prevention
facilities. In the figure, we estimate equation (5.45) but substitute Flood Exposure with a
series of IDL indicators. The regression with different indicators of IDL can reveal the
non-linear relationship between the level of potential flood damage and the provision
of flood defense systems. Across all coefficients in Figure 5, we see a positive and signif-
icant increase in the provision of flood defense systems with all IDL indicators relative
to unexposed grids. However, the magnitude is not linear with IDL. In particular, we
observe an inverted U-shape pattern for the change in flood defense systems and the
level of potential flood damage, as shown by IDL indicators. Flood prevention facilities
first increase with the IDL until it is equal to 4 m; after this point, it starts to decrease.
This means that in our sample, the increase in flood protection coverage dominates up
to an IDL of 4 m. After this point, the increase in the marginal cost of providing an ad-
ditional unit of flood prevention dominates and leads to a decrease in flood prevention,
even with a higher level of flood hazard or potential flood damage. Another expla-
5.3. EMPIRICAL ANALYSIS 159

nation probably is the zoning policies in highly inundated areas, resulting in lower
population, housing construction and lower flood mitigation infrastructures.

5.3.5 Land use for local public goods within flooded zone

In this section, we study the effect of flood hazard on the provision of amenity land. The
theoretical model shows an ambiguous relationship between the provision of amenity
land and the level of flood hazard (see Proposition 4).

Table 5.9: Regression results: Amenity land uses and flood hazards

Dependent variable: Share of Amenity Land Use

(1) (2) (3) (4) (5) (6)
Flood Exposure 0.003∗ 0.006∗∗
(0.002) (0.003)
Flood Frequency 0.030∗∗∗ 0.062∗∗∗
(0.010) (0.016)
Flood Exp. × Dist. to CBD 0.005∗∗ −0.007
(1st quartile) (0.002) (0.007)
Flood Exp. × Dist. to CBD −0.0004 0.015∗∗∗
(2nd quartile) (0.003) (0.004)
Flood Exp. × Dist. to CBD 0.002 0.005
(3rd quartile) (0.004) (0.004)
Flood Exp. × Dist. to CBD 0.006 0.005
(4th quartile) (0.006) (0.004)
Dist. to CBD Y Y Y Y Y Y
Dist. to CBD (sq) Y Y Y Y Y Y
City FE Y Y Y Y Y Y
Geographical controls Y Y Y Y Y Y
Dist. to river Y Y Y Y Y Y
Natural Amenity Controls Y Y Y Y Y Y
River Buffer Zones < 10 km < 10 km < 10 km < 10 km < 10 km < 10 km
Grid Sample Urban grids Within 20% cut-off border
Adjusted R2 0.228 0.228 0.228 0.227 0.227 0.227
N 22,239 22,239 22,239 11,269 11,269 11,269
Note: This table reports the OLS estimates where the dependent variable is the share of amenity land
within 1km × 1km grids. The explanatory variable of interest is the Flood Exposure which equal to 1 if
the cell is in floodplain area and 0 otherwise. We, then, further use the flood annual frequencies (which
is the inverse of flood return periods). Column (1) to (3) includes all the grids with urban status equal
to 1. The cells that have urban status equal to 1 are those with at least 20% of artificial land use within
the grid. Column (4) to (6), we consider all cells within a circular urban boundary. Urban boundaries are
defined as the circular cut-off at which the total urban fabric (or residential land use) are of at least 20% of
total land within its rings. For more details, see Picard and Tran (2019). Other controls include the coastal
grid indicator. Significance levels are denoted by ∗ for p<0.1, ∗∗ for p<0.05 and ∗∗∗ for p<0.01. Standard
errors are clustered at city level and reported in parentheses.

We follow the same strategies and regressions as with flood prevention infrastruc-
ture in the previous section. In particular, we run the following regressions:
AmenityLandirc = β f FloodDirc + αc Xc + γirc + eirc (5.47)
4
AmenityLandirc = FloodDi × ∑ ( β f k DistQuartilekc ) + αc Xc + γirc + eirc (5.48)
k =1
160 CHAPTER 5. URBAN FLOOD, AMENITY LAND AND CITY STRUCTURE

where AmenityLandirc is the share of urban green space and sport and leisure facility
land uses for each grid in our samples. We also consider different quartile levels as in
the previous section.
Table 5.9 shows the results. Columns (1) and (4) show a significant and positive co-
efficient for Flood Exposure with amenity land use. However, the magnitude is small. A
grid located in a floodplain area induces a 0.3% to 0.6% increase in the amenity land. A
similar regression with flood frequency level also shows a significant and positive sign.
Specifically, one additional flood event every 10 years would increase the provision of
amenity land in the same grid by 0.3% to 0.62% relative to unexposed grids. Surpris-
ingly, the interaction between Flood Exposure and the distance quartiles do not result in
a significant coefficient except for the first and second quartiles.

Figure 6: Amenity land uses and IDL exposure

This figure demonstrates the relationship between the change in amenity land uses relative to the unexposed location
and the IDL exposure categories. These coefficients are based on the regression of the stock of the flood prevention
infrastructures and five different categories of IDL from Very Low level (0-1m) and Very High level (above 6m).

Figure 6 shows the effect of different inundation levels on the provision of amenity
land. We obtain the coefficients by rerunning regression (5.47) but replace the exposure
indicators with IDL indicators. As shown in Figure 6, there is a mostly linear relation-
ship between IDL and the provision of amenity land for both urban grids and those
within a circular urban border. However, there is a switch in sign from negative to pos-
itive from very low IDL below 1 m and IDL above 1 m. This pattern could be related
to the positive coefficient in population density for very low IDL, as shown in Figure 4.
5.4. CONCLUSION 161

The gradual increase in amenity land with respect to IDL suggests that the cost reduc-
tion effect dominates in equation (5.28) at a higher IDL. Only at very low IDL (below 1
m) does less housing construction dominate, leading to less amenity land at very low
IDL.

