ESSAYS ON APPLIED MICROECONOMICS

By

Kyung Ho Song

A DISSERTATION

Submitted to

Michigan State University

in partial fulﬁllment of the requirements

for the degree of

Economics — Doctor of Philosophy

2019

ABSTRACT

ESSAYS ON APPLIED MICROECONOMICS

By

Kyung Ho Song

Chapter 1. Education and Neighborhood Premium in Housing Price and its Dy-

namics

From a newly-assembled panel dataset, this study estimates the implicit prices of better

education and neighborhood environments through models in which unobservable housing

attributes are allowed to vary over time. The results of this study show that, ﬁrst, past

records of school performances have a long-lasting reputation eﬀect, which has not been

considered in the previous literature. Without taking the reputation eﬀect into account,

estimation results can be biased. Second, education and neighborhood premiums account

for more than 43% of the housing price gaps. This paper also analyzes the dynamics of the

housing market by using the panel VAR, a new approach in urban economics. The dynamic

model shows more remarkable eﬀects of education and neighborhood environments on hous-

ing price than a static model through bilateral – direct and indirect – relations among school

performance, neighborhood composition, amenities, and housing price over time. This result

supports the idea that education and neighborhood environments have more substantial im-

pacts on housing prices and city segregation in a dynamic framework.

Chapter 2. Impact of Resale Price Maintenance (RPM) Regulation on the Book

Market of Korea

The reinforced resale price maintenance (RPM) regulation applied to books took eﬀect from

November 2014 in Korea under the policy purpose of preserving cultural variety and knowl-

edge by supporting small publishing companies and bookstores. However, empirical evidence

implies that the new regulation should be revised to achieve its policy purpose. After the

new regulation, the number of newly-published books, which can be used as the index for

cultural variety, was decreased by 7% annually after de-trending. The proﬁt ratio of the

six largest online ﬁrms was enhanced by more than 5% and the amount of their proﬁt also

increased signiﬁcantly. However, the top 70 publishing companies’ sales decreased by more

than 10% and the sales of small book stores also decreased. Based on the empirical analyses,

this paper suggests that other methods such as direct subsidy should be reviewed instead of

price regulation to achieve the policy’s purpose.

Chapter 3. Impact of Resale Price Maintenance (RPM) Regulation on the Used

Book Market

A new bill came into eﬀect in November 2014 that further strengthens the existing resale

price maintenance (RPM) regulation on the book market in Korea under the policy purpose

to foster diversity of culture and knowledge by protecting small publishing companies and

oﬄine bookstores. However, after the new regulation took eﬀect, some studies ﬁnd evidence

that is contrary to the intention of the regulation from various data. Most importantly, the

gap in the market share between large online ﬁrms and small-scale oﬄine bookstores has ex-

panded even more, and the number of new publications has decreased signiﬁcantly. Another

noticeable change since the introduction of the new RPM regulation is the rapid growth of

the used book market that has also beneﬁted large online bookstores because the large online

ﬁrms have dominated the used book market. Theoretically, this study constructs a compre-

hensive model including the used book market, which has properties of a two-sided market

and is linked with the new book market. From the model, this paper explains various eﬀects

of new RPM regulations on the used book market.

Copyright by
KYUNG HO SONG
2019

To my parents,

my wife, Dayeon Lee

and sons, Yejun and Yeonjun

v

ACKNOWLEDGEMENTS

I am very grateful to my advisor Kyoo il Kim for his constant support, encour-

agement and helpful feedback. I would also like to thank Jay Pil Choi and Thomas

Jeitschko for their invaluable help and advice. This paper also beneﬁted from com-

ments by seminar participants at Michigan State University.

I am deeply indebted to my parents, Samyoung Song and Nasoon Sun, for their

enormous support and love. My two sons, Yejun and Yeonjun, became the joy of my

life and the source of motivation and encouragement.

Last but by no means least, my wife, Dayeon Lee, has gladly endured a hard

time by my side during doctoral program while taking care of two sons with sincere

love. Today’s accomplishments and the happiness of our family are all thanks to her

sacriﬁce, eﬀort, and love.

vi

TABLE OF CONTENTS

LIST OF TABLES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

LIST OF FIGURES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

1.3 Hedonic Price Regressions

Chapter 1. Education and Neighborhood Premium in Housing Price
and its Dynamics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.2 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.2.1 Data Matching . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.2.2 Panel Dataset Generating Process
. . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . .
1.3.1 School Reputation Eﬀect
. . . . . . . . . . . . . . . . . . . . . .
1.3.2 Education Environment and City Segregation . . . . . . . . . . .
1.4 Policy Simulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.5 Dynamic Model
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
APPENDIX . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
A.1 Data Appendix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . .
A.2 Table & Figure Appendix . . . . . . . . . . . . . . . . . . . . . . . . . .
A.3 Structural Panel VAR . . . . . . . . . . . . . . . . . . . . . . . . . . . .

A.1.1 Housing Environment in Seoul
A.1.2 Education and Neighborhood Statistics
A.1.3 Panel Dataset Generating Process

ix

xi

1
1
8
15
17
19
22
27
32
38
47
49
50
50
51
58
60
73

Chapter 2. Impact of Resale Price Maintenance (RPM) Regulation

on the Book Market of Korea . . . . . . . . . . . . . . . . . . . . . . . . 76
76
2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
82
2.2 Impact of RPM Regulation on the Book Market
. . . . . . . . . . . . .
2.2.1 Market Size and Online Share . . . . . . . . . . . . . . . . . . . .
82
85
2.2.2 Variety of New Book Publications
. . . . . . . . . . . . . . . . .
88
2.2.3 Price of Books . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.3 Eﬀect of RPM on Firms . . . . . . . . . . . . . . . . . . . . . . . . . . .
96
2.3.1 Online Firm . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
96
2.3.2 Publishing Company . . . . . . . . . . . . . . . . . . . . . . . . . 102
2.3.3 Oﬄine Book Stores
. . . . . . . . . . . . . . . . . . . . . . . . . 106
2.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
APPENDIX . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110

vii

Chapter 3. Impact of Resale Price Maintenance (RPM) Regulation

on the Used Book Market . . . . . . . . . . . . . . . . . . . . . . . . . . 117
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
3.2 Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120
3.2.1 Structure of Used Book Market . . . . . . . . . . . . . . . . . . . 120
3.2.2 Demand of Used Book Market
. . . . . . . . . . . . . . . . . . . 123
3.2.3 Supply of Used Book Market . . . . . . . . . . . . . . . . . . . . 134
3.3 Eﬀects of New RPM Regulation and the Equilibrium of Market . . . . . 141
3.3.1 Proﬁt of Oﬄine Stores . . . . . . . . . . . . . . . . . . . . . . . . 141
3.3.2 Equilibrium of the Used Book Market . . . . . . . . . . . . . . . 143
3.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 144
APPENDIX . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146

BIBLIOGRAPHY . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148

viii

LIST OF TABLES

Table 1.1 Summary Statistic . . . . . . . . . . . . . . . . . . . . . . . .

Table 1.2 School Reputation Eﬀect

. . . . . . . . . . . . . . . . . . .

Table 1.3 Willingness to Pay for Educational Environments . . . .

Table 1.4 Housing Price Gap Decomposition . . . . . . . . . . . . .

Table 1.5 WTP for the School Zone Adjustment . . . . . . . . . . .

Table 1.6 The Changes in Experimental Area . . . . . . . . . . . . .

Table A1 Summary Statistic . . . . . . . . . . . . . . . . . . . . . . . .

Table A2 School Reputation Eﬀect – 25 District Fixed Eﬀect – . . . . .

Table A3 Reputation Eﬀect (House Fixed-Eﬀect) – 11 School Zone – . . .

Table A4 Reputation Eﬀect (Time-varying) – 11 School Zone – . . . . . .

Table A5 Reputation Eﬀect (House Fixed-Eﬀect) – 25 District –

. . . . .

Table A6 Reputation Eﬀect (Time-varying) – 25 District –

. . . . . . . .

Table A7 WTP for Educational Environments – 25 District –

. . . .

Table A8 Housing Price Gap Decomposition – 25 District –

. . . . .

Table A9 WTP for the School Zone Adjustment without Weight .

Table A10 The Changes in Experimental Area without Weight

. .

Table 2.1 Amendment of RPM Regulation . . . . . . . . . . . . . . .

Table 2.2 The Book Market of Korea . . . . . . . . . . . . . . . . . .

Table 2.3 The 7 Biggest Online Firms . . . . . . . . . . . . . . . . . .

Table 2.4 The Number of Newly-Published Books by Categories .

Table 2.5 Eﬀect of RPM on the Number of New Publications . . .

Table 2.6 The Average List Price of Books by Category . . . . . .

Table 2.7 Eﬀect of RPM on List Price . . . . . . . . . . . . . . . . .

Table 2.8 Eﬀect of RPM on List Prices (RPM Trend)

. . . . . . .

Table 2.9 Eﬀect of RPM on List Prices (Panel) . . . . . . . . . . . .

Table 2.10 Eﬀect of RPM on List Prices (Panel, RPM Trend) . . .

Table 2.11 Eﬀect of RPM on Online Firm’s Proﬁt Ratio . . . . . . .

13

23

28

30

37

37

60

63

64

65

66

67

68

69

70

70

78

82

83

86

87

90

91

92

94

95

99

Table 2.12 Eﬀect of the New RPM on Online Firms

. . . . . . . . . 101

Table 2.13 Eﬀect of RPM on Publishing Firm’s Sales . . . . . . . . . 104

ix

Table 2.14 Eﬀect of RPM on Publishing Firm’s Proﬁt Ratio . . . . 105

Table 2.15 Eﬀect of RPM on Publishing Firm’s Proﬁt . . . . . . . . 106

Table 2.16 Survey Result of Oﬄine Stores . . . . . . . . . . . . . . . . 107

Table B1 The Number of Newly-Published Books . . . . . . . . . . 111

Table B2 The Average Price of Newly-Published Books . . . . . . 111

Table B3 The Number of Circulation . . . . . . . . . . . . . . . . . . 112

Table B4 Eﬀect of RPM on List Price . . . . . . . . . . . . . . . . . 113

Table B5 Eﬀect of RPM on List Price (RPM Trend) . . . . . . . . 114

Table B6 Eﬀect of RPM on List Price (Panel) . . . . . . . . . . . . 115

Table B7 Eﬀect of RPM on List Price (Panel, RPM Trend) . . . . 116

Table 3.1 Eﬀect of Used Book Market Growth on Oﬄine Stores . 120

x

LIST OF FIGURES

Figure 1.1 School Zone and School Performance Gap . . . . . . . .

Figure 1.2 Neighborhood Segregation and Housing Price Gap . . .

Figure 1.3 School Zone Adjustment . . . . . . . . . . . . . . . . . . .

Figure 1.4 School Zone Adjustment Simulation . . . . . . . . . . . .

Figure 1.5 (Cumulative) IRF – (Impulse) : (Response) – . . . . . . . . .

Figure 1.6 (Cumulative) IRF – Impulse = Top Univ (1%p) –

. . . . . .

Figure 1.7 (Cumulative) IRF – Response = Top Univ (%p) –

. . . . . .

Figure 1.8 (Cumulative) IRF – Impulse = % BA or Above (1%p) –

. . .

Figure 1.9 (Cumulative) IRF – Response = % BA or Above (%p) –

. . .

Figure A1 School Zone, Top Performing High School

. . . . . . . .

Figure A2 Neighborhood Segregation and Housing Price . . . . . .

Figure A3 (Cumulative) IRF – Response = Housing Price (%) – . . . . .

Figure A4 (Cumulative) IRF – Impulse = Housing Price (1%) – . . . . .

Figure A5 Structural Panel VAR, (Cumulative) IRF . . . . . . . .

Figure A6 Panel VAR, (Cumulative) IRF – (Impulse) : (Response) – .

Figure 2.1 Market Size and On-line Share . . . . . . . . . . . . . . .

Figure 2.2 Top 6 Online Firm’s Sales, Cost and Proﬁt . . . . . . . .

Figure 2.3 The Proﬁt Ratio of online Firms . . . . . . . . . . . . . .

9

10

33

36

40

42

43

44

46

61

62

71

72

74

75

84

97

98

Figure 2.4 Proﬁts and Sales of Top 6 Online Firm . . . . . . . . . . 100

Figure 2.5 The Publishing Industry Production Index . . . . . . . . 103

Figure 3.1 Used Book Market . . . . . . . . . . . . . . . . . . . . . . . 121

Figure 3.2 Discrete Used Book Quality Model . . . . . . . . . . . . . 127

Figure 3.3 The Eﬀect of Increasing Used Book Price . . . . . . . . 129

Figure 3.4 The Eﬀect of Changing Parameters

. . . . . . . . . . . . 130

Figure 3.5 Probability of Buying Each Book for Vi

. . . . . . . . . . 132

Figure 3.6 Changes in Probability Functions by Increased P (qL)

. 133

Figure 3.7 Demand Curve of High Quality Used Book . . . . . . . 134

Figure 3.8 Locations and Individual Values of Each Seller . . . . . 136

Figure 3.9 Network Eﬀect in Transaction Cost . . . . . . . . . . . . 139

xi

Figure 3.10 Reaction Curves of Online Firm and Oﬄine Store . . . 140

Figure 3.11 Eﬀect of Increasing Demand on Reaction Curves . . . . 142

Figure 3.12 Eﬀect of Increasing Demand on Used Book Market . . 144

xii

Chapter 1. Education and Neighborhood Premium in Housing

Price and its Dynamics

1.1 Introduction

The hedonic price regression, using housing market data, has been widely used as

the methodology to estimate the implicit prices of local public goods such as school

quality, crime rate, and air pollution.1 Economists, especially, have been trying to

estimate households’ valuations for better school performance and neighborhood envi-

ronments, given their relevance to important empirical issues.2 The primary purposes

of estimating the implicit prices of better school quality and neighborhoods can be

summarized into two categories. First, estimates illustrate households’ valuation for a

better performing school and for living around more educated neighbors. Second, es-

timates allow us to ﬁgure out how school quality and neighborhood composition aﬀect

micro-geographic patterns of housing prices. Also, estimating households’ preferences

for schools and neighbors is essential to study school and residential segregation.