5.4 Conclusion
To what extent do city structure and land use planning reflect climatic risks, specif-
ically urban flood risks? This issue is of growing interest due not only to its policy
importance (Anderson et al., 2018) but also to its relation to the central question re-
garding the increasing vulnerability of urban zones. Flooding has long been one of the
costliest natural disasters worldwide, and risks are set to increase with the effects of
climate change and rising sea levels over the coming decades.
In this paper, we study the effect of flood risks on urban structures and optimal
land uses. In particular, we examine theoretically (i) whether population density, hous-
ing demand, and housing constructions concentrate disproportionately in flood-prone
urban areas; (ii) whether higher risks lead to higher levels of private and public flood
protection efforts; and (iii) whether the amenity provision is changed due to flood risk
within the urban area. Our model anticipates a lower population density and a sparser
residential construction area in a hazardous area. It also predicts a nonmonotonic re-
lationship between the level of flood risks and the optimal provision of public flood
defenses and amenity land. In the second-best case, the city government collects taxes
to cover the expenses of flood defense systems as well as the provision of amenity land.
It faces budget balance constraints. We find that a locational tax is efficient while an
income tax system would result in over-concentration of population and housing con-
struction in a flood-prone area.
The empirical studies across European cities confirm the persistent and negative
impacts of flood risk on urban population density and residential land covers, which
confirms the prediction from our theoretical model. Second, the provision of amenity
lands is nonmonotonic with the level of flood risks and depends on the relative location
within the urban area. Specifically, the fraction of amenity land is more likely to be built
in a hazardous and compact area than in a risky but sparse area near the urban fringe.
These results further contribute to our understanding of the effects of river flood risk
on land use planning policies.
Cities are enormous and intricate economic engines, but ultimately, they are crea-
tures of human free will and respond to people’s desires for a livable environment.
People can optimize the use of land within urban locations to help mitigate the draw-
backs and embellish the strengths. We believe that the environments that we have lived
in are those that we created.
162 APPENDIX

Appendix A: Second-best world with locational tax system

Here, we assume that the burden of the locational tax τ (r, φ) is on the land develop-
ers.18 We first study the housing demand from households’ side, then move to the land
developers’ problem once the bid-rent function is determined, and the decisions of the
city government over the provision of public flood mitigation and amenity land.

Housing demand

The household chooses z, h, e. During a unique stage, the household finds z, s, e that
Z1  
max U (z, h, a) s.t. z + t + Rh + h l ( x )(1 − m(e, p)) + C H (e) dF ( x; r, φ) = w.
z,h,e
0
(5.49)
One can interpret this as the consumer considering her permanent income (she can
borrow to smooth her consumption).
R1
Denote again l = l (r, φ) ≡ l ( x )dF ( x; r, φ). We will drop the location indicator
0
(r, φ) where there is no confusion in the notation. Each household will bid for the
location of their preference, and the slot will be given to the highest bidder. Hence, we
can write the bid rent function as follows The household program is given by equation
(5.49). The bid rent function is given by
w−t−z
 
Υ = max H
− l (1 − m(e, p)) − C (e) s.t. U (z, h, a) = u (5.50)
z,h,e h
We have utility function U is strictly increasing in z; therefore, we can write ẑ(u, h, a)
as the unique solution of equation U (z, h, a) = u, at each level of housing consump-
tion and amenity level of indifferent curve. For general form of utility, we have zh =
−Uh /Uz < 0, z a = −Ua /Uz < 0 as utility is increasing function in all z, h, a; therefore,
the consumption of composite good along the indifferent curve is decreasing with the
housing floor service h and amenity level a. Then, the bid rent function can be defined
as follows:
w − t − ẑ(u, h, a)
 
Υ = max H
− l (1 − m(e, p)) − C (e) (5.51)
h,e h
The first-order conditions give

ẑ − hẑh = y (5.52)
l × me (e, p) = CeH (e) (5.53)

where y = w − t is the disposable income for each household. We obtain h(y, a, u),
z(y, a, u) and e( p, l ), which gives the equilibrium housing floor value
y−z
R = V (y, a, u, p, l ) ≡ − l (1 − m(e, p)) − C H (e). (5.54)
h
18 It is similar in the case of locational tax for households.
APPENDIX 163

Housing supply

On the supply side, we assume the land is owned by absentee landlords, who have
two options to either lend the piece of land for land developers to construct residential
houses or to the agriculture sectors at R A . At each location, a land developer pays a
land price of Ψ to the landlords and a locational tax τ to build q units of housing floors
that can be rented out for households at R.19 The construction cost is κqα with α > 1 as
in the previous section. The residential developers, then, choose the quantity housing
floor q to maximize the total profit
max Π = Rq − κqα − Ψ − τ
q

Taking the first-order condition w.r.t q, we have

  α−1 1
R
q=
ακ
which is at the same level as in the first best case.
Additionally, the free entry condition dictates
 α−1 1
α−1

R
Π = 0 ⇐⇒ Ψ = Rq − κq − τ = Rα
− τ.
α ακ
The rent is at least equal to or higher than the outside opportunity cost of land which is
the agricultural land rent R A . Therefore, we can denote the actual land rent at location
(r, φ) is as follows:
R(Lr,φ) = max{ R A ; Ψ(r,φ) }
Therefore, a location will be developed if and only if
  α−α 1
α 1
Ψ(r,φ) ≥ R A ⇐⇒ R ≥ R( R A , τ ) ≡ (R A + τ) (ακ ) α (5.55)
α−1
From equation (5.54), we can see that the flood risk level reduces the bid rent of house-
holds, leading to lower house rent Rl < 0. If the effect of the flood risk is strong enough
to reduce the bid rent of household R at some location to be lower than the threshold
R, then this location will be left undeveloped. This corner equilibrium can lead to the
discontinuity in land use within the urban area. Specifically,
y − z(y, a, u)
R ≥ R( R A , τ ) ⇐⇒ L( p, l ) ≡ l (1 − m(e, p)) + C H (e) ≤ − R( R A , τ )
| {z } h(y, a, u)
Loss due to flood hazard
y−z
− R( R A , τ ) − C H (e)
⇐⇒ l ≤ l ( p, R A , τ ) where l ( p, R A , τ ) ≡ h
(1 − m(e, p))
Any location with average financial loss of flood risk l higher than l is not developed.
One notices that l is decreasing with distance to the city center, R A and the locational
tax τ, meaning the restriction is stricter near urban fringe and with higher outside op-
portunity cost of land, all else being equal.
19 As it is clear that the tax and land price are at each location (r, φ), so we drop the notation.
164 APPENDIX

City government

The city government has three choices over τ (r, φ), p(r, φ) and g(r, φ), and she chooses
to minimize the total social cost while maintaining the same level of utility for every city
inhabitant. We drop the location notation where it is not caused misunderstanding. The
objective of the urban planner, now, can be written as follows.

Zb Z2π 
ES = max [V (y, a, u, p, l )q − κqα − CU ( p)](1 − g) − R A − cg `dφdr (5.56)
p,g
0 0

such that
U (z, h, a) ≥ u.

and balanced budget at each location

τ = CU ( p)(1 − g) + cg (5.57)

As the budget constraint is always binding at each location, and it is just an accounting
problem, we can solve the city government problem first with the choice of public flood
prevention p and amenity g. Taking the first order conditions of equation (5.56) w.r.t. p
and g, we get

Vp × q = CU
p (5.58)
θqVa (1 − g) = Vq − κqα − CU ( p) + c (5.59)

which are identical to the first-best conditions. One also notice that in safe area, we have
Vp = l × m p = 0, leading to zero public flood protection effort p; hence, the locational
tax in safe zone only cover the cost for amenity provision τsa f e = cg. Under locational
taxation rule (5.57) and condition (5.55) binding for all locations within the urban area,
the outcomes of the first-best optimization are decentralized through market equilib-
rium through locational taxation.
APPENDIX 165

Appendix B: Flood risk and river networks

Figure B1: Flood Risk, River Networks and Land Use in Paris
(a) River Networks with Flood Hazard

(b) Continuous Urban Fabric, Urban Boundary and River Buffer Zones
166 APPENDIX

Figure B2: Paris - Flood Prevention Infrastructures

Paris - Ouvrage de Protection

Ouvrage de protection

1 : 500 000 © IGN, © TELEATLAS, © BRGM

Appendix C: Robustness checks

Figure C1: Inundation Depth Level and population density (5 km buffer)

This figure denotes the relationship between the % change in population density of the grids within the zones with
flood exposure (in comparison with unexposed grids), partitioned by different levels of inundation depth level. We
categorize the flood hazard according to the exposure level from Very Low to Very High, which are defined as the
average Inundation Depth Level (IDL) in all return periods. We set the IDL between 0 and 1 meter as very Low level
and above 6 meters in Very High level.
 Table C1: Population density and flood hazard - Robustness check

Dependent variable: Ln(Pop)

Panel A: (1) (2) (3) (4) (5) (6) (7) (8)
Flood Exposure 0.077∗∗ −0.068∗ −0.127∗∗∗ −0.101∗∗∗ −0.091∗∗∗ −0.205∗∗∗ −0.055 −0.041
(0.032) (0.038) (0.041) (0.033) (0.035) (0.049) (0.037) (0.036)
APPENDIX

Dist. to CBD Y Y Y Y Y Y Y Y
City FE Y Y Y Y Y Y Y Y
Geographical controls N Y Y Y Y Y Y Y
Dist. to river N N Y Y Y Y Y Y
Natural Amenity Controls N N N Y Y N Y Y
Artificial Amenity Controls N N N N Y N N Y
River Buffer Zones < 5 km < 5 km < 5 km < 5 km < 5 km < 5 km < 5 km < 5 km
Grid Sample Urban grids Urban grids Urban grids Urban grids Urban grids Within 20% Within 20% Within 20%
cut-off border cut-off border cut-off border
Adjusted R2 0.399 0.421 0.424 0.612 0.638 0.457 0.701 0.726
N 15,014 15,014 15,014 15,014 15,014 7,964 7,964 7,964
Panel B: (1) (2) (3) (4) (5) (6) (7) (8)
Flood Frequencies 0.229 −0.550∗∗∗ −0.929∗∗∗ −0.745∗∗∗ −0.639∗∗∗ −1.680∗∗∗ −0.702∗∗∗ −0.565∗∗∗
(0.185) (0.211) (0.230) (0.183) (0.191) (0.326) (0.199) (0.193)
Dist. to CBD Y Y Y Y Y Y Y Y
City FE Y Y Y Y Y Y Y Y
Geographical controls N Y Y Y Y Y Y Y
Dist. to river N N Y Y Y Y Y Y
Natural Amenity Controls N N N Y Y N Y Y
Artificial Amenity Controls N N N N Y N N Y
River Buffer Zones < 5 km < 5 km < 5 km < 5 km < 5 km < 5 km < 5 km < 5 km
Grid Sample Urban grids Urban grids Urban grids Urban grids Urban grids Within 20% Within 20% Within 20%
cut-off border cut-off border cut-off border
Adjusted R2 0.399 0.422 0.425 0.612 0.639 0.461 0.702 0.727
N 15,014 15,014 15,014 15,014 15,014 7,964 7,964 7,964
Note: This table reports the OLS estimates where the dependent variable is the natural log of population count in 1k × 1km grid. The explanatory variable of interest
is the Flood Exposure which equal to 1 if the cell is in floodplain area and 0 otherwise. We, then, further use the flood annual frequencies (which is the inverse of
flood return periods). Column (1) to (5) includes all the grids with urban status equal to 1. The cells that have urban status equal to 1 are those with at least 20% of
artificial land use within the grid. Column (6) to (8), we consider all cells within a circular urban boundary. Urban boundaries are defined as the circular cut-off at
which the total urban fabric (or residential land use) are of at least 20% of total land within its rings. For more details, see Picard and Tran (2019). Other controls
include the coastal grid indicator. Significance levels are denoted by ∗ for p<0.1, ∗∗ for p<0.05 and ∗∗∗ for p<0.01. Standard errors are clustered at city level and
167