In one representative study, Black (1999) examines the parental valuation of

elementary education by developing the boundary discontinuity design (BDD). The

study ﬁnds that a 5% increase in elementary school test scores leads to a 2.1% higher

housing price. Bayer, Ferreira, and McMillan (2007) embed the BDD in both the

sorting model, which considers residential location decision of each household, and the

hedonic regression model. The study shows that households are willing to pay less

than 1% more for housing when the average test score of the local school increases by

5%. They also ﬁnd that households prefer to self-segregate based on both race and

education.3

1 Gibbons and Machin (2008) review the empirical literature on hedonic analysis valuing school quality,

better transportation, and lower crime.

2 Black and Machin (2011), Machin (2011) and Nguyen-Hoang and Yinger (2011) oﬀer comprehensive

reviews of this literature. Black and Machin introduce and classify 54 papers studying this topic.

3 Also, Kane, Riegg, and Staiger (2006) investigate school districts under a court-imposed desegregation
order and its eﬀect on housing price. According to their results, 1 SD increases in school quality increases
housing price by 10%. Clapp, Nanda, and Ross (2008) show 1 SD increase in students’ math scores increases
housing price by 1.3∼1.4% while neighborhood characteristic (the portion of Black and Hispanic) becomes
more important in the long run. Figlio and Lucas (2004), Fiva and Kirkeboen (2011), Gibbons, Machin, and
Silva (2013), and Imberman and Lovenheim (2015) investigated the eﬀect of newly released information such
as school report card, school ranking, and school and teacher value-added on the housing price. Boustan
(2012) shows the desegregation of public schools in central cities reduced the demand for urban residence,
leading urban housing prices and rents to decline by 6% relative to neighboring suburbs.

1

Reliable estimates for the implicit prices are also required in advance to construct

a relevant demand function, which is adopted for counterfactual welfare analyses and

policy simulations for various education and school policies. For instance, if a local

government considers an increase in expenditure for school quality enhancement in

their municipal area, before making the decision, they would compare its planned

spending with residents’ willingness to pay for better school quality.

However, analyzing the demand for housing is challenging empirical work be-

cause a house is a bundle of various types of characteristics. The ﬁrst type pertains

to the physical attributes of a house (size, age, the number of rooms, and type of

building). The second attribute concerns local public goods of a community such as

education level of the surrounding neighborhood, school quality of the nearby school,

transportation systems, crime rate, and air pollution in the area. The third character-

istic includes amenities around a house (e.g., supermarkets, shopping centers, sports

facilities, hospitals, restaurants, ﬁre and police stations, and etc.), and the fourth

type of characteristic is the distance from a house to other locations (e.g., workplaces,

schools, and subway stations).

The challenge for economists is that it is impossible to control for all these char-

acteristics in the analysis due to the lack of data, and many of the factors aﬀecting

housing price are diﬃcult to measure by numerical value (e.g., curb appeal and the

quality of landscaping). Then, such unobservables and omitted variables cause bias

problems because they are likely to be correlated with other attributes included in

the model.

In the case when a cross-sectional dataset is available, instrument vari-

ables or the boundary discontinuity design approach can be adopted to handle the bias

problem. In practice, however, it is hard to ﬁnd appropriate instrumental variables

in the housing market that satisfy exclusion restrictions in the IV framework, and

those instruments are not usually free from an ad-hoc selection problem. Exploiting

discontinuities at a jurisdictional boundary, the use of BDD has been widely adopted

in modern empirical research. To estimate the preference for school quality, the BDD

investigates both sides of the narrow area that borders a school zone boundary.

In

most applications, however, school zone districts are coterminous with administrative

districts and geographical discontinuity by a broad road or a river. Also, the possibility

2

of household self-segregation – based on the diﬀerent school quality – leads to other

critical diﬀerences (e.g., neighborhood composition) between both sides, while making

it hard to obtain a clean variation in school quality.

On the other hand, if a panel dataset with repeated housing transactions is avail-

able, we can handle the endogeneity problem by eliminating the time-constant omitted

variables from the diﬀerenced models. However, another problem in the housing mar-

ket is that repeated transactions for the same house rarely happen, thereby requiring

the use of long-term panel dataset to obtain a suﬃcient sample size. For this reason,

if we use such a long-term panel dataset, the possibility that unobservable information

and omitted variables could vary over time precludes the diﬀerenced model, solving

the bias problem. In particular, when we analyze the housing market of an urban area

where many amenities usually change fast compared to rural areas, it is not reasonable

to assume that the unobservables and omitted variables are ﬁxed.

The ﬁrst main methodological feature of this paper, as the solution for the above

problems, involves constructing a short-term (two consecutive years) panel dataset. It

is more plausible that the unobservables causing the bias problem are ﬁxed over time

as the dataset covers a shorter period. From the dataset, this paper analyzes the

housing market in Seoul, Korea with assorted amenity information of houses obtained

by merging various data sources. The short-term panel dataset – with suﬃcient sample

size – is generated from using a characteristic of the apartments, which is that an

apartment building contains many housing units that are almost identical within the

same building. The details of this method are explained in Section 1.2.2. Along

with constructing a short-term panel dataset, for more robust estimations, this study

considers the case that unobservables vary within the two years under the assumption

that it follows the ﬁrst-order Markov process. For the Time-varying Unobservable

Model, this paper closely follows the method developed by Bajari, Fruehwirth, Kim,

and Timmins (2012).

In the previous literature, housing price hedonic models usually include physical

attributes of a house, neighborhood characteristics, other variables of interest according

to a topic (e.g., school quality, air quality, and crime rate), and regional- and time-ﬁxed

eﬀects to control an omitted variable bias. As noted by Bayer, Ferreira, and McMil-

3

lan (2007), however, including regional- and time-ﬁxed eﬀects may not be enough to

control all omitted variables and unobservable information for an unbiased estima-

tion. According to the study, the inclusion of precise neighborhood socio-demographic

information reduces the coeﬃcient of school quality in a hedonic regression by 50%,

compared to the model simply including boundary ﬁxed eﬀects. This result implies

that it is hard to represent region-speciﬁc omitted information and unobservables with

regional-ﬁxed eﬀects. For this reason, this study tries to control as much information

as possible along with typical regional- and time-ﬁxed eﬀects.4 The models in this

paper also controls the number of similar sized houses within a certain radius from a

house to consider the supply side of the housing market.5 The variables used in this

paper are listed in Table 1.1 and Table A1 in the appendix.

Another feature of this study is to adopt a spatial modeling technique using the

distance-weight matrices to match a particular house with amenity information.

In

the housing market literature, amenity information is usually matched with a house by

census tract level when the exact location of a house is not identiﬁed. However, given

that people can beneﬁt from amenities located nearby but in diﬀerent census areas,

using the spatial distance-weight is a more precise way to model the housing markets.6

To combine the spatial information with hedonic and dynamic models, this study uses

the exact locations of houses, schools (elementary, middle, and high school), subway

stations, department stores, supermarkets, centroids of census tracts, and air quality

monitors by changing the address information into longitude and latitude.

From the newly-assembled panel dataset with the above features, this paper

shows the evidence that there exists a long-lasting reputation eﬀect in the implicit price

of school quality. When we evaluate a school, it is natural to consider the past records

of school performance together with the most recent ones. Let’s suppose three schools

(“A,” “B,” and “C” ) have the same performance records in the most recent year. The

4 This paper considers regional ﬁxed eﬀect for each of the 11 school zone level, and monthly time ﬁxed
eﬀects for each region. As a robustness check, this study also tries generating regional and time ﬁxed eﬀect
for 25 district level. The district area is smaller than a school zone. Two or three districts consist of a school
zone in Seoul.

5 Models include the number of similar sized houses within 0.5 km, 1 km and 2 km radius from a house.
i) less than 60 m2, ii) 60∼100 m2, iii)

In this process, the size of house was divided into four categories.
100∼130 m2, and iv) larger that 130 m2

6 More details about distance-weight matrices are explained in Section 1.2.1.

4

school “A” is a prestigious school which has stable, long-standing good performance

records. However, the performance of the school “B” has improved recently and the

previous records are not good. Last, the school “C” has no previous records because the

school is newly-established. In this case, parents may have a higher willingness to pay to

live near the prestigious school than the other two schools, since parents also care about

the past records of a school. Even though it is intuitively natural, I could not ﬁnd any

previous research explicitly considering the above reputation eﬀect, which comes from

past records of a school. According to the result of this study, people consider the past

records of a school up to ten years while placing more weight on recent performances

of a school, and the weight fades out as time passes by. The estimated implicit price

for one school year result is relatively smaller; however, its accumulated value for long-

term good performances of a school becomes more substantial than estimates of prior

studies. The previous studies have used the school performance statistics of a speciﬁc

year without considering previous records, and some studies used a two-year, three-year

or four-year average of the school statistics to reduce any year-to-year noise. However,

such averaging is diﬀerent from ﬂexibly considering the above reputation eﬀect; thereby

it can cause bias estimation results for the implicit prices of school performances.7

The second purpose of estimating households’ preferences for schools and neigh-

bors is linked to studies about segregation issues among schools, neighborhoods, and

housing prices. Much of the previous literature focuses on residential segregation driven

by race. For example, Card, Mas, and Rothstein (2008) ﬁnd evidence of a tipping point
(5∼20% minority share) and study the dynamics of racial segregation within a city.8
Another category of segregation studies analyzes the eﬀects of neighborhood and school

segregation – also by race – on school achievement gaps9 or school segregation itself

from the viewpoint of students’ ethnic composition.10 Card and Rothstein (2007) study

7 For example, if we use the variable averaging test score results from the past three years, the estimated
coeﬃcient of the variable is the same with the result from the model that imposes the restriction – all coeﬃcients
for the past three year results are the same when we include each three previous results independently in the
model.

8 About this topic, see also Bajari and Kahn (2005), Bayer, McMillan and Rueben (2004), Bayer, Fang
and McMillan (2014), Bayer and McMillan (2012), Easterly (2009), Zhang (2009), Bischoﬀ and Reardon
(2014), Reardon and Yun (2001) and Boustan (2010, 2012).

9 Bifulco and Ladd (2007), Reardon (2011, 2016) and Reardon and Galindo (2009) are also included in

this category.

10 See Reardon, Yun and Eitle (2000), Reardon, Grewal, Kalogrides, and Greenberg (2012), and Cartano

and Maheshri (2017).

5

the eﬀects of school and neighborhood segregation on the relative SAT scores of black

students across diﬀerent metropolitan areas. Lastly, there are a few studies that in-

vestigate residential segregation driven by school segregation. Baum-Snow and Lutz

(2011) examine the residential location and school choice responses to the desegrega-

tion of large urban public school districts. Overall, the common feature of previous

literature is a focus on the unilateral eﬀect from one to another ((i) neighborhood seg-
regation ⇒ school segregation or school achievement gaps, (ii) school achievement gaps
⇒ neighborhood composition, and (iii) school quality or neighborhood composition ⇒
housing price) or an analysis of the residential and school segregations driven by race.

Diﬀerent from previous literature, the contribution of this study is to analyze the

bilateral – direct and indirect – dynamic relations among school quality, neighborhood

characteristics, housing prices, and other amenities over time, based on the panel VAR

(vector autoregression) model and its impulse-response analysis. A coeﬃcient in the

hedonic regression is interpreted as an implicit price of a variable and shows the eﬀect on

the market price of increasing a particular attribute while holding the other attributes

ﬁxed. Hence, we can think of a hedonic model as showing the static picture of a housing

market in an equilibrium status. In practice, however, if one of the housing attributes

changes, other attributes of the same house could be aﬀected and change over time.

For example, enhanced school qualities may increase housing prices (school qual-
ity ⇒ housing price) while attracting more educated neighbors around the school
(school quality ⇒ neighborhood composition) if more educated people prefer a bet-
ter performing school. Then, the increased neighborhood education level caused by
the improved school quality can increase the housing price further (school quality ⇒
neighborhood composition ⇒ housing price) because a neighborhood composition is
another critical housing attribute aﬀecting housing prices. This is the indirect eﬀect of

enhanced school qualities on housing price by attracting another good attribute – more

educated neighborhoods. It is also likely that a higher neighborhood education level,

in the opposite direction, can improve performances of a school (neighborhood com-
position ⇒ school quality) because parental education level may aﬀect their children’s
academic performance. This is the bilateral relationship between school performances
and neighborhood education level (school quality ⇔ neighborhood composition). Ad-

6

ditionally, the increased portion of highly educated people in the neighborhood itself
can also attract more educated people in the region (neighborhood composition ⇒
neighborhood composition) due to household self-segregation based on education level.

To summarize, there could be bilateral and indirect relationships among important

housing attributes over time. Bayer, Ferreira, and McMillan (2007) refer to these ef-

fects as ‘second-round social multiplier.’ To the best of my knowledge, this is the ﬁrst

paper analyzing the above dynamics in the housing market.

Recently, there have been some studies analyzing the dynamics of the housing

market. Bayer, Keohane, and Timmins (2009) examine the dynamics of households’

residential choice considering moving cost and ﬁnd that the estimates are three times

greater than the marginal willingness to pay, which was estimated by a conventional

hedonic model. Bayer, McMillan, Murphy, and Timmins (2016) introduce the dynamic

of the neighborhood choice problem through two channels: wealth accumulation and

moving costs. In their setting, households are treated as making a sequence of location

decisions that maximize the discounted sum of expected per-period utilities. That

is, the dynamic analysis in the previous literature is based on the theoretical model

considering dynamic decisions such as maximizing the households’ lifetime utility from

the given utility function.

From the short-term panel regression model and dynamic model, this paper ﬁnds

evidence that there exists a long-lasting reputation eﬀect in the implicit price of school

quality. The second ﬁnding comes from the results of price gap decomposition analysis.

Similar with other metropolitan cities, residential segregation also has been occurring

in Seoul.

In Seoul, the Gangnam area is preferred to the other regions for many

reasons. Since there is no racial diﬀerence in Seoul, the residential segregation is driven

by parents’ education level, income, and wealth. Regarding neighborhoods’ education

level, school performance, and average housing price (as a proxy for incomes and wealth

of an area), there are substantial diﬀerences between Gangnam and the other districts

in Seoul. According to the estimation results of this study, more than 43% of the

housing price gap between the areas is explained by the diﬀerences in education and

neighborhood environments.11 Based on the estimation results, this study simulates

11 The top university entrance rate (Seoul National University), the portion of BA or above degree holder

7

an education policy that adjusts school zone boundaries. The government of Seoul is

now considering school zone adjustment as an option for immoderate housing price

gaps and education performance gaps within the metropolitan area. The simulation

predicts the eﬀects of school zone change binding two diﬀerent areas (one of the best

performing areas and a below average area) that are geographically adjacent to each

other. In the new school zone, the housing price gaps are estimated to be diminished

by 13% in the long-term.