reported in parentheses.
APPENDIX

Table 5.10: Table C2: Housing construction and flood hazard - Robustness check
Panel A: Dependent variable: Building Height
(1) (2) (3) (4) (5) (6) (7) (8)
Flood Exposure −0.175 −0.517∗∗∗ −0.586∗∗∗ −0.524∗∗∗ −0.464∗∗∗ −0.552∗∗∗ −0.496∗∗∗ −0.434∗∗∗
(0.144) (0.161) (0.147) (0.118) (0.119) (0.116) (0.129) (0.109)
Adjusted R2 0.422 0.458 0.467 0.522 0.600 0.485 0.540 0.626
N 2,903 2,903 2,903 2,903 2,903 2,375 2,375 2,375
Panel B: Dependent variable: Share of Continuous Residential Categories (S.L > 80%)
(1) (2) (3) (4) (5) (6) (7) (8)
Flood Exposure −0.006 −0.015∗∗∗ −0.025∗∗∗ −0.023∗∗∗ −0.021∗∗∗ −0.036∗∗∗ −0.029∗∗∗ −0.024∗∗∗
(0.005) (0.005) (0.005) (0.004) (0.004) (0.007) (0.006) (0.006)
Adjusted R2 0.228 0.238 0.246 0.352 0.429 0.386 0.425 0.516
N 15,093 15,093 15,093 15,093 15,093 8,026 8,026 8,026
Dist. to CBD Y Y Y Y Y Y Y Y
City FE Y Y Y Y Y Y Y Y
Geographical controls N Y Y Y Y Y Y Y
Dist. to river N N Y Y Y Y Y Y
Natural Amenity Controls N N N Y Y N Y Y
Artificial Amenity Controls N N N N Y N N Y
River Buffer Zones < 5 km < 5 km < 5 km < 5 km < 5 km < 5 km < 5 km < 5 km
Grid Sample Urban grids Urban grids Urban grids Urban grids Urban grids Within 20% Within 20% Within 20%
cut-off border cut-off border cut-off border
Note: This table reports the OLS estimates where the dependent variable is the natural log of population count in 1km × 1km grid. The explanatory variable of
interest is the Flood Exposure which equal to 1 if the cell is in floodplain area and 0 otherwise. We, then, further use the flood annual frequencies (which is the
inverse of flood return periods). Column (1) to (5) includes all the grids with urban status equal to 1. The cells that have urban status equal to 1 are those with at
least 20% of artificial land use within the grid. Column (6) to (8), we consider all cells within a circular urban boundary. Urban boundaries are defined as the circular
cut-off at which the total urban fabric (or residential land use) are of at least 20% of total land within its rings. For more details, see Picard and Tran (2019). Other
controls include the coastal grid indicator. Significance levels are denoted by ∗ for p<0.1, ∗∗ for p<0.05 and ∗∗∗ for p<0.01. Standard errors are clustered at city level
and reported in parentheses.
168
APPENDIX 169

Table C3: Flood prevention infrastructure and flood hazards - Robustness check

Dependent variable: Flood Prevention Infrastructures Level (km)

(1) (2) (3) (4) (5) (6)
Flood Exposure 0.082∗∗∗ 0.122∗∗∗
(0.018) (0.031)
Flood Frequency 0.542∗∗∗ 0.856∗∗∗
(0.114) (0.175)
Flood Exp. × Dist. to CBD 0.155∗∗∗ 0.214∗
(1st quartile) (0.038) (0.125)
Flood Exp. × Dist. to CBD 0.013 0.137∗∗∗
(2nd quartile) (0.019) (0.050)
Flood Exp. × Dist. to CBD 0.022 0.119∗∗
(3rd quartile) (0.023) (0.048)
Flood Exp. × Dist. to CBD −0.006 0.076∗
(4th quartile) (0.021) (0.041)
Dist. to CBD Y Y Y Y Y Y
City FE Y Y Y Y Y Y
Geographical controls Y Y Y Y Y Y
Dist. to river Y Y Y Y Y Y
Natural Amenity Controls Y Y Y Y Y Y
Artificial Amenity Controls Y Y Y Y Y Y
River Buffer Zones < 5 km < 5 km < 5 km < 5 km < 5 km < 5 km
Grid Sample Urban grids Within 20% cut-off border
Adjusted R2 0.159 0.164 0.176 0.180 0.190 0.183
N 4,390 4,390 4,390 2,636 2,636 2,636
Note: This table reports the OLS estimates where the dependent variable is the flood prevention infras-
tructure within 1km × 1km grids. The explanatory variable of interest is the Flood Exposure which equal
to 1 if the cell is in floodplain area and 0 otherwise. We, then, further use the flood annual frequencies
(which is the inverse of flood return periods). Column (1) to (3) includes all the grids with urban sta-
tus equal to 1. The cells that have urban status equal to 1 are those with at least 20% of artificial land
use within the grid. Column (4) to (6), we consider all cells within a circular urban boundary. Urban
boundaries are defined as the circular cut-off at which the total urban fabric (or residential land use) are
of at least 20% of total land within its rings. For more details, see Picard and Tran (2019). Other controls
include the coastal grid indicator. Significance levels are denoted by ∗ for p<0.1, ∗∗ for p<0.05 and ∗∗∗
for p<0.01. Standard errors are clustered at city level and reported in parentheses.
170 APPENDIX

Figure C2: Flood prevention infrastructures and IDL exposure (5km buffer)

This figure demonstrates the relationship between the change in flood prevention infrastructure (measured as kilome-
ters of “Ouvrage de protection”) relative to the unexposed location and the IDL exposure categories. These coefficients
are based on the regression of the stock of the flood prevention infrastructures and five different categories of IDL from
Very Low level (0-1m) and Very High level (above 6m).

Figure C3: Amenity land uses and IDL exposure (5km buffer)

This figure demonstrates the relationship between the change in amenity land uses relative to the unexposed location
and the IDL exposure categories. These coefficients are based on the regression of the stock of the flood prevention
infrastructures and five different categories of IDL from Very Low level (0-1m) and Very High level (above 6m).
APPENDIX 171

Table C4: Amenity land uses and flood hazards - Robustness check

Dependent variable: Amenity Land Use (%)

(1) (2) (3) (4) (5) (6)
Flood Exposure 0.006∗∗ 0.009∗
(0.003) (0.005)
Flood Frequency 0.051∗∗∗ 0.086∗∗∗
(0.014) (0.025)
Flood Exp. × Dist. to CBD 0.009∗∗ −0.003
(1st quartile) (0.003) (0.009)
Flood Exp. × Dist. to CBD −0.0003 0.021∗∗∗
(2nd quartile) (0.003) (0.006)
Flood Exp. × Dist. to CBD 0.005 0.006
(3rd quartile) (0.005) (0.007)
Flood Exp. × Dist. to CBD 0.006 0.006
(4th quartile) (0.011) (0.006)
Dist. to CBD Y Y Y Y Y Y
Dist. to CBD (sq) Y Y Y Y Y Y
City FE Y Y Y Y Y Y
Geographical controls Y Y Y Y Y Y
Dist. to river Y Y Y Y Y Y
Natural Amenity Controls Y Y Y Y Y Y
River Buffer Zones < 5 km < 5 km < 5 km < 5 km < 5 km < 10 km
Grid Sample Urban grids Within 20% cut-off border
Adjusted R2 0.227 0.228 0.227 0.215 0.216 0.216
N 15,093 15,093 15,093 8,026 8,026 8,026
Note: This table reports the OLS estimates where the dependent variable is the flood prevention infras-
tructure within 1km × 1km grids. The explanatory variable of interest is the Flood Exposure which equal
to 1 if the cell is in floodplain area and 0 otherwise. We, then, further use the flood annual frequencies
(which is the inverse of flood return periods). Column (1) to (3) includes all the grids with urban sta-
tus equal to 1. The cells that have urban status equal to 1 are those with at least 20% of artificial land
use within the grid. Column (4) to (6), we consider all cells within a circular urban boundary. Urban
boundaries are defined as the circular cut-off at which the total urban fabric (or residential land use) are
of at least 20% of total land within its rings. For more details, see Picard and Tran (2019). Other controls
include the coastal grid indicator. Significance levels are denoted by ∗ for p<0.1, ∗∗ for p<0.05 and ∗∗∗
for p<0.01. Standard errors are clustered at city level and reported in parentheses.
Bibliography

[1] Alfieri, L., Salamon, P., Bianchi, A., Neal, J., Bates, P.D., Feyen, L. (2014). Ad-
vances in pan-European flood hazard mapping, Hydrol. Process., 28 (18), 4928-4937,
doi:10.1002/hyp.9947

[2] Alonso W. (1964). Location and Land Use. Harvard Univ. Press, Cambridge, MA,
1964.

[3] Anderson S.E., T.L. Anderson, A.C. Hill, M.E. Kahn, H. Kunreuther, G.D. Libecap,
H. Mantripragada, P. Merel, A. Plantinga, and V.K. Smith (2018). The Critical Role
of Markets in Climate Change Adaptation. NBER Working Paper No. w24645.

[4] Anderson S.T. and S.E. West (2006). Open Space, residential property values, and
spatial context. Regional Science and Urban Economics 36, 773-789.

[5] Arnott, R. J., and Stiglitz, J. E. (1979). Aggregate land rents, expenditure on public
goods, and optimal city size. The Quarterly Journal of Economics, 93(4), 471-500.

[6] Atreya A., Ferreira S., and Kriesel (2013). Forgetting the Flood? An Analysis of
the Flood Risk Discount over Time. Land Economics, Vol 89, No.4, Nov 2013, pp.
577-596.

[7] Bairoch P., Batou J. and Chèvre P. (1988). La Population des villes européennes de
800 à 1850 : banque de données et analyse sommaire des résultats. Publications
d’histoire économique et sociale internationale. Ed. Droz. France.

[8] Bakkensen L. A., and Barrage L. (2018). Flood Risk Belief Heterogeneity and Cotal
Home Price Dynamics: Going Under Water? NBER Working Paper 23854.

[9] Barthel S., C. Folke, J. Colding (2010). Social–ecological memory in urban gardens:
retaining the capacity for management of ecosystem services. Global Environmental
Change 20, 255-265.

[10] Bayer B. and Timmins C. (2005). Estimating equilibrium models of sorting across
locations. Economic Journal 117(3). 353-374.

173
174 BIBLIOGRAPHY

[11] Berman M.G., J. Jonides, and S. Kaplan (2008). The Cognitive Benefits of
Interacting With Nature. Psychological Science 19, 1207. DOI: 10.1111/j.1467-
9280.2008.02225.x

[12] Bernstein A., Gustafson M. T., and Lewis R. (2018). Disaster on the horizon: the
price effect of sea level rise. Journal of Financial Economics, forthcoming.

[13] Bolitzer B. and N.R. Netusil (2000). The impact of open spaces on property values
in Portland, Oregon. Journal of Environmental Management 59, 185-193.