The third ﬁnding comes from the dynamic model. This study ﬁnds evidence

that positive bilateral dynamic relations among school quality, neighborhood educa-

tion level, amenities, and housing price exist. That is, those housing attributes turn out

to aﬀect each other making the eﬀects of school performance and neighborhood com-

position more signiﬁcant by bilateral relationships over time in a dynamic framework.

These results imply the impact of school performance and neighborhood composition

on housing price and city segregation can be underestimated in a static model if dy-

namic eﬀects are ignored.

This paper proceeds as follows: Section 1.2 describes the data and explains

the panel dataset generating process. Section 1.3 presents results from the models

which allow unobservables to vary over time. Section 1.4 shows results from the policy

simulation (school zone adjustment). Section 1.5 analyzes dynamics of the housing

market by using the panel VAR approach. Section 1.6 concludes.

1.2 Data

This section describes the main data sources and explains how they are matched with

the housing transactions in the Seoul metropolitan area in Section 1.2.1. For the

matching process, this paper adopts a spatial modeling technique frequently used in a

GIS analysis. Section 1.2.2 shows the process of constructing the panel dataset.

As the spatial summary, Figure 1.1 and 1.2 show the primary statistics of this

paper on the map of Seoul. Figure 1.1a shows the school zone map of Seoul. Areas

with the same color and pattern represent the same school zone. There are 25 districts
in 40∼49 age group, and private education environment are the three components that explain almost 50% of
the housing price gaps in Seoul.

8

(a) School Zone Map

(b) Top University Entrance Rate

Figure 1.1 School Zone and School Performance Gap

9

(a) University Graduates (%, Age 40∼49)

(b) Housing Price per Size ($/m2)

Figure 1.2 Neighborhood Segregation and Housing Price Gap

10

and 11 school zones in Seoul. The Seocho and Gangnam districts form the Gangnam

school zone (the area with oblique lines).

In terms of housing price, neighborhood

composition, and school performance gaps, the rest of the ﬁgures show the substantial

diﬀerences between the Gangnam school zone and the other areas of Seoul. In Figure

1.1b, each circle represents the location of a high school, and the size of the circle is

proportionate to the top university entrance rate of a high school.12 Figure 1.2a shows
the percent of university graduates (BA or above degree holders) in the 40∼49 age
group in the 423 census tracts. As the color gets darker, it means a higher portion of

well-educated neighbors reside in the area. Figure 1.2b represents the housing price

per size (m2). Each circle identiﬁes the location of a house, and darker points mean

more expensive houses. Note that large circles (Figure 1.1b), dark areas (Figure 1.2a),

and dark points (Figure 1.2b) are highly concentrated in the Gangnam school zone.

To clearly illustrate the gaps between regions within Seoul, Figure A1 and A2 in

the appendix show the locations of top performing high schools (Figure A1b), census

tracts where the portion of university graduates is higher than 62% (Figure A2a), and

the locations of expensive houses within the top 5% (Figure A2b). Also, to compare

both areas, Table 1.1 and A1 show summary statistics of variables for all areas of

Seoul, and for Gangnam and the other areas separately. The next paragraphs explain

the main data sources where the variables in this paper are drawn from to construct

the panel dataset.

A. Housing Transaction and Location Data

The Ministry of Land, Infrastructure, and Transportation Department of Korea pro-

vides transaction data on houses.13 This dataset contains information about the real

transaction price, housing type (apartments or single detached house), speciﬁc address,

size, construction year, and transaction date. In the case of apartment transactions,

it also includes the building name or brand of apartments and the ﬂoor on which the

apartment is located. Table 1.1 – (A.1) describes the summary statistics of the phys-

ical attributes of a house (price, ﬂoor, age, size, complex, and brand).14 From the

12

In this paper, the top university entrance rate means the portion of students who entered Seoul National

University (SNU) by high school.

13 http://rt.molit.go.kr
14 The details about complex (the type of apartment) and brand of apartments are explained in the data

11

speciﬁc address information, I generate distance variables from a house to some impor-

tant places and amenities. Table 1.1 – (A.2) shows the distance variables used in this

study. The distances from a house to the nearest school (high, middle, and elementary

school),15 subway station, Han-river, department store, and a large supermarket chain

are used as distance variables. More detailed explanations about some unique charac-

teristics of the housing market of Seoul, Korea are in the data appendix.

B. Population and Housing Census

The Population and Housing Census data is used to obtain neighborhood demographic

information surrounding a house. From the data, I calculate the portion of people

by education level and age group, such as the portion of university graduates (BA or
above degree holders) in the 40∼49 age group by each census tract. This study uses the
portion of BA or above degree holders in the 40∼49 age group as the primary index for
neighborhood education level. The ﬁrst reason is that most people in this age group

are school parents who have a stake in the educational environment. Second, due to

the rapidly increasing trend in the portion of BA or above degrees in Korea, the BA

degree no longer represents a high education level in younger age groups. Also, this
paper uses the population of all residents, students, people aged 35∼55, and those over
65 and calculates each portion by census tract. From the census data, I use the number

of households by house type and its composition (lives alone, a married couple, and

a family with children) by census tract. The above neighborhood characteristics are

matched with a house by adopting the distance-weight matrices that use the distances

from a house to centroids of census tracts. This paper explains the matching method

in Section 1.2.1.

C. Business Census

The Business Census oﬀers information about the number of businesses by 228 indus-

try classiﬁcations and the number of employees by the types of employment (full-time,

appendix.

15

I calculated the distances to the third nearest middle and high schools because the school allocation
system does not guarantee assignment to the nearest middle and high school from their house. Meanwhile, in
the case of an elementary school, there exists one-to-one matching between a house and a school. More details
are in Section 1.2.1

12

Table 1.1 Summary Statistic

A. House Attributes & Distance

A.1 Apartments Attributes
Price ($)
Floor
Age (Year)
Size (m2)
Complex (0 or 1)
Brand (0 or 1)

A.2 Distance to (km)
High School 1
High School 2
High School 3
Middle School 1
Middle School 2
Middle School 3
Elementary School
Subway
Han-river
Department Store
Supermarket

A.3 New House Supply
Radius : 0.5 km
Radius : 1 km
Radius : 2 km

All

Gangnam

The Other

Average Median

SD

Average Median

SD

Average Median

SD

426,655
8.83
16.6
77.96
0.03
0.15

357,273
8.00
15.0
83.86
-
-

282,497
5.93
8.9
28.58
0.18
0.36

752,714
8.78
19.2
87.83
0.12
0.14

680,000
7.00
18.0
84.79
-
-

409,688
6.65
10.9
38.53
0.32
0.35

357,002
8.84
16.1
75.86
0.02
0.15

327,273
8.00
15.0
82.53
-
-

182,854
5.77
8.3
25.47
0.13
0.36

0.73
1.14
1.49
0.49
0.84
1.13
0.33
0.60
4.47
2.39
1.35

0.64
1.07
1.43
0.44
0.78
1.07
0.30
0.52
3.51
2.15
1.19

0.44
0.49
0.55
0.29
0.37
0.44
0.16
0.38
3.70
1.43
0.83

0.63
1.07
1.41
0.49
0.85
1.11
0.35
0.50
2.43
1.87
1.72

0.56
0.97
1.34
0.43
0.76
1.03
0.33
0.46
2.62
1.60
1.59

0.41
0.51
0.53
0.31
0.40
0.53
0.19
0.28
1.59
1.18
0.92

0.75
1.15
1.51
0.49
0.84
1.13
0.32
0.62
4.91
2.50
1.27

0.67
1.09
1.44
0.44
0.78
1.07
0.29
0.53
3.97
2.25
1.11

0.45
0.49
0.55
0.28
0.36
0.42
0.16
0.40
3.88
1.45
0.79

14.32
37.32
131.96

-
-
-

116.35
176.84
333.97

38.86
80.77
223.41

-
-
-

238.20
326.31
497.03

9.27
28.37
113.12

-
-
-

66.99
123.73
285.55

Obs.

199,673

35,146

164,527

B. School Performance

B.1 High School
Top Univ Entrance (persons)
Top Univ. Entrance Rate (%)
N-SAT
Students per School
Students per Class
Students per Teacher
Full-time Teacher (%)

Obs.

B.2 Middle School
Top High School (persons)

– Entrance Rate (%)

Private High School (persons)

– Entrance Rate (%)

Obs.

C. Demographics

BA or Above 40s (%)
BA or Above 50s (%)
Population
Average Age
Portion of Student (%)
Portion of 35∼55 Age (%)
Portion of 65 Age Above (%)
Full-time Employee
Part-time Employee
Self-Employee

Obs.

All

Gangnam

The Other

Average Median

SD

Average Median

SD

Average Median

SD

3.51
0.81
100.8
384
31.1
14.3
85.1

5.8
2.02
23.9
8.57

2.00
0.54
101.3
384
31.4
14.4
86.1

220

4.7
1.90
18.0
7.10

379

4.55
1.00
19.2
105
4.3
2.2
6.6

4.2
1.03
19.6
6.29

9.40
2.04
109.8
442
33.4
15.6
83.4

7.7
2.51
36.5
12.11

8.00
1.76
114.4
445
33.7
15.7
84.8

42

7.3
2.44
29.0
10.60

66

6.55
1.46
15.5
79
4.1
2.2
8.7

4.2
1.08
24.1
6.98

2.12
0.53
98.6
370
30.6
14.0
85.6

5.4
1.92
21.3
7.83

2.32
0.53
19.4
106
4.1
2.1
5.9

4.1
0.99
17.4
5.87

1.00
0.36
99.2
355
31.2
14.1
86.3

178

4.3
1.77
16.3
6.16

313

All

Gangnam

The Other

Average Median

SD

Average Median

SD

Average Median

SD

46.5
32.8
22,616
40.8
13.3
32.5
13.0
8,174
1,559
1,794

43.7
27.8
21,867
40.8
12.9
32.7
13.0
3,524
914
1,452

423

18.3
18.5
9,045
2.1
3.7
2.8
2.6
15,024
2,003
1,632

68.1
56.8
23,300
39.7
15.1
33.9
11.1
16,935
3,316
1,917

18.5
21.2
7,497
1.6
4.2
2.5
2.3
19,659
3,205
1,327

68.4
56.6
21,611
39.4
14.6
33.8
10.6
9,252
2,731
1,725

66

42.3
28.2
22,394
41.1
13.0
32.3
13.3
6,717
1,257
1,784

41.4
24.9
21,867
41.0
12.6
32.6
13.3
2,986
830
1,451

357

14.9
13.8
9,193
2.1
3.5
2.7
2.5
13,673
1,543
1,686

13

part-time, and self-employed) for each business by census tract level. There are 423

census tracts in Seoul, and the average area is 1.4km2 (0.54mile2). Given that housing

prices are aﬀected by the amenities surrounding the house, among the 228 industry

classiﬁcations, this paper selected several classiﬁcations that could aﬀect the housing

price. Table A1 shows the list of selected variables from the Business Census data, and

the employment information of businesses is presented in Table 1.1 – (C). The infor-

mation assembled from the Business Census is also matched with a house by using the

distance-weight matrices as described in Section 1.2.1.

D. Education Information

As variables of interest, this study uses the top university (Seoul National University)

entrance rates and the results of Nationwide Scholastic Achievement Test (N-SAT)

by a high school and entrance rates for private and top high schools by each middle

school.16 In addition, school characteristics such as the number of students per class

and per teacher, as well as the portion of full-time teaching staﬀ are controlled in

the analysis. In most previous literature analyzing the implicit price of school perfor-

mance, researchers usually control the characteristics of a school along with the test

scores, such as the racial composition in a school (e.g., White, Black, Hispanic, and

Asian) and the portion of international students. However, racial diversity does not

exist as such in Korea,17 thereby it oﬀers a better environment to reveal the house-

holds’ willingness to pay for better school performances. Another critical factor in the

educational environments is the distance between a house and a school. This paper

uses the latitude and longitude information of houses and schools and calculates the

exact distances from houses to schools. This paper also uses the number of private

academies by census tract from the Business Census as a proxy for the private educa-

tion environment. There were 13,149 private academies preparing various school exams

16

In the data appendix, this paper explains further information about these statistics. In this study, I use
“top high school” instead of its oﬃcial name “special-purpose high school.” In general, only the top 2∼3%
ranked students in each middle school enter special-purpose high schools.

17 The percent of other race and foreigner is negligible in Seoul. According to the statistics published
by Seoul Metropolitan Oﬃce of Education (https://kess.kedi.re.kr), the portion of international students was
0.42% (elementary), 0.15% (middle school) and 0.09% (high school) among all students in Seoul. Moreover,
most of them attend separate international schools for foreigners. Thus, the portion of international students
in general schools is close to zero.

14

as well as university entrance exams in Seoul in 2015. According to a recent survey

(The Private Education Expenses Survey) by Statistics Korea, 75% of students in Seoul

(more than a million) attended private academies after school.18 As a result, private

academies also become another important factor aﬀecting the educational environment.

E. Air Quality and House Supply

This paper additionally assembles information about the air quality around a house.

Many recent studies show that households are willing to pay to avoid air pollution.

According to Bajari, Fruehwirth, Kim, and Timmins (2012), a home buyer would be

willing to pay 0.36% of the housing price to avoid 1 µg/m3 increase in PM10 (particles

less than 10 micrometers in diameter) concentration.19 This paper uses the average

annual PM10 concentration (µg/m3) statistic, which has recently been classiﬁed as

Category 1 carcinogen by the World Health Organization (WHO) and is also known

to cause various health problems such as asthma, bronchial pneumonia, and heart

attacks. The PM10 concentration is measured at 25 monitors in Seoul.20 Additionally

note that this study also controls for the number of new housing units to consider the

supply side of the housing market. To be speciﬁc, each model controls for the number

of similar sized apartment units within a certain distance (0.5 km, 1 km, and 2 km)

from a particular house.