[14] Bowler D.E., L. Buying-Ali, T.M. Knight, and A.S. Pullin (2010). Urban Greening to
cool towns and cities: a systematic review of the empirical evidences. Landscapes
an Urban Planning 97, 147-155.

[15] Brander L.M. and M.J Koetse (2011). The value of urban open space: Meta-analyses
of contingent valuation and hedonic pricing results. Journal of Environmental Man-
agement 92(10), 2763-2773.

[16] Brack, C. L. (2002). Pollution mitigation and carbon sequestration by an urban

forest. Environmental Pollution, 116, 195–200.

[17] Brakenridge, G.R. Global Active Archive of Large Flood Events, Dartmouth Flood
Observatory, University of Colorado.

[18] Brueckner J. K. (1983). The Economics of Urban Yard Space: An ‘Implicit-Market’

Model for Housing Attributes. Journal of Urban Economics 13, 216-234 (1983).

[19] Brueckner J., J-F. Thisse, and Y. Zenou (1999). Why is central Paris rich and down-
town Detroit poor? An amenity-based theory. European Economic Review 43, 91-107.

[20] Capozza, D. R., & Helsley, R. W. (1990). The stochastic city. Journal of urban Eco-
nomics, 28(2), 187-203.

[21] Caruso G., D. Peeters, J. Cavailhes, and M. Rounsevell (2007). Spatial configura-
tion in a periurban city. A cellular automata-based microeconomic model. Regional
Science and Urban Economics 37(5), 542-567.

[22] Cheshire P. and S. Sheppard (2002). The welfare economics of land use planning.
Journal of Urban Economics 52, 242-269.

[23] Cho S., J. Bowker, and W. Park (2006). Measuring the contribution of water and
green space amenity to housing values: An application and comparison of spa-
tially weighted hedonic models. Journal of Agricultural and Resource Economics 31(3),
485-507.
BIBLIOGRAPHY 175

[24] Cohen D.A., T.L. MacKenzie, A. Sehgal, S. Williamson, D. Golinelli, and N. Lurie
(2007). Contribution of Public Parks to Physical Activity. American Journal of Public
Health, 97(3), pp. 509-514.

[25] Cohen D.A., B. Han, C.J. Nagel, P. Harnik, T.L. MacKenzie, K.R. Evenson, T. Marsh,
S. Williamson, C. Vaughan, and S. Katta (2016). The First National Study of Neigh-
borhood Parks: Implications for Physical Activity. American Journal of Preventive
Medicine, 54(4), pp. 419-426.

[26] Colding J. and S. Barthel (2013). The potential of ‘Urban Green Commons’ in the
resilience building of cities. Ecological Economics 86, 156-166.

[27] Cremer H., A. De Kerchove, and J-F. Thisse (1985). An economic theory of public
facilities in space. Mathematical Social Sciences 9, 249-262.

[28] Cribari-Neto F. and Da Silva W.B. (2011). A New Heteroskedasticity-Consistent

Covariance Matrix Estimator for the Linear Regression Model. Advances in Statis-
tical Analysis, 95(2), 129-146.

[29] Davis, M.A, and Ortalo-Magne, F. (2011). Household Expenditures, wages, rents.
Review of Economic Dynamics 14 (2), 248-261.

[30] Dottori, F., Salamon, P., Bianchi, A., Alfieri, L., Hirpa, F.A., Feyen, L. (2016a). De-
velopment and evaluation of a framework for global flood hazard mapping. Ad-
vances in Water Resources 94, 87-102.

[31] Dottori, Francesco; Alfieri, Lorenzo; Salamon, Peter; Bianchi, Alessandra; Feyen,
Luc; Lorini, Valerio (2016b): Flood hazard map for Europe - 100-year return period.
European Commission, Joint Research Centre (JRC).

[32] E-OBS database (2017 release version). We acknowledge the E-OBS dataset from
the EU-FP6 project ENSEMBLES (http://ensembles-eu.metoffice.com) and the
data providers in the ECA&D project (http://www.ecad.eu).

[33] Ellison G., E.L. Glaeser, and W.R. Kerr (2010). What Causes Industry Agglomera-
tion: Does Natural Advantage Explain Agglomeration? American Economic Review
100, 1195-1213.

[34] Elsasser P. (1996) Der Erholungwert des Waldes: Monetare Bewertung der Er-
holungsleistung ausgewählter Wälder in Deutchland (The recreational value of
forests: Monetary valuation of the recreational performance of selected forests in
Germany). Schriften zur Forstökonomie Bd. 11. Sauerländer, Frankfurt am Main,
(in German).
176 BIBLIOGRAPHY

[35] EEA (2009). WISE Large River and Large Lakes. Water Pattern Europe, scale 10
million, version 2, from EUROSTAT GISCO database; Water Framework Directive
article 3 data on rivers and lakes from countries; Joint Research Centre catchment
database CCM1.

[36] Epple D. and H. Sieg (1999). Estimating equilibrium models of local jurisdictions.
Journal of Political Economy 107 (4), 645-681.

[37] EU DG-ENVI (2012). The multi-functionality of green infrastructure. Science for

Environmental Policy, In-depth Reports.

[38] Eurostat (2016). Urban Europe: Statistics on cities, towns and suburbs. 2016 Edi-
tion. ISBN 978-92-79-60139-2.

[39] Evenson K.R., F. Wen, D. Golinelli, D.A. Rodriguez, and D.A. Cohen (2013). Mea-
surement Properties of a Park Use Questionnaire. Environmental and Behavior,
45(4), pp. 526-547.

[40] Fahey, T., Nolan, B., Maitre, B. (2004) Housing expenditures and income poverty
in EU countries. Journal of Social Policy 33 (3), 437-454.

[41] Fallick B., C.A. Fleischmann, and J.B. Rebitzer (2006). Job-Hopping in Silicon Val-
ley: Some Evidence Concerning the Microfoundations of a High-Technology Clus-
ter. Review of Economics and Statistics 88(3), 472-481.

[42] Frame, D.E. (1998). Housing, Natural Hazards, and Insurance. Journal of Urban
Economics 44, 93-109 (1998).

[43] Freeman A.M., J.A. Herriges, and C.L. Kling (2014). The measurement of Environ-
mental and Resource Values: Theory and Method. Third Edition. Resources for the
Future - RFF Press, Routledge, NY.

[44] Fujita M. (1986). Optimal Location of public facilities - Area Dominance Approach.
Regional Science and Urban Economics 16, 241-268.

[45] Fujita M. (1989). Urban Economic Theory: Land Use and City Size. Cambridge
Univ. Press.

[46] Fujita M. and J-F. Thisse (2002). Economics of Agglomeration: Cities, Industrial
Location, and Regional Growth. Cambridge Univ. Press.

[47] Gallagher, Justin (2014). Learning about an Infrequent Event: Evidence from Flood
Insurance Take-Up in the United States. American Economic Journal: Applied Eco-
nomics 2014, 6(3): 206-233.
BIBLIOGRAPHY 177

[48] Geoghegan J., L.A. Waigner, and N.E. Bockstael (1997). Spatial landscape indices
in a hedonic framework: an ecological economics using GIS. Ecological Economics
23, 251-264.

[49] Glaeser E.L., J. Kolko, and A. Saiz (2001). Consumer City. Journal of Economic Geog-
raphy 1, 27-50.

[50] Glaeser E.L., and J. Gottlieb (2009). The Wealth of Cities: Agglomeration
Economies and Spatial Equilibrium in the United States. Journal of Economic Lit-
erature 47(4), 983-1028.

[51] Glaeser E.L. (2011). Cities, Productivity and the Quality of Life. Science 333: 592-
594

[52] Glaeser E.L. (2012). The Triumph of the City: How Our Greatest Invention Makes
Us Richer, Smarter, Greener, Healthier, and Happier. New York: The Penguin Press.

[53] Haylock, M.R., N. Hofstra, A.M.G. Klein Tank, E.J. Klok, P.D. Jones, M. New. 2008:
A European daily high-resolution gridded dataset of surface temperature and pre-
cipitation. J. Geophys. Res (Atmospheres), 113, D20119, doi:10.1029/2008JD10201

[54] Hedblom M., B. Gunnarsson, B. Iravani, et al. (2019). Reduction of physiological

stress by urban green space in a multisensory virtual experiment. Nature, Scientific
Reports 9, 10113. https://doi.org/10.1038/s41598-019-46099-7

[55] Heidt V. and M. Neef (2008). Benefits of Urban Green Space for Improving Urban
Climate. In: Carreiro M.M., Y.C. Song, J. Wu (eds) Ecology, Planning, and Manage-
ment of Urban Forests, Springer, New York, NY.

[56] IFPRA (2013). Benefits of Urban Parks - A systematic review. Copenhagen, Inter-
national Federation of Parks and Recreation Administration. 68p.

[57] Indaco Agustı́n, Francesc Ortega, and Süleyman Taspmar (2019). The Effects of
Flood Insurance on Housing Markets. Working paper 2019. [Older version: IZA
DP No. 11810].

[58] Kabisch N. and D. Haase (2014). Green justice or just green? Provision of green
urban spaces in Berlin, Germany. Landscape and Urban Planning 122 (0), 129-139.

[59] Klaiber H. A. and D.J. Phaneuf (2010). Valuing open space in a residential sorting
model of the Twin Cities. Journal of Environmental Economics and Management 60,
57-77.

[60] Kocornik-Mina A., McDermott, G. Michaels, F. Rauch (2019). Flooded Cities. Amer-
ican Economic Journal: Applied Economics, forthcoming.
178 BIBLIOGRAPHY

[61] Kuminoff N.V., V:K: Smith, and C. Timmins (2013). The New Economics of Equi-
librium Sorting and Policy Evaluation Using Housing Markets. Journal of Economic
Literature 51(4), 1007-1062.

[62] Lee C. and M. Fujita (1997). Efficient configuration of a greenbelt: Theoretical mod-
elling of greenbelt amenity. Environment and Planning A 29, 1999-2017.

[63] Lucas R.E. and E. Rossi-Hansberg (2002). On the Internal Structure of Cities. Econo-
metrica 70, 1445-1476.

[64] Luechinger S. and Raschky P. A. (2009). Valuing flood disaster using the life satis-
faction approach. Journal of Public Economics 93 (2009) 620-633.

[65] Luttik, J. (2000). The value of trees, water and open space as reflected by house
prices in the Netherlands. Landscape and Urban Planning 48, 161-167.

[66] Lutzenhiser M. and N.R. Netusil (2001). The effect of open spaces on a home’s sale
price. Contemporary Economic Policy 19, 291-298.

[67] Manes F., G. Incerti, E. Salvatori, M. Vitale, M. Ricotta, and R. Constanza (2012).
Urban ecosystem services: tree diversity and stability of tropospheric ozone re-
moval. Ecological Application 22, 349-369.

[68] McFadden D. (1978). Modeling the choice of residential location. in: Krlqvist A., L.
Lundqvist, and F. Snickars (Eds.) Spatial Interaction Theory and Planning Models.
Amsterdam, North Holland, 1978.

[69] Mossay P. and P.M. Picard (2011). On spatial equilibria in a social interaction
model. Journal of Economic Theory, 146(6), 2455-2477.

[70] OECD (2013). Definition of Functional Urban Areas (FUA) for the
OECD metropolitan database. Link: https://www.oecd.org/cfe/regional-
policy/Definition-of-Functional-Urban-Areas-for-the-OECD-metropolitan-
database.pdf (acessed on Mar 01, 2018).