1.2.1 Data Matching

For the estimation of the implicit prices of housing attributes, all the assembled data

should be matched with an individual house. For the matching process, unlike most

of the previous literature, this study uses spatial weight matrices used in a recent GIS

analysis. When the location of a house is identiﬁed by census tract level resolution,

researchers usually match the amenity information to a house based on the census tract

level (this study deﬁnes this method as the simple matching method). Then, identical

18 Also, according to the survey, on average, a student spends $ 5,180 for the private education in a year.
19 Sieg, Smith, Banzhaf, and Walsh (2004), Chay and Greenstone (2005), Bayer, Keohane, and Timmins
(2009), Tra (2010), Grainger (2012), Bento, Freedman, and Lang (2015), and Hamilton and Phaneuf (2015)
examine this topic.

20 The Korea Environment Corporation

publishes

this

data

and

it

is

available

at

https://www.airkorea.or.kr

15

information is matched with all the houses located in the same census tract. However,

if the speciﬁc location of a house is identiﬁed, better methods can be considered for

the matching technique.

To illustrate the problem of simple matching, we can suppose the following

example. If a house is located in “A” census tract but is close to the boundary line

of “B” census tract, then the amenities in “B” could also aﬀect the housing price in

“A.” However, this reality – people can take advantage of amenities near their house

irrelevant from census boundary – is ignored when the simple matching method is

adopted.21 To solve the problem of interrupting the space created artiﬁcially by census

tract boundaries, this study adopts various distance-weight matrices for the matching

between surrounding amenity information and a house. This method uses amenity

information by the census tract level while merging the information with a house by

considering distances. That is, amenities near a house have larger weights than those

far away from the house. The neighborhood and amenity information (variables in

Table 1.1 – (C) and Table A1 – (D)) is matched with a house using this method.

This paper adopts diﬀerent distance-weight matrices for the matching between

school information and a house to consider school zone and the student allocation

systems of Seoul. Note that one-to-one mapping does not exist between a residential

location and a school in Seoul, since there are multiple middle and high schools in the

same school zone.22 The proximity of a residence to a school and students’ preference

are primarily considered in the student allocation process; however, there is no guar-

antee that students are assigned to the school nearest to their houses. To consider the

above properties, this paper uses the distance-weight method for matching between a

school and a house. By using this method, a school near a house has a larger weight

than those far away from the house in the data matching process. There are several

reasons for justifying the use of this method. First, even though uncertainties exist in

the student allocation process, it is more likely students are allocated to the nearest

21

I could ﬁnd some extreme cases that reveal this problem. For instance, according to the Business Census
2015, there were 817 private academies in the Dachi census area while there were 37 in the Yeoksam census
area. The distance between the centroid of both areas is less than 1 km. Accordingly, people in the Yeoksam
area can beneﬁt from facilities located in the Daech area, and those facilities may also aﬀect the housing price
in the Yeoksam area.

22

In the case of an elementary school, however, there exists one-to-one matching according to the address

of a house.

16

school since the allocation system considers the distances from a house to schools. Sec-

ond, when considering students’ commute cost, the performance of a school far away

from a house should be discounted compared to a nearby school with the same per-

formance statistics. Third, when households decide where to live, they consider both

the possibility of being allocated to the school where they want their child to attend

and the ease of commute. Thus, they tend to live near a school where they want to

apply, or to be assigned while considering the performances of schools near the house.

This paper argues the school information is appropriately matched with a house while

considering the properties of the school zone and student allocation systems by using

the distance-weight matching method. More details about school zone and the stu-

dent allocation systems along with the matching methods are explained in the data

appendix.

For the weights, this study adopts the inverse-squared distance-weight matrices

using the distances from a house to i) centroids of census tracts (for the matching of

neighborhood demographics and amenity information), ii) schools (for the matching of

school information)23, and iii) air quality measuring monitors. For robustness checks,

this study also tries other weight matrices with some cut-oﬀ radius distances for the

matching of neighborhood and amenity information surrounding a house, and the N-

nearest information matching methods for school statistics.24

1.2.2 Panel Dataset Generating Process

As discussed in the introduction, the problem of housing panel dataset is that it in-

evitably covers the long-term to have a suﬃcient sample size because repeated transac-

tions for the same house are rare events. Using the long-term panel dataset, however,

makes the panel regression (using mean-diﬀerenced or ﬁrst-diﬀerenced models) less re-

liable because the analysis is based on the unreasonable assumption that unobservables

or omitted attributes of a house – which cause the bias estimation – stay ﬁxed for the

23 There are two types of weight matrices (one for high schools and another weight matrix for middle

schools). The distance weight matrices are changed over time due to a new establishment of a school.

24 This paper additionally tried 1 km and 2 km radius cut-oﬀs from an individual house for generating
the distance-weight matrix. For the matching of school statistics, the three-nearest school matching method
is adopted for distance-weights along with the overall weights using statistics of all schools within the same
school zone.

17

long-period that the panel dataset covers.

To overcome the above problem, this study generates a short-term panel dataset

by using the apartments’ characteristic that an apartment building includes multiple

and almost identical housing units in the same building. Given this property, it is

reasonable to assume a group of equal sized housing units in the same apartment

complex as identical houses. In this study, if apartments have the same address, size,

construction year, and are located in the same ﬂoor group, they are treated as the same

housing unit in the panel dataset.25 A critical point for using this method to generate

a short-term panel dataset is that it makes much more frequent repeated transactions

in a given period than using the repeated transactions of exactly the same house, since

multiple housing units are included in the same group.

As a possible problem of this method, there could be some diﬀerences within an

identical group of apartments owing to renovations and interiors. However, the same

issues happen even when we use the panel dataset using the repeated transactions of

the same single detached house because such renovations and interiors are also mainly

hidden information for researchers. On the other hand, there are some advantages of

using this method. First, apartments have more restrictions than the single detached

house for renovations because a unit is a part of the entire building. For instance,

structural renovations such as the extension of house size or making more rooms are not

possible due to safety restrictions. Also, enhancement in curb appeal could be another

critical change in the unobservable attribute of a single detached house; however, this

is not the case with apartments. Put diﬀerently, apartments have less possibilities for

some fundamental changes in unobservable attributes that make the panel analysis less

reliable compared to a single detached house.

This study also uses diﬀerent numbers of transactions for each group of apart-

ments as the sample weight, and the diﬀerent weights act as an additional control to

mitigate the possible diﬀerences in unobservable attributes within a group of apart-

ments. For example, “A” group of apartments has 30 transactions while “B” group of

apartments has three transactions within a year. Then, the weight of observation “A”
25 A ﬂoor is divided into ﬁve groups. Group 1 includes the ﬁrst ﬂoor and lower. Group 2: ﬂoors 2∼5,
Group 3: ﬂoors 6∼10, Group 4: ﬂoors 11∼20, Group 5: ﬂoors higher than 20. Also, this paper also tried
other group compositions; however, there were no meaningful diﬀerences in the results.

18

is ten times larger than the weight of observation “B” in the panel regression analysis.

Since some unobservables – such as renovation, interiors, and bargaining power between

a buyer and a seller – can be diﬀerent within the same group of apartment transac-

tions, it is reasonable to assume that an observation generated from more frequent

transactions reveals a more accurate price of the group of apartments. In this respect,

the diﬀerent number of transactions are used as weights to consider the accurate price

of a group of apartments.26

It is important to note the limitation of this method to generate a panel dataset.

This method can be justiﬁed only when the apartment transaction samples represent

the whole housing market of the area. If the portion of apartments among entire houses

is low, and if the implicit price of housing attributes vary depending on the house type,

there may be a sample bias problem. For instance, we could expect that a household

with children would have a higher willingness to pay for better educational environ-

ments than a household without children. According to the Housing and Population

Census 2015, due to the high population density, less than 15% of households live in

a single detached house in Seoul and almost 90% of housing transactions come from

apartments.27 Also, the portion of families with children among apartment residences

(74%) is similar to that of single-detached house residences (67%). Based on the above

statistics, this study assumes that apartments can represent the entire housing market

of Seoul.

1.3 Hedonic Price Regressions

In this section, this paper uses a hedonic price regression to investigate education

and neighborhood premiums in housing prices. For the analysis, the following base-

line model is used to estimate the implicit prices for educational and neighborhood

environments:

log(price)ijt = WitXitβt + Ziφt + Dististδt + Djt + ξit

(1)

26

In the process for generating the price of a group of apartments in a speciﬁc year, this study controls
monthly-regional ﬁxed eﬀect to consider monthly price ﬂuctuation by each region. For more technical details,
refer to the data appendix.

27 The population density of Seoul (16,204/km2) is higher than New York City (10,725/km2)

19

where log(price)ijt is the log of the transaction price of house i in a school zone j at time

t. Xit includes time-varying variables such as amenities, neighborhood demographics,

various education statistics, air pollution observations around a house i, and the num-

ber of new housing units within a certain radius from a house i at time t. Wit is the

spatial weight matrix and it has diﬀerent forms for amenity information (W a
statistics (W s,m

it), school
it). Zi
includes time-constant variables such as physical attributes of a house (size, ﬂoor, age,

(high school)), and air pollution levels (W p

(middle school), W s,h
it

it

brand, and type of house) and distances to a ﬁxed location.28 Distist includes various
distances from house i to location s at time t.29 Djt represents the school zone ﬁxed-

eﬀect (11 school zones by monthly level), thereby this study ﬂexibly allows the regional

ﬁxed-eﬀect to vary separately over time by each school zone. The role of school zone

ﬁxed-eﬀect is to control housing price ﬂuctuations by macro factors – speculation and

asset investment decision by forecasting future prices – that are irrelevant to changes

in housing attributes. Lastly, ξit denotes the omitted or unobservable attributes of the

house i at time t.

By using two consecutive years of data and additionally assuming that the un-

observable attributes of a house and the implicit prices for time-varying variables do
not change over two years (βt(cid:48) = βt = β, δt(cid:48) = δt = δ, ξit(cid:48) = ξit, (t(cid:48) > t)), the equation
(1) can be re-written as the following model by ﬁrst diﬀerencing:

House Fixed-Eﬀect Model30

∆log(price)ij = (∆WiXi) β + (∆Distis) δ + Zi (φt(cid:48) − φt) + ∆Dj

(2)

This study also ﬂexibly allows the unobservables to vary within the two consecutive

years given the assumption that it follows the ﬁrst-order Markov process:

ξit(cid:48) = γ ξit + ηit(cid:48)

(t(cid:48) > t)

(3)

28 Straight distance from a house i to the Han-river.
29

s ∈ {the nearest school (high, middle, and elementary school), subway station, department store,

30 Note that the coeﬃcient of time constant variable (Zi) shows the estimated change in the implicit price:

supermarket}
φt(cid:48) − φt

20

here, γξit is the expected value of the unobservable attributes at time t(cid:48), and ηit(cid:48) is
the stochastic innovation in the unobservable attributes. Note that this study assumes
E[ηit(cid:48)|It]= 0, where It denotes the available information at time t. This assumption
is based on the eﬃciency house market by Case and Shiller (1989) implying that it is

impossible to earn excess returns by using the available information at time t in the

housing market. Then, equation (1) can be re-written as follows by using the Markov

process assumption. For the Time-varying Unobservable Model, this paper closely fol-

lows the method developed by Bajari, Fruehwirth, Kim, and Timmins (2012).

Time-varying Unobservable Model

log(price)ijt(cid:48) = Wit(cid:48)Xit(cid:48)β + Ziφt(cid:48) + Distist(cid:48)δ + Djt(cid:48) + ξit(cid:48)

(4)

= Wit(cid:48)Xit(cid:48)β + Ziφt(cid:48) + Distist(cid:48)δ + Djt(cid:48) + γξit + ηit(cid:48)

= Wit(cid:48)Xit(cid:48)β + Ziφt(cid:48) + Distist(cid:48)δ + Djt(cid:48)

+ γ[log(price)ijt − WitXitβ − Ziφt − Dististδ − Djt] + ηit(cid:48)

= (Djt(cid:48) − γDjt) + γlog(price)ijt + [Wit(cid:48)Xit(cid:48) − γWitXit]β
+ [Distist(cid:48) − γDistist]δt + Zi(φt(cid:48) − γφt) + ηit(cid:48)

This paper also considers the case that Xit(cid:48) could be correlated with ηit(cid:48) and use

two-stage nonlinear least squares (2SNLS) to recover parameters in equation (4) under

the above assumptions. In the ﬁrst stage, all the exogenous variables and predetermined

variables are used as instruments to replace Xit(cid:48) with its projected value. Intuitively,

the above Time-varying Unobservable Model uses the housing price information of

the previous period to impute the value of unobservable attributes of the house. For

example, if the housing price of i is unusually high after controlling the variables (Xi,

Zi, Distis, Dj) in the model, it means the value of unobservable attributes of the

house ξi – observed by the buyer of the house – is high. From this, the value of the

unobservable attributes is used as additional information to analyze the housing price

of the next period.

By using the above models, this paper analyzes the long-term reputation eﬀect

of school performance on household implicit price in Section 1.3.1, and Section 1.3.2

21

examines the implicit price for better educational environments and shows how much

of housing price gaps within a city are explained by the diﬀerences in educational

environments.

1.3.1 School Reputation Eﬀect

People naturally consider both the past records of a school and the most recent infor-

mation when assessing the quality of a school. In this respect, however, I could not ﬁnd

any previous studies that explicitly consider the ‘reputation eﬀect,’ which comes from

past records of a school. To the best of my knowledge, all previous research analyzing

the willingness to pay for a better educational quality use a one-year statistic or aver-
aged statistics over a short-term (e.g., averaged over 2∼4 years), not to consider the
reputation eﬀect, but to reduce any year-to-year noise in the school quality variable.31

In this section, I show that there exists a long-lasting reputation eﬀect in willing-

ness to pay for the better performance of a school, and also that estimating an implicit

price by using an one-year or averaged statistics over the short-term can be biased.

Suppose a model uses a three-year averaged test score of a school as the variable of

interest. In this case, the model implicitly assumes the implicit prices are all the same

for the test scores of the most recent three years.32 In practice, however, people can

consider the history of school performance longer than three years while placing more

weight on the recent results than the past results.33 As such, the estimated model

should include longer past records than three years without restriction on coeﬃcients

for unbiased estimations.

Table 1.2 shows various results about the school reputation eﬀect. All models
(column (1)∼(12)) include all the other variables listed in the summary Table 1.1 and
A1 in addition to the top university entrance rate for the past ten years represented in

31

2-year average: Bayer, Ferreira, and McMillan (2007). 3-year average: Gibbons and Machin (2003),
Kane, Riegg, and Staiger (2006), Clapp, Nanda, and Ross (2008). 4-year average: Reback (2005), and most
other research use merely one-year statistics.

32 Let’s suppose the true model is: Log(P rice)t = βtSt + βt−1St−1 + βt−2St−2 + δXt (St denotes test
score of a school at time t, and Xt contains other control variables), and the ﬁtted model is: Log(P rice)t =
α · ASt + δXt (here, ASt =
, the three-year averaged test score) Then, α = β ∗ 3 under the
3
restriction that βt = βt−1 = βt−2 = β

St + St−1 + St−2

33

In the equation, if βt > βt−2 (people think the recent record (St) is more important than the past

record (St−3)), and if the true model includes until St−i (i > 3), the estimated α should be biased.

22

Table 1.2 School Reputation Eﬀect

Dependent Variable = Log(Price)

Panel A : Including all previous (10 years) results of schools

Variable

Cross Section

(1)

(2)

House Fixed

(3)

(4)

Time-varying

(5)

(6)

Top Univ.
Entrance Rate (t)

8.370***
(0.597)

7.232***
(0.675)

2.068***
(0.314)

1.528***
(0.363)

2.099***
(0.331)

1.503***
(0.273)

(t − 1)

(t − 2)

(t − 3)

(t − 4)

(t − 5)

(t − 6)

(t − 7)

(t − 8)

(t − 9)

Sum(t ∼ t − 7)

Sum(t ∼ t − 9)

γ

W/ All lags
Obs.
Adj R

0.0311
(0.621)

-1.601*
(0.748)

3.270***
(0.301)

1.928***
(0.322)

3.693***
(0.318)

2.159***
(0.349)

-2.999***

-4.138***

(0.628)

(0.747)

2.040***
(0.265)

1.908***
(0.363)

2.577***
(0.252)

2.021***
(0.259)

3.040***
(0.530)

0.663
(0.637)

2.080***
(0.240)

1.596***
(0.326)

2.661***
(0.274)

1.812***
(0.314)

5.312***
(0.518)

7.772***
(0.594)

1.229***
(0.219)

1.728***
(0.281)

1.856***
(0.255)

1.63***
(0.359)

2.954***
(0.605)

6.173***
(0.645)

1.316***
(0.315)

1.297***
(0.340)

1.690***
(0.318)

1.674***
(0.292)

-2.156***

-3.393***

(0.556)

(0.617)

1.239***
(0.245)

-4.193***

-4.587***

(0.534)

(0.616)

-2.965***

-3.028***

(0.428)

(0.507)

0.479*
(0.194)

0.635**
(0.207)

2.525***
(0.502)

2.396***
(0.552)

0.786***
(0.177)

1.180**
(0.372)

0.801**
(0.246)

0.157
(0.276)

0.662**
(0.226)

1.499***
(0.237)

1.016***
(0.345)

0.814***
(0.198)

0.727***
(0.224)

0.590**
(0.223)

0.331
(0.263)

0.861***
(0.184)

1.069***
(0.215)

9.920***

7.487***

13.720***

11.965***

16.888***

12.541***

(1.089)

(1.093)

(1.473)

(1.754)

(1.567)

(1.643)

10.360***

8.120***

15.141***

12.784***

18.339***

13.942***

(1.345)

(1.343)

(1.632)

(1.971)

(1.761)

(1.874)

144,492

0.872

Yes

144,492

0.886

72,246
0.183

Yes

72,246
0.230

72,246
0.991

Yes

72,246
0.993

0.961***
(0.003)

0.976***
(0.002)

Panel B : Including previous 3-year averaged results of schools

Variable

Cross Section

(7)

(8)

Top Univ. Entrance
3-Year Average

9.358***

8.366***

(0.964)

(0.981)

W/ All lags
Obs.
Adj R2

144,492

0.869

Yes

144,492

0.884

House Fixed

Time-varying

(9)

2.259***

(0.518)

72,246
0.174

(10)

0.832
(0.688)

Yes

72,246
0.225

(11)

2.562***

(0.824)

72,246
0.987

(12)

0.914
(0.589)

Yes

72,246
0.993

Standard errors clustered at IDs are in parentheses

*** p < 0.01, ** p < 0.05, * p < 0.1

Note: All coeﬃcients are reported as β*100. All models include ﬁxed eﬀects (11 school zones

by monthly level) and the other control variables listed in Table 1.1 and A1 in the model.

23

Table 1.2. Columns (3), (4), (9), and (11) represent the results from the House Fixed-

Eﬀect Model using the equation (2), and columns (5), (6), (11), and (12) show the

estimation results from the equation (4), the Time-varying Unobservable Model. Each

column in Panel A is the result when the model includes the past records of a school

respectively until ten years ago, while Panel B shows the results when each model

includes only the three-year averaged statistics of a school instead of ten variables

represented in the Panel A. As other explanatory variables, regression models use the

same variables for the comparison between Panel A and B. All models in columns (1)

to (12) also control monthly time ﬁxed eﬀects by 11 school zones.

According to the results, in the cross-section model ((1) and (2)), positive and

negative coeﬃcients are mixed, and it is hard to ﬁnd any patterns among the implicit

prices of the past records. However, in the results from the House Fixed-Eﬀect Model

((3) and (4)) and the Time-varying Unobservable Model ((5) and (6)), all coeﬃcients

show a reasonable positive sign, and the size of the coeﬃcients becomes smaller as

time passes. These results can be interpreted as people evaluating a school with higher

weight placed on recent results, while also taking into account the past records. It is

important to note that the implicit price of the top university entrance rate for each

year is relatively small, but the accumulated value for eight years and ten years becomes

over 12.5% of a housing price according to the most conservative results (column (6)).

This result implies people do not respond instantly to a transient, one-time performance

record of a school; however, they have a large amount of willingness to pay for long-

term, stably accumulated results. That is, the reputation of a school has a signiﬁcant

eﬀect on people’s willingness to pay for better school performance.

A comparison of the cross-section model with the other two models suggests that

although the cross-section model controls for the monthly time ﬁxed eﬀects by each 11

school zones – and the adjusted R2 almost reaches 0.9 – the model still suﬀers from

the endogeneity problem. If we look at the results from the House Fixed-Eﬀect Model
and the Time-varying Unobservable Model (column (3)∼(6)), the model in columns
(3) and (5) includes the top university entrance rate variables for the last ten years

and other explanatory variables only at time t. However, one could expect signiﬁcant

correlations between the past records of top university entrance rate and the past

24

amenity information around a house to cause a biased estimation. To control this

problem, the models (2), (4), and (6) in Panel A, and (8), (10), and (12) in Panel B
include the equivalent time lengths (from t to t − 9) of other explanatory variables in
the model in addition to the top university entrance rate variables of the past ten years.

If we compare the results of column (4) with (3), and (6) with (5), the cumulative value

of coeﬃcients for eight years decreases 13% (House Fixed-Eﬀect Model, from 13.72 to

11.97) and 26% (Time-varying Unobservable Model, from 16.89 to 12.54) respectively.

This result implies that positive correlations exist between school performance and

good amenities around a house in the past records.

Note that results from both models – House Fixed-Eﬀect Model and Time-

varying Unobservable Model – are similar. The estimated γ in equation (4), which

comes from the Markov process assumption (equation (3)), is 0.976 in the model (6).

First, this result means that the unobservables and omitted attributes of a house stay at

almost the same level during the two-year period, given the Markov process assumption

about unobservables. Second, it also veriﬁes this paper’s expectation that it is more

likely that unobservables stay at the same level if a housing panel covers the short-

term. Note that if the γ is equal to one, the Time-varying Unobservable Model is

converted into the House Fixed-Eﬀect Model while justifying the use of a mean- or

ﬁrst-diﬀerenced model. Because the estimated γ is close to one, both models presented

in Table 1.3 have similar results. However, the estimated γ – statistically diﬀerent from

one34 – also provides a reason to use the Time-varying Unobservable Model instead of

the House Fixed-Eﬀect Model for unbiased estimations and more robust results.

Panel B shows the results when the model includes the recent three-year average

top university entrance rate instead of its records for the last ten years respectively.

When we use a cross-sectional model, if most of the schools’ previous records have been

stable – a statistically high correlation among the past records of a school – then the

estimation bias may not be severe even though it does not consider the reputation eﬀect.

However, a cross-sectional model is not free from omitted variable bias. On the other

hand, if we use a diﬀerenced model to control time-invariant omitted variables while

using the short-term averaged school performance statistics to estimate its implicit

34 The estimate is statistically diﬀerent from 1 even under 0.1% signiﬁcance level.

25

price, the bias can become severe because of the diﬀerencing. The intuitive reason is

that the diﬀerencing used to delete time-invariant omitted variables can also eliminate

the reputation eﬀects that come from all previous school performance results. For

instance, if a recent three-year averaged test score is used in the diﬀerenced model, the

diﬀerencing deletes a large portion of the reputation eﬀects coming from the test score

more than four years ago. If we compare the results in the House Fixed-Eﬀect Model

(column (3) and (9)), the estimated coeﬃcient decreases 84% (from 13.72 to 2.26) and

85% (from 16.89 to 2.56) in the Time-varying Unobservable Model (column (5) and

(11)). In case of the boundary discontinuity design (BDD) based on a cross-sectional

model – which controls omitted variable bias by including boundary ﬁxed-eﬀects and

other controls in the region – the bias can be trivial, since the correlation among

past performance results of a school is usually high. Even in this case, however, this

study suggests considering a long-term performance history of a school to obtain more

accurate estimates for the implicit prices associated with better school performance.

As robustness checks, this study uses the regional ﬁxed-eﬀect for smaller area (25-

district) instead of 11 school zone level ﬁxed-eﬀect (Table A2, Table A5, and Table A6).

Also, I try diﬀerent distance-weight matrices for the matching of amenity information

and school statistics with a house. To be speciﬁc, this study adopts the distance-

weight matrices with 1 km and 2 km radius cut-oﬀs for the matching between amenity

information and a house instead of using the matrices without radius cut-oﬀs (columns

(1) and (2) in Table A3 and Table A4).35 Also, for the matching between school

statistics and a house, this paper tries another matching matrix using the information

of the nearest three schools (three schools for each middle and high school) instead of

using all schools in the school zone where the house is located (columns (3) and (4)

in Table A3 and Table A4). As a distance-weight, the same metric – inverse squared

distance – is used. According to the results, the primary results, long-lasting reputation

eﬀects and more weight on recent results, are maintained although there are diﬀerences

in the estimated values of accumulated coeﬃcients.36

35

I do not report the results from the model using the 2 km radius cut-oﬀ distance-weight matrix because

the results are quite similar to the presented results from using a 1 km radius cut-oﬀ matrix.

36 Columns (5) and (6) in Table A3 and Table A4 use both a 1 km radius cut-oﬀ for amenity information

matching and the nearest three schools statistics for school information matching with a house.

26

When the Time-varying Unobservable Model with the three-nearest school match-

ing method (column (4) in Table A4) is adopted, the estimated coeﬃcient decreases

to 10.058 compared to the result in column (6) in Table 1.2. Note that the coeﬃcient

decreases by 20% (from 12.541 to 10.058), however, the three-nearest school matching

method also increases the gaps in top university entrance rate between the Gangnam

and the other areas in Seoul by 19%. As a result, the diﬀerence in the school statistics

matching methods does not have a signiﬁcant eﬀect on housing price gaps caused by

the diﬀerences in top university entrance rate results. This paper explains the details

about the housing price gap decomposition in the next section.

1.3.2 Education Environment and City Segregation

This section analyzes the implicit prices of better educational environments and shows

how much the housing price gaps within a city are explained by the diﬀerences in

educational environments.

Table 1.3 shows the estimated implicit prices and calculated willingness to pay

for better educational environments in addition to the top university entrance rate

from the model (4) (House Fixed-Eﬀect Model) and the model (6) (Time-Varying

Unobservable Model) in Table 1.2. Note that given the estimated γ is close to one, the

results from both models are similar. In discussing the results, we focus on the Time-

varying Unobservable Model. The ﬁrst column shows the estimated coeﬃcients, and

the second column represents the calculated willingness to pay for a one-unit increase

of each variable per year. For the calculation, the mean housing price ($ 438,656) is

used and annualized at the rate of 7%.

According to the results, an average household has the highest willingness to

pay for the one-unit (1%p) higher top university entrance rate among the presented

variables, and distances from a house to schools also have substantial negative implicit

prices per unit (1 km). Next, the results of WTP for the one standard deviation

increase show more realistic ﬁgures to compare relative importance among variables.

The top university entrance rate is still the most critical attribute among all of the

educational environment variables in the model. To live around a school which has one-

SD (0.71%p) higher entrance rate to a top university, an average household is willing

27

Table 1.3 Willingness to Pay for Educational Environments

Dependent Variable = log(price)

House Fixed-Eﬀect Model

Time Varying Unobservable

VARIABLE

Coeﬃcient

WTP ($)

(1 unit)

(1 SD)

Coeﬃcient

WTP ($)

(1 unit)

(1 SD)

Top Univ
Entrance Rate

N-SAT
Score

Top High School
Entrance Rate

Private High School
Entrance Rate

Neighborhood
% BA or Above

Private
Academy

Student
per Class

Student
per Teacher

% Regual
Teacher

Distance to
Secondary School

Distance
Elementary School

PM10
Air Pollution

11.965***

3,903

2,781

12.541***

4,103

2,924

0.017

0.035

0.050

0.167

0.114

5

11

15

51

35

50

10

73

590

757

0.053**

0.062

0.061

16

19

19

159

17

89

0.338**

104

1,200

0.103**

32

683

-0.844***

- 258

- 596

-0.625***

- 191

- 442

-0.423***

- 130

- 174

-1.233***

- 376

- 504

0.011

3

13

0.073

23

89

-2.85**

- 859

- 377

-5.32**

- 1,560

- 247

-0.833***

- 255

- 292

-0.918***

- 281

- 321

All coeﬃcients are report as β*100.

Willingness to pay is annualized at rate of 7% for mean house price of $ 438,656

*** p < 0.01, ** p < 0.05, * p < 0.1

28

to pay $ 2,924 per a year. The portion of the highly educated in a neighborhood, the

number of private academies (as a proxy for a better environment for private educations

after school) around a house, and distances from a house to schools also become critical

attributes aﬀecting housing prices.

Recently, rapid air quality deterioration in Seoul has caused many people to

become interested in PM10 (particles less than 10 micrometers in diameter) concen-

tration. Note that, according to the estimation result, a home buyer would be willing

to pay 0.9% of the housing price to avoid 1µg/m3 increase in PM10 concentration.

This estimated value is higher than the estimates from Bayer, Fruehwirth, Kim, and
Timmins (2012). They estimate the (negative) implicit price of PM10 as 0.36∼0.58%
of a housing price. To estimate the price of air pollution, they analyze the housing

market of California’s Bay Area. The critical diﬀerence is that the average annual

PM10 concentration in Seoul (47.5µg/m3) during the sample period is much higher

than California’s Bay Area (23µg/m3) in their study. Due to the increasing marginal

disutility of non-goods such as air pollution, the higher average level of air pollution

could increase households’ willingness to pay to avoid the one unit increase in air

pollution.

In contrast to the previous literature, in this study, the test score (N-SAT) of a

school only has a marginal eﬀect on housing prices. However, when the top university

entrance rate variables were excluded from a model, the N-SAT score became the most

critical variable. Given the meaning of a coeﬃcient in a hedonic price regression –

the eﬀect of a one unit increase of a variable on the market price while holding the

other attributes ﬁxed – the above results can be interpreted as a higher N-SAT score

not being meaningful if the top university entrance rate is the same. Also note that,

as with the performance of a middle school, the estimated implicit price of top high

school entrance rate is not signiﬁcant, and the private high school entrance rate only

has a marginal implicit price.

Next, we investigate the city segregation and the eﬀect of educational environ-

ments on housing price gaps within a city. As a geographical summary, see Figure

1.1 and 1.2. The ﬁgure shows the overall segregation status in school performances,

neighborhood composition, and housing price on the map of Seoul. Note that the

29

Table 1.4 Housing Price Gap Decomposition

Variable

Gangnam The other

Coeﬃcient

Decomposition Diﬀerence(Mean)

Mean value of Variable

($) Price Gap

% of

Top Univ
Entrance Rate

N-SAT
Score

Top High School
Entrance Rate

Private High School
Entrance Rate

Neighborhood
% BA or Above

Private
Academy

Student
per Class

Student
per Teacher

% Full-time
Teacher

Distance to
Secondary School

Distance
Elementary School

Mean Housing Price

(900f t2)

2.02

0.60

12.541***

101,269

29.40%

109.45

100.07

0.053**

2,833

2.71

2.09

12.83

8.53

0.062

0.061

220

1,487

0.82%

0.06%

0.43%

70.01

43.44

0.338**

51,100

14.83%

50.61

32.74

0.103**

10,441

3.03%

33.76

31.40

-0.625***

- 8,362

-2.43%

15.85

14.47

-1.233***

- 9,707

-2.82%

84.44

86.36

0.073

- 802

-0.23%

1.01

1.13

-2.85**

1,820

0.53%

0.33

0.32

-5.32**

- 443

-0.13%

$757,076

$412,610

SUM

149,855

43.5%

All coeﬃcients are report as β*100.
*** p < 0.01, ** p < 0.05, * p < 0.1

better performing schools, well-educated people, and more expensive houses are highly

concentrated in the Gangnam school zone area.

Numerically, the ﬁrst two columns in Table 1.4 show the diﬀerences in average

statistics between the two areas. The Gangnam area has 1.4%p higher top university

entrance rate and 9%p higher N-SAT score than the other area.

In neighborhood

composition, there is a 25%p gap in the portion of BA or above degree holders in the
40∼49 age group between the two areas. Also note that, on average, there are 18 more
private academies around a house in the Gangnam school zone; however, there are also

more students per class and teacher in Gangnam school zone compared to the other

area. Last, there is little diﬀerence in distances from a house to schools.37

37 The summary Table 1.1 and Table A1 compare all variables between the Gangnam and the other area.

30

Table 1.4 also shows the housing price gap decomposition resulting from diﬀer-

ences in educational environments between the Gangnam school zone and other areas

in Seoul. For the calculation of the price gap decomposition, this study uses the es-

timated coeﬃcients from the model (6) in Table 1.2 and uses the mean housing price

of mean-size apartments (900f t2) in each region. The right two columns show the

number of price gaps explained by each attribute and the ratio of it to the total av-

erage housing price gap.38 Note that the diﬀerence in top university entrance rates

explains almost 30% of the housing price gap. The neighborhood composition and

the number of private academies also account for 14.83% and 3.03% of the housing

price gap respectively. In summation, almost 50% of the housing price gap comes from

diﬀerences in the above three components between the two areas. On the other hand,

more students per class and teacher in the Gangnam school zone diminish the housing

price gap by 5.25%. Other attributes have relatively marginal eﬀects on the price gap,

and overall, the educational and neighborhood environments account for 43.5% of the

housing price gap.

If the Time-varying Unobservable Model uses the three-nearest school matching

method for the matching of school statistics (column (4) in Table A4), the estimated

coeﬃcient of the top university entrance rate decreases by 20% (from 12.541 to 10.058).

However, note that the gaps in the top university entrance rate between the Gangnam

and the other areas increase by 19% from 1.42 (= 2.02 – 0.60) to 1.69 (= 2.2 – 0.51)

when the three-nearest school matching is adopted. Thus, the top university entrance

rate still explains 28.27% of the mean housing price gap between the two areas un-

der the diﬀerent matching method. As a robustness check, this paper also adopts the

regional ﬁxed-eﬀect by smaller areas (25 district level) instead of 11 school zone level

for the housing price gap decomposition. Table A7 and A8 show the willingness to

pay for better educational environments and the results of the housing price gap de-

composition when models use a 25-district-level ﬁxed eﬀect. According to the results,

the estimated implicit prices of primary variables decrease; however, the diﬀerences

in educational environments still explain the signiﬁcant portion (34%) of the housing

38 Because the dependent variable is log(price), we calculate the portion of the price gap explained by

each variable as follows:

, here the subscript G denotes the Gangnam school

β ∗ (XG − XN G)

log(P RICEG) − log(P RICEN G)

zone and NG means the other areas)

31

price gaps between the Gangnam and the other areas.

1.4 Policy Simulation

As discussed in the above section, educational environments have a signiﬁcant eﬀect on

housing prices based on households’ substantial willingness to pay for them. Accord-

ingly, educational policies such as school desegregation by changing school catchment

areas and student allocation systems inevitably aﬀect housing prices in the area where

the new policy is applied. Hence, it is impossible and undesirable to separate out an

urban housing policy from such educational policies. Indeed, the government of Seoul

is currently considering various educational policies such as school zone adjustment and

changes in student allocation methods to mitigate severe segregations in both housing

prices and school performance gaps within the city.

In this regard, we can refer to many precedents and studies on this topic. In

most cases, the purpose of the educational policies was not originally intended to

aﬀect housing prices; however, it turned out to signiﬁcantly aﬀect housing prices in the

region. Baum-Snow and Lutz (2011) analyze the public-school desegregation policy in

US cities and its eﬀect on residential location choice by race. Machin and Salvanes

(2016) investigate the change in the student allocation system of Oslo, Norway in 1997

and show its eﬀect on housing prices. Boustan (2012) examines the changes in housing

price gaps in US metropolitan areas with and without public school desegregation

policies. Ries and Somerville (2010) use school zone adjustment in Vancouver as a

quasi-experiment to measure housing price capitalization of school quality.

In this section, as the policy simulation, this paper changes the school zone

boundaries and forecasts what would happen in the new school zone area. In the case

of educational policy, such as school zone and student allocation changes, there should

be households who support the new policy and those who oppose it. Therefore, it

might be meaningful if we can predict the result of the vote for the new policy and

calculate households’ willingness to pay for the policy change. The expected number

of people in favor of the new policy and their estimated willingness to pay can provide

a valuable criterion for evaluating the superiority of policies.

Figure 1.3 presents the school zone change simulation. The Gangnam school zone

32

(a) Current

(b) Simulation

Figure 1.3 School Zone Adjustment

(shaded areas in the left ﬁgure) currently includes the Gangnam and Seocho districts.

This study changes the school zone boundaries as in Figure 1.3b, which includes the

Seocho, Dongjak, and Gwanak districts in the new school zone. That is, the new

school zone policy is designed to tie up adjacent areas with signiﬁcant gaps in the top

university entrance rate.39 Schools in the Dongjak and Gwanak districts have lower top

university entrance rates compared to the schools in the Seocho district. School zone

adjustment will aﬀect the matching between a house and schools and the distances from

a house to schools in the same school zone immediately after the change. For example,

suppose house “A” is located in the Seocho district and the nearest high school is “B,”

however, the school is located in the Dongjak district, which is located in the other

school zone. In this situation, the house “A” is not aﬀected by the performance of high

school “B” or the distance to get there under the current school zone system. However,

after the school zone is adjusted like the above simulation, the high school “B” has the

largest eﬀect on house “A” of all the high schools.

In the long-term, diﬀerent student composition – students coming from diﬀerent

regions – can aﬀect the performance of schools and consequently aﬀect the neighbor-

hood composition due to a household self-segregation according to school performance.

For these reasons, it is expected that the performance of schools and the neighborhood

39

In this case, the Gangnam district can form another school zone with other districts such as Seongdong

or Gwangjin.

33

composition could vary in the new school zone over time.

Accordingly, the simulation considers two possible scenarios based on time. The

ﬁrst scenario considers the short-term situation immediately after the policy takes ef-

fect. In the short-term, the eﬀects come from the diﬀerences in matching a house to

a school under the new school zone while assuming that school performance statistics

and neighborhood composition stay at the same level. The second scenario allows the

possibility that school statistics and neighborhood composition can vary after the new

school zone system takes eﬀect in the region. This scenario can be considered as a

long-term situation. For the analysis of the long-term, we need virtual school per-

formance statistics and neighborhood composition after the new regime. To generate

those numbers, this paper uses the following method.

First, I calculate the average standard deviation of school performance (the top

university entrance rate) and neighborhood composition (% BA or above degree hold-
ers in the 40∼49 age group) within the same school zone from the current statistics.
Next, virtual values for each school and census block in the new school zone are gen-

erated based on the assumption that standard deviation of school performance and

neighborhood composition in the new school zone will converge to the average stan-

dard deviation of school zones in Seoul. To be speciﬁc, the standard deviation of the

top university entrance rate is 0.8% in the new school zone area; whereas, the average

standard deviation within the same school zone (calculated from 11 school zones) is

0.4%. In neighborhood demographics, the standard deviation in the portion of BA or
above degree holders in the 40∼49 age group is 19% in the new school zone, and that
value from the average of 11 school zones is 14%. This study generates the virtual

top university entrance rate and the portion of highly educated in a neighborhood by

using the initial and target standard deviations and the Z-value in the standard normal

distribution.40

This study also considers the sample weights by using the total housing popula-

tion data of the region to increase the accuracy of the policy simulation for the voting

experiment and estimating overall households’ WTP in the region. As discussed in the

40 That is, this paper assumes the standard deviations within the new school zone – 0.8% (top university
entrance rate) and 19% (neighborhood composition) – will converge to 0.4% and 14% respectively in the
long-term.

34

previous Section 1.2.2, transaction data of apartments is used to construct the short-

term panel dataset. However, the samples from the transactions are a part of the total

number of apartments in the policy area, and there are also other types of houses in

the area. Given that the portion of actual transactions among all of the housing units

can be diﬀerent depending on the region, one might expect that the transaction data

could not represent the whole households in the region due to sample bias.

To overcome this problem, I use Housing Census data and compare the number

of real transactions to the total number of apartment units by census tract level then

generate sample weights to complement the real transaction samples. Another problem

with the policy simulation is that other types of houses are in the region, such as

single detached houses. Note that all households should be included in the analysis to

precisely forecast the result of the vote for the new school zone policy. For this, the

simulation model also includes the transaction data of all types of houses along with

apartments and generates the sample weights by using the method as for apartments to

ﬁx the sample bias. The results, which do not consider the above sample weights, are

presented in Table A9 and A10 in the appendix. In conclusion, there is no signiﬁcant

diﬀerence in outcomes between the two methods.

Figure 1.4 shows the kernel density estimation of households’ willingness to pay

for the school zone adjustment in the region.41 The upper ﬁgure shows the WTP as the

percent of housing prices and the lower ﬁgure shows it as the amount of money. The

line and dashed graphs represent the short-term and long-term situations respectively.

In the short-term, again note that the eﬀect comes from the diﬀerent matching between

a house and a school as the school zone changes. Households in the lower score zone

area (Dongjak and Gwanak districts) beneﬁt from the new matching to schools with

better performances, and the opposite happens in the higher score zone area (Seocho

district).42 In the long-term, the gaps in WTP between two regions becomes broader

than in the initial short-term situation, since the simulation assumes that school per-

41 The Epanechnikov kernel function is adopted for the estimation.
42 Additionally, note that the short-term eﬀect also comes from the diﬀerent distances from houses to
schools and the number of student per class and teacher. The higher score zone area beneﬁts from the lower
number of student per class and teacher. However, its eﬀect is much smaller than the eﬀect from the top
university entrance rate of a school, since there are small diﬀerences in the number of students per class and
teacher between the two regions.

35

Figure 1.4 School Zone Adjustment Simulation

formance (top university entrance rate) and neighborhood composition (% BA or above
in the 40∼49 age group) will be homogenized within the new school zone over time.
Even though there are no signiﬁcant diﬀerences in the absolute value of WTP as the

percentage of a housing price, there are more substantial gaps in WTP in the amount

of money between the two areas because the average housing price in the higher score

zone area is more than double that of the lower score zone area.

Tables 1.5 and 1.6 show the summary statistics for the policy simulation results.

Overall, regarding the voting experiment, median WTP as a percent and money value

show positive numbers, which means there would be more households supporting the

new school zone in both the short-term and long-term. However, the sum of all house-

36

Table 1.5 WTP for the School Zone Adjustment

Area

All

Lower Score

Higher Score

WTP %

WTP($) / Year

Time

Median Mean

Median

Mean

Short-term
Long-term

Short-term
Long-term

Short-term
Long-term

0.18
2.77

0.42
3.50

0.24
– 0.11

0.61
3.39

55
468

89
697

10
– 677

155
754

– 0.37
– 5.24

– 0.52
– 5.63

– 193
– 2,488

– 220
– 2,942

Obs

183,504

112,458

71,046

Table 1.6 The Changes in Experimental Area

Housing Price ($1,000)

Top Univ. Entrance(%)

BA or Above(%)

Area

Time

Median Mean

SD

Median Mean

All

Lower Score

Higher Score

Initial
Short-term
Long-term

Initial
Short-term
Long-term

Initial
Short-term
Long-term

401.9
402.7
406.9

304.0
305.9
314.0

696.9
693.1
657.1

485.0
485.2
475.4

330.6
332.8
341.4

729.5
726.3
687.4

286.1
284.7
262.9

135.7
137.0
138.9

291.2
291.1
273.2

0.61
0.75
0.86

0.44
0.49
0.69

2.24
2.12
1.66

1.20
1.20
1.18

0.51
0.59
0.77

2.29
2.15
1.84

SD

0.97
0.90
0.69

0.28
0.34
0.24

0.59
0.64
0.66

Median Mean

SD

52.81

56.08

13.13

53.48

55.81

10.25

48.39

47.24

6.48

50.06

48.94

5.26

70.36

70.06

7.71

66.84

66.68

5.88

Obs

183,504

112,458

71,046

holds’ WTP in the long-term becomes negative (the mean WTP is – 0.11% and $ –

677) since the negative WTP in the higher score area ($ – 2,942) is almost four times

larger than the positive WTP in the lower score area ($ 754). Table 1.6 shows the

changes in housing prices, top university entrance rates, and neighborhood composi-

tions in the new school zone. The standard deviation of the top university entrance

rate and neighborhood composition decreases in the long-run by the assumption of the

second scenario, and consequently, housing price gaps within the area are diminished.

Note that the standard deviation of housing price decreases by 8.1% in the long-term

compared to the initial status. The mean housing price and the standard deviation of

housing prices increase in the long-term by 3.27% and 2.35% respectively in the lower

score area and decrease by 5.77% (mean housing price) and 6.18% (standard devia-

tion) in the higher score area. As a result, the housing price gap in median housing

prices decreases by 12.67% (from $ 392,900 to $ 343,100). In conclusion, the results of

the policy simulation show that the school desegregation policy, through school zone

adjustment, signiﬁcantly reduces the housing price gaps within the city.

37

1.5 Dynamic Model

In this section, I investigate the dynamics of the housing market. The analysis in the

above sections using a hedonic price regression shows the implicit prices of housing

attributes in a static equilibrium. Note that the estimated implicit price is interpreted

as the eﬀect of an attribute on market price while holding the other attributes ﬁxed.

However, there could be bilateral relationships among housing attributes over time. For

instance, we might expect that better school performance can attract more educated

people – which is another critical factor among households’ implicit prices for housing
attributes – around the school (school quality ⇒ neighborhood composition). Given
that parents’ education level could also aﬀect the children’s academic performance, the

higher education level of the neighborhood – attracted by improved school performance

– would additionally enhance school performance over time (neighborhood composition
⇒ school quality).

As a consequence, the better school performance indirectly increases the housing
price by attracting a more educated neighborhood (school quality⇒ neighborhood
composition ⇒ housing price) along with its own direct eﬀect on the housing price
(school ⇒ housing price). Also, we can think of the second and the third round bilateral
indirect eﬀects, such as the echo eﬀect (school quality ⇒ neighborhood composition
⇒ school quality ⇒ neighborhood composition). As such, in a dynamic framework, it
is likely that a particular attribute can play a more important role if it attracts and

increases other good attributes of a house. This study uses the panel VAR (vector

autoregression) model as follows to consider the dynamics, which show direct and

indirect eﬀects from the bilateral relationships between housing attributes.

T(cid:88)

Yit =

ΓpYit−p + ΘXit + Djt + ξi + εit

(5)

p=1

Here, Yit represents the vector of endogenous variables in the VAR system, and Γp

and Θ are coeﬃcient matrices. The Yit includes 19 variables (log of housing price;
top university entrance rate; percent of BA or above degree holders in the 40∼49 age
group; log of population; log of employment (full-time and part-time employment), log

of self-employed; the number of student per class and per teacher; and the number

38

of private academies, libraries, supermarkets, culture related stores, shopping stores,

clinics, restaurants, sport facilities, banks, and grocery stores around a house). Xit

denotes time-variant exogenous variables.

In the case of some variables, it is hard

to think that there exists bilateral relationships among variables. For example, the

distance to the nearest school or subway station are more likely to act as exogenous

variables. Those variables can aﬀect the other attributes of a house; however, it is hard

to think of a case where other attributes aﬀect those distances. The vector Xit includes

distances from a house to the nearest subway station, department store, supermarket,

and schools (elementary, middle, and high school). Also, as in the above hedonic price

regression analysis, the model takes into account Djt, the ﬁxed eﬀects over time (t) in
each region (j).43 ξi represents a vector of the house-speciﬁc unobserved heterogeneity.

Note that the model uses the ﬁrst-diﬀerenced variables to control the endogeneity

caused by correlations between ξi and Yit−1, while eliminating the unit root in the
variables.44 Last εit denotes a vector of shock.

From the model selection criteria, the equation (5) is estimated for T = 3 us-

ing the system GMM estimator developed by Arellano and Bover (1995) and Blundell

and Bond (1998), which is applied to the panel VAR by Binder, Hsiao, and Pesaran

(2005). The system GMM also handles the remaining bias problem caused by the ﬁrst-

diﬀerenced error term (Nickell’s bias, Nickell (1981)) by using the lagged variables as

instruments. This estimator is consistent when the panel dataset has a large number of

observations i in a given time dimension. For the estimation, the dataset is constructed

using the same method as the hedonic price regressions in the previous sections; how-
ever, it covers a more extended period (2006∼2015) than the hedonic model for the
panel VAR analysis.45 This study also tries the structural panel VAR model with the

smaller set of endogenous variables while considering the contemporaneous relations

among endogenous variables in the appendix A.3.

The panel VAR system includes 19 endogenous variables, so the system gener-

ates 361 impulse-response functions to show the relations among endogenous variables

43 This paper considers the monthly time ﬁxed eﬀect by 11 school zones, which is the same with the

hedonic regression analysis in the previous sections.

44 After the variable transformation, this study checks the panel VAR system stability condition, and the

system strongly satisﬁes the condition.

45 The house real transaction data does not exist before 2006.

39

over time. Among them, this paper focuses on the impulse-response functions related

to school performance, neighborhood composition, and housing price. Note that the

results of the panel VAR cannot be directly compared with the estimation results of the

hedonic price regression because, ﬁrst, each panel VAR equation includes the lags of

dependent variables while the hedonic regression does not. Second, owing to the data

limitation, the panel VAR model does not include all variables used in hedonic price

regressions. Accordingly, in this section, the analysis focuses on showing the dynamic

properties of the housing market rather than comparing the estimated numbers with

hedonic regressions.

(a) BA : House Price (%)

(b) Top Univ : House Price
(%)

(c) BA : Top Univ (%p)

(d) Top Univ : BA (%p)

Figure 1.5 (Cumulative) IRF

– (Impulse) : (Response) –

Note: “BA” denotes the percent of BA or above degree holders in the 40∼49 age group.
“Top Univ” means the top university (Seoul National University) entrance rate of a high
school.

Figure 1.5 shows the cumulative impulse response functions among three vari-

ables of interest: school performance, neighborhood composition, and housing price.

40

The left side of the colon is an impulse variable and the right side denotes a response

variable. The shock is a one-unit increase, and the line represents the estimated im-

pulse response function. The shaded area is the 95% conﬁdence interval generated by

the 500 Monte Carlo draws. It is important to note that there exists a long-term eﬀect

from two variables (top university entrance rate and the portion of BA or above degree
holders in the the 40∼49 age group) on housing price over time as Figures 1.5a and 1.5b
demonstrate. The one-time shock at the initial time signiﬁcantly aﬀects housing price

even up to ten years later. Also, Figures 1.5c and 1.5d show that there are signiﬁcant

bilateral long-term eﬀects between school performance and neighborhood composition.

Figures 1.5a shows that 1%p increase in the percent of BA or above degree hold-
ers in the 40∼49 age group increases the housing price by 11% in 10 years (in standard
deviation, a 1 SD increasing shock from the “BA” variable leads to the 1.8 SD increase

in the housing price in ten years). Its eﬀect on housing price at time t+1, only including

the direct eﬀect on the housing price, is estimated as 2.4%; however, its indirect eﬀect

through the bilateral relationships among variables consistently increases housing price

for ten years. Meanwhile, according to Figure 1.5b, 1%p higher top university entrance

rate increases a housing price by 1.7% and 4.6% at time t+1 and t+10 respectively (in

standard deviation, a 1 SD impulse shock from “Top Univ” causes a 0.1 SD increase

of the housing price). Figures 1.5c and 1.5d represent the bilateral relationships be-

tween top university entrance rate and the neighborhood composition. Importantly,

they increase and attract each other. The higher portion of educated parents increases

the primary school performance (top university entrance rate), and the better school

performance consistently attracts more educated neighborhoods around a school.

Speciﬁcally, a 1 SD (2%p) increase in the portion of BA or above degree holders
in the 40∼49 age group in neighborhood enhances the top university entrance rate of
schools in the region by 0.96 SD (0.28%p) for ten years after eliminating the regional

time trend.46 In the opposite direction, a 1 SD (0.29%p) higher top university entrance

rate attracts more educated people to a neighborhood by 0.06 SD (0.11%p) in ten years.

In the case of the impulse response from school performance to the neighborhood

composition as shown in Figure 1.5d, the overall long-term eﬀect (0.38%p, at time

46 The panel VAR model includes the regional time ﬁxed eﬀects (11 school zone – monthly level)

41

t+10) – including direct and indirect eﬀects among endogenous variables – becomes

signiﬁcantly larger than its short-term direct eﬀect (0.09%p, at time t+1). On the

other hand, Figure 1.5c demonstrates that the 95% conﬁdence interval of the overall

long-term eﬀect [0.089, 0.175] at time t+10 includes the size of its direct eﬀect (0.105)

at time t+1. In other words, the long-term eﬀect of school performance on the portion

of the well-educated in a neighborhood is more signiﬁcant than the opposite direction

(from neighborhood composition to school performance). Whereas regarding the size

of the eﬀect, the neighborhood composition has a more substantial eﬀect on school

performance (1 SD impulse to the response of 0.96 SD) than the opposite direction

from school performance to neighborhood composition (1 SD impulse to the response

of 0.06 SD).

(a) Culture

(b) Library

(c) Population (%)

(d) Top Univ (%p)

Figure 1.6 (Cumulative) IRF
– Impulse = Top Univ (1%p) –

Figures 1.6, 1.7, 1.8, and 1.9 show the more speciﬁc relationships among endoge-

nous variables regarding school performance and neighborhood composition. Figure 1.6

shows the cumulative response functions from the one-unit (1%p) increase in shock of

the top university entrance rate. Enhanced school performance increases the num-

ber of some amenities such as cultural stores (book, music and sporting goods stores)

and libraries. Also, it increases the population growth while attracting more educated

people to a neighborhood in the region. Figure 1.6d shows the cumulative response

function of the top university entrance rate from the same variable’s shock. Note that

the one-time increase in the top university entrance rate of a school tends to converge

to the original level, which shows the mean-reversion property; however, it does not

decrease to zero and half of the initial shock is maintained.

42

Figure 1.7 shows what statistically induces better school performance. The

increasing number of private academies (as a proxy for a private educational envi-

ronment) and cultural stores such as book-stores enhance school performance in the
region. Along with a higher portion of educated people (% BA or above in the 40∼49
age group) in a neighborhood, the improvement in student-class ratio also increases

the top university entrance rate. All those eﬀects are signiﬁcant and have time-lagged

long-term eﬀects.

(a) Academy

(b) BA (1%p)

(c) Student/Class

(d) Culture

Figure 1.7 (Cumulative) IRF
– Response = Top Univ (%p) –

Figure 1.8 presents the response functions corresponding to the impulse shock
(1%p) of neighborhood composition (% BA or above degree holders in the 40∼49 age
group). The results imply that the higher portion of educated people in a neighborhood

attracts many amenities such as private academies, libraries, culture stores, clinics,

banks, and restaurants. If we focus on the educational environment more speciﬁcally,

a 1 SD (2.05%p) increase in the portion of BA or above degree holders in the par-

ents’ age group results in a 1.78 SD (7.9) increase in the number of private academies.

Interestingly, the increasing portion of the highly educated in a neighborhood leads

to improvements in the ratio of students per class and teacher; however, in the case

of students per teacher, its eﬀect is somewhat insigniﬁcant due to the relatively wide

conﬁdence interval. We can interpret these results as more educated parents searching

for and moving to an area where those ratios are expected to be enhanced shortly be-

cause of a new school opening in the region.47 The increased neighborhood education

level also tends to create more jobs in the region, and the above results are meaningful

47 The plan for establishing a new school usually becomes available to the public a couple of years before

it opens.

43

(a) Academy

(b) Student/Class

(c) Student/Teacher

(d) Library

(e) Culture

(f ) Employment (%)

(g) Clinic

(h) Bank

(i) Restaurant

(j) Supermarket

(k) Top Univ (%p)

(l) BA (%p)

Figure 1.8 (Cumulative) IRF

– Impulse = % BA or Above (1%p) –

44

in that the model also controls for population growth. Given that there is a positive

correlation between education level and income, the increasing number of highly edu-

cated people who are likely to have more disposable income could have more eﬀect on

amenities and businesses around the region compared to a population growth which

does not change the neighborhood education level.

Lastly, Figure 1.8l shows the response function of neighborhood composition

from the same variable’s shock, and we can ﬁnd the critical diﬀerence between the

corresponding impulse response function of the top university entrance in Figure 1.6d.

In the case of top university entrance rate, it shows the mean reversion trend. That is,

if the top university entrance rate increases in a school, then it tends to lower back to its

original level in the future. On the other hand, neighborhood composition shows strong

persistency, implying that if the portion of highly educated in a neighborhood increases

in a region, it tends to increase more in the future. Note that a 1%p increase in the

portion of BA or above degree holders (at time t) induces a 2.28%p increase after ten

years. Because of this diﬀerent property between school performance and neighborhood

composition – mean reversion and persistency – neighborhood composition plays a more

critical role than school performance in a dynamic framework based on more ﬂuent and

signiﬁcant relationships with other endogenous variables.

Next, the results in Figure 1.9 show what attracts the more educated parents to

the region. Importantly, as I discuss in the above main panel VAR results, better school

performance attracts more educated parents to the region. If we look into the impulses

from two ratios (students per class (Figure 1.9b) and teacher (Figure 1.9c)), in the case

of the number of students per class, the result is not signiﬁcant, which is diﬀerent from

the result of the opposite impulse-response relationship shown in Figure 1.8b. On the

other hand, the enhanced ratio of students per teacher attracts the highly educated to

a neighborhood even though its conﬁdence interval is relatively wide, and the result is

similar to the impulse response function of the opposite direction as shown in Figure

1.8c.

If we examine the amenities attracting more educated parents, the increasing

number of culture stores (book, stationery, music, and sporting goods stores), sports

facilities, commercial banks, and restaurants tends to attract more educated parents

to the region.

45

(a) Top Univ (1%p)

(b) Student/Class

(c) Student/Teacher

(d) Culture

(e) Sport

(f ) Bank

(g) Restaurant

Figure 1.9 (Cumulative) IRF

– Response = % BA or Above (%p) –

Figures A3 and A4 in the appendix show the other impulse response functions

related to the housing price. Figure A3 presents the responses of housing price from

shocks of other variables. The increase in the number of banks, culture stores, sports

facilities, and restaurants has a signiﬁcant long-term eﬀect on the housing price. Note

that those variables coincide with the factors that attract more educated people to a

neighborhood and those that are increased by the higher portion of the educated in a

neighborhood. Thereby, we could expect that those amenities have a long-term positive

eﬀect on housing prices due to the bilateral relationship with neighborhood composition

over time. Figure A4 shows the response functions corresponding to the shock from

a housing price, which demonstrates the deﬁnite diﬀerence from the hedonic model.

Note that the increasing housing price usually crowds out many amenities including the

portion of highly educated people in a neighborhood and the top university entrance

rate. However, given that the size of the eﬀects is quite small for all variables, we could

conclude that the housing price has a strong property as an endogenous variable.

To summarize, this study veriﬁes the signiﬁcant bilateral relationships between

the critical attributes of a house.

Importantly note that neighborhood composition

46

(the percent of BA or above degree holders in the 40∼49 age group) leads the key
dynamics of housing attributes while attracting good amenities and enhancing school

performance. The school performance also has an important role in the dynamics;

however, the strong mean-reversion trend weakens the eﬀect whereas the persistence

in neighborhood compositions strengthens the dynamic eﬀect.

1.6 Conclusion

This study shows that past records of school performance have a long-lasting reputation

eﬀect from the hedonic price regression model using a short-term panel dataset that

overcomes the problems of the housing market panel dataset. The results imply that

people consider the past records of school performance as well as the most recent results

and that households have diﬀerent weights on previous results. That is, people place

more weight on recent results of a school, and its weight fades out as time passes by.

This study examines households’ willingness to pay for better educational envi-

ronments and a more educated neighborhood and shows that they account for more

than 43% of housing price gaps in Seoul. Based on the estimates of hedonic analysis,

this paper tries a policy simulation that changes school zone boundaries. According

to the simulation results, the school zone adjustment has signiﬁcant eﬀects on housing

prices and neighborhood compositions. The simulation model also predicts that the

policy – changing school zone boundaries – would help to mitigate huge gaps in housing

prices and school performance between the Gangnam and the other areas in Seoul.

This study also investigates the dynamics of the housing market by using the

panel VAR model. Unlike the hedonic analysis showing the eﬀect of a variable on the

market price of a house while holding the other attributes ﬁxed, the panel VAR model

allows bilateral relationships among housing attributes and examines the direct and

indirect long-term eﬀects among them. According to the results, the overall eﬀects

of school performance in the long-term becomes more signiﬁcant than its eﬀects in

the short-term. Also, neighborhood composition shows a more remarkable eﬀect in

the dynamic framework compared to the static hedonic model. Because of its per-

sistence, which is diﬀerent from the mean reversion property of school performance,

neighborhood composition plays a critical role in the long-term based on ﬂuent and

47

signiﬁcant relationships with other housing attributes. To summarize, this paper veri-

ﬁes households’ signiﬁcant willingness to pay for better educational and neighborhood

environments and also ﬁnds evidence that they play a more critical role if we consider

their dynamic eﬀects on housing attributes over time.

48

APPENDIX

49

APPENDIX A

A.1 Data Appendix

A.1.1 Housing Environment in Seoul

Seoul Metropolitan Area

The population of Seoul is over 10 million, and its population density (16,204/km2) is

higher than New York City (10,725/km2). Due to the high population density, most

people live in apartments, and less than 15% of people live in a single detached house.

1.2 million people commute from outside the city, and on average, 7 million people use

the subway each day to avoid traﬃc jams. Therefore, the distance from a house to

the nearest subway station becomes a critical attribute aﬀecting housing prices.48 In

2015, there were over 22 department stores and 59 large supermarket chains in Seoul,

and people frequently visited them for groceries and shopping. Distance to the nearest

department store and supermarket also turns out to be an important factor in housing

price. The Han-river runs through the middle of Seoul. In Seoul, it has a roughly 1km

(0.62miles) river width on average, and the houses alongside the river oﬀer an expan-

sive river view. Two expressways and many parks located along both sides of the river

oﬀer good transportation environments and recreation areas for nearby residences. For

these reasons, there could be a premium in the price of a house near the Han-river.

This paper generates three dummy variables according to the distance to the Han-river
(0∼0.5km, 0.5∼1km, and 1∼1.5km from the Han-river).

Brand of Apartments and High-rise Residential Complexes

By using the name or brand of the apartment, I generate a brand dummy variable if

the apartment has a major brand name.49

In Korea, almost every apartment has a

brand name, similar to cars, and the brand name represents the construction ﬁrm of

48 There were 285 subway stations in Seoul in 2015
49

I selecte major apartment brands using BSTI (Brand Stock Top Index, www.brandstock.co.kr) infor-
mation. It publishes the brand ranking of various products and ﬁrms. I make the brand dummy variable 1 if
the apartment brand is Ramian(Samsung), Xi(GS), I’PARK(Hyundai), Hill-state(Hyundai), We’ve(Doosan),
Prugio(Daewoo), The Sharp(Posco), Lotte Castle(Lotte), E-Pyunhan Sesang(Daerim), or SK-VIEW(SK), and
0 if otherwise. (in parenthesis is the construction ﬁrm of the brand).

50

the apartment building. According to a recent survey,50 the brand of apartments was

the ﬁrst consideration for people given apartments are located in the same place. That

is because the brand shows the construction company, and people prefer to have their

apartments constructed by famous conglomerate ﬁrms such as Samsung, Hyundai,

GS, or Doosan. Additionally, the top-ranked brand apartments usually have better

conditions (e.g., better landscape and curb appeal, better location, spacious parking

lots, security, and more amenities), which are hard to measure by numerical values and

are usually unobservable attributes to an econometrician. Consequently, controlling

the brand eﬀect can help capture one of the signiﬁcant unobservable attributes of

apartments.

In addition to the brand of apartments, there were 84 high-rise residential com-

plexes in Seoul in 2015. In the complex building, lots of amenities are located on the

lower level and apartments units on high ﬂoors oﬀer an expansive view. To consider the

eﬀect from diﬀerent types of apartments, this paper also generates a dummy variable

for high-rise residential complexes.

A.1.2 Education and Neighborhood Statistics

Top University Entrance Rate

This paper uses the top university entrance rate of a high school as the primary variable

of interest representing the performance of a school, which is diﬀerent from previous

studies that use various test score results as a proxy for school performance.51 Ac-

cording to the results of this paper, both the top university entrance rate and the test

score (N-SAT) are critical factors regarding household willingness to pay. However,

the eﬀect of top university entrance statistics turns out to overwhelm that of test score

results in households’ implicit price as discussed in Section 1.3.2.

That being said, there are some problems in using the university entrance statis-

tics in the analysis. First of all, the university entrance statistics by a high school are

usually not publicly available. Second, the universities considered prestigious can be

diﬀerent from region to region. For example, in a vast country like the US, a prestigious

50 http://www.rcast.co.kr/
51 To the best of my knowledge, I could not ﬁnd any previous literature using prestigious university

entrance rate as a variable of interest representing the performance of a school.

51

school which the top students enter could be diﬀerent from the west side to the east

side. However, such variations do not exist in small countries like Korea, and more-

over, this study focuses only on the housing market of Seoul, Korea. As the primary

variable, this paper uses the portion of students who entered Seoul National University

(SNU) by high school. The prestige university ranking is stable in Korean society, and

Seoul National University (SNU) is the top-ranked university. It would be better if the

data about the entrance statistics of other prestige universities were also available to

be used together to generate performance statistics of high schools. Unfortunately, this

data is not available to the public, and the entrance statistics of SNU by high school

becomes available only due to a member of the National Assembly who requested the

data for parliamentary inspection of Seoul National University. Given data limitations,

this paper assumes there would be a strong positive correlation between the number of

students who enter SNU and other prestigious universities when considering the strong

university hierarchy in Korea. Under this assumption, the top university entrance rate

can be a good proxy to evaluate the university entrance results of a high school, which

is one of the most critical performance results.

Nationwide Scholastic Achievement Test

The results of the N-SAT52 are available from the school information system of Korea53

by each school. This site oﬀers information about the number of students taking the

test and test results for each exam subject (Korean, Math, and English). The results

are provided after dividing students into three categories (above normal, normal, and

below normal) and showing each percentage by subject in a high school. Based on the

data, I calculated the average “Above Normal” ratio of three subjects and divided it

with the total average “Above Normal” ratio of all high schools in Seoul. For example,

N-SAT score “120” means that there are 20% more “Above Normal” category students

in the high school compared to the average high school in Seoul. The average test score

is an imperfect quality measure. Nonetheless, it has the advantage of being easily mea-

52 This test is diﬀerent from the SAT which is used as the university entrance exam. The purpose of this
test is to measure the average academic levels of students and schools. However, this test score is free from
the selection bias problem which can occur in scores of an exam that students can choose whether to take the
exam because almost all students take the exam and there is no choice for students.

53 http://www.schoolinfo.go.kr

52

sured by numbers, observed by both parents and researchers, and compared with other

schools; as a result, it has been used in most analyses that attempt to measure the

implicit price for school quality.

Types of High Schools and the Matching Process of School Performance

Statistics

There are three types of high schools in Seoul: public, private, and special-purpose high

schools. Since private and special-purpose high schools adopt an open enrollment sys-

tem, students can apply for admission to those schools regardless of residence location

while public high schools use the allocation system following school catchment area.

Private high schools usually have better performance results in top university entrance

rate and N-SAT scores compared to nearby public schools. Note that also, in general,
only the top 2∼3% ranked students in each middle school enter special-purpose high
schools, and consequently, the test score results and top university entrance rates of

special-purpose high schools are outstanding. In this study, I use “top high school”

instead of its oﬃcial name “special-purpose high school.” The bottom line is that

the three types of high schools and diﬀerent enrollment systems create obstacles for

analyzing implicit prices of better educational environments.

To overcome the above problem, this study tried various methods as follows.

The ﬁrst method used the high school entrance rate for three types of high schools

by each middle school. For the ﬁrst step, I matched the high school entrance rate of

middle schools to a house by distance-weight matrices considering the middle school

enrollment system,54 and used the matched statistics (diﬀerent entrance rate for each

type of high school by a house) as a weight to combine performance statistics of diﬀerent

types of high schools.

Next, I matched high school performance statistics with a house by each high

school type. In the process, this paper used a diﬀerent method to reﬂect the diﬀerent

enrollment systems (open enrollment and catchment area). For the matching of private

and top high schools, which adopt open enrollment systems, with a house, this paper

used distance-weight matrices for each type of high school, and each matrix is generated

54 Diﬀerent from the high school enrollment system, middle schools only use the school zone system

following catchment area.

53

by calculating the distances from a house to all private and top high schools in Seoul.

Even though the location of a house is irrelevant in applying to those schools, in

practice, the beneﬁt from and probability of attending nearby schools becomes higher

as the distance to the school gets closer because of daily commute costs. Moreover,

households generally select a house near a school which they want their child to attend

when they choose the location of the house.

For the performance statistics of public high schools, this paper also used distance-

weight matrices. However, for this step, it also considered the school zone and stu-

dent allocation system of public schools along with the distance from houses to public

schools.55 As the last step, I combined the matched statistics by high school type by

using the weight among private, public, and top high schools calculated during the ﬁrst

step.

For example, suppose that the matched high school entrance rate statistics of

a house “A” are as follows: 80% (public high school), 17% (private high school), and

3% (top high school), which means 80%, 17%, and 3% of students respectively enter

public, private, and top high schools from middle schools around the house “A.” Next,

we have high school performance statistics for each type of high school matched with

the house “A” generated during the second step. Last, we can combine the performance

statistics from three diﬀerent high school types by the weight – 0.8, 0.17, and 0.03.

However, the critical problem of the above method is that people could hold dif-

ferent levels of importance among diﬀerent types of high school. For example, suppose

that on average 17% and 80% of students enter private and public high schools in the

region. Then, in this case, the performance of public high schools should have a more

signiﬁcant eﬀect than private high schools when people evaluate school performance

in the region. However, if households over-estimate the probability of attending the

private high school than the actual rate, then the actual weight of the household, which

is based on their subjective probability and aﬀecting the willingness to pay, could be

entirely diﬀerent from the actual entrance rate of each type of high school.

This paper found strong evidence that households’ implicit weight among dif-

ferent types of schools could be substantially diﬀerent from the actual probability of

55

I further explain school zone and the students’ allocation system of public high schools in the next

section.

54

entering each type of high school. I used the same model (6) in Table 1.2, and included

the last ten years’ records of top university entrance rate of public, private, and top

high school respectively. According to the results, the implicit price of top high school

performance was not signiﬁcant. Given that the entrance rate for the top high school
is around 2∼3%, the result seems reasonable. However, households’ willingness to pay
for the better performance of private high schools was more than twice than that of

public high schools. Note that 85% of students enter public high school whereas only

12% of students are admitted to private high schools in Seoul. Hence, the severe level

of discrepancy between high school entrance rate statistics and estimated implicit price

of school performance