[71] Orford S. (2002). Valuing locational externalities: A GIS and multilevel modelling
approach. Environment and Planning B: Planning and Design 29(1), 105-127.

[72] Ortega Francesc and Suleyman Taspinar (2018). Rising Sea Levels and Sinking
Property Values: The Effects of Hurricane Sandy on New York’s Housing Market.
Journal of Urban Economics 106 (2018) 81-100.

[73] Owens, R., Rossi-Hansberg, E., Sarte, P. D. (2017). Rethinking Detroit. Federal Re-
serve Bank of Richmond Working Papers (2017), Vol. 17, Issues 04, 1-45.

[74] Picard P.M. and Tabuchi T. (2013). On microfoundations of the city. Journal of Eco-
nomic Theory 148(6), 2561-2582.
BIBLIOGRAPHY 179

[75] Picard P.M., and Tran, T.T.H (2019) Green Urban Spaces. CREA Discussion Papers
2019.

[76] Purdy J. (2015). After Nature - A politics for the Anthropocene. Harvard Univ.
Press.

[77] Rappaport J. (2008). Consumption Amenities and City Population Density. Re-
gional Science and Urban Economics 38(6), 533-552.

[78] Sakashita N. (1987). Optimum Location of Public Facilities under the influence of
the land market. Journal of Regional Science 27.

[79] Saldivar-Tanaka L. and M. Krasny (2004). Culturing community development,

neighborhood open space, and civic agriculture: the case of Latino community
gardens in New York City. Agriculture and Human Values 21, 399-412.

[80] Schindler M., Le Texier M., and Caruso G. (2018). Spatial Sorting, attitudes and the
use of green space in Brussels. Urban Forestry & Urban Greening 31, 169-184.

[81] Sieg H., V.K. Smith, H.S. Banzhaf, and R. Walsh (2004). Estimating the general
equilibrium benefits of large changes in spatially delineated public goods. Interna-
tional Economic Review 45(4), 1047-1077.

[82] Smith S.K., C. Poulus, and H. Kim (2002). Treating open space as an urban amenity.
Resource and Energy Economics 24, 107-129.

[83] Strohbach M.W., E. Arnold and D. Haase (2012). The carbon footprint of urban
green space - A life cycle approach. Landscape and Urban Planning 104, 220-229.

[84] Tiebout C. M. (1956). A Pure Theory of Local Expenditures. The Journal of Political
Economy 64(5), 416-424.

[85] Thisse J-F. and D. Wildasin (1992). Public facility location and urban spatial struc-
ture - Equilibrium and welfare analysis. Journal of Public Economics, 48, 83-118.

[86] Timmins C. (2007). If you can’t take the heat, get out of Cerrado... recovering the
equilibrium amenity cost of nonmarginal chimate change in Brazil. Journal of Re-
gional Science 47(1), 1-25.

[87] Turnbull, G. K. (1991). The spatial demand for housing with uncertain quality.
Journal of Real Estate Finance and Economics, 4, 5-17 (1991).

[88] Turner M. (2005). Landscape preferences and patterns of residential development.

Journal of Urban Economics 57, 19-54.

[89] Turner M.A., A. Haughwout, and W. Van der Klaauw (2014). Land Use Regulation
and Welfare. Econometrica 82(4), 1341-1403.
180 BIBLIOGRAPHY

[90] Tyrväinen L. and A. Miettinen (2000). Property prices and urban forest amenities.
Journal of Environmental Economics and Management 39(2), 205–223.

[91] Tyrväinen L. (2001). Use and valuation of urban forest amenities in Finland. Journal
of Environmental Management 62, 75–92.

[92] Tyrväinen L., Pauleit S., Seeland K., de Vries S. (2005) Benefits and Uses of Urban
Forests and Trees. In: Konijnendijk C., Nilsson K., Randrup T., Schipperijn J. (eds)
Urban Forests and Trees. Springer, Berlin, Heidelberg.

[93] Tzoulas K., K. Korpela, S. Venn, V. Yli-Pelkonen, A. Kaźmierczak, J. Niemela, and

P. James (2007). Promoting ecosystem and human health in urban areas using
Green Infrastructure: A literature review. Landscape and Urban Planning 81, 167-
178. doi:10.1016/j.landurbplan.2007.02.001.

[94] UN DESA (2014). World Urbanization Prospects: The 2014 Revision.

[95] UN DESA (2018). World Urbanization Prospects: The 2018 Revision, custom data
acquired via website. Accessed on June 19th , 2018.

[96] van den Bosch M. and A.O. Sang (2017). Urban natural environments as nature-
based solutions for improved public health - A systematic review of reviews. En-
vironmental Research 158, 373-384.

[97] Walsh R. (2007). Endogenous open space amenities in a locational equilibrium.

Journal of Urban Economics 61, 319-344.

[98] Warziniack T. (2010). Efficiency of public goods provision in space. Ecological Eco-
nomics 69, 1723-1730.

[99] White H. (1980). A Heteroskedasticity-Consistent Covariance Matrix and a Direct

Test for Heteroskedasticity. Econometrica 48, 817-838.

[100] White M.P., I. Alcock, B.W. Wheeler, and M.H. Depledge (2013). Would you be
happier living in a greener urban area? A fixed-effects analysis of panel data.
Psychological Science 24, 920-928.

[101] Wu J. and S.H. Cho (2003). Estimating households’ preferences for environmental
amenities using equilibrium models of local jurisdictions. Scottish Journal of Politi-
cal Economy 50(2), 189-206.

[102] Wu J. and A. Plantinga (2003). The influence of public open space on urban spatial
structure. Journal of Environmental Economics and Management 46, 288-309.

[103] Zhang L. (2016). Flood hazards impact on neighborhood house price: A spatial
quantile regression analysis. Regional Science and Urban Economics 60 (2016) 12-19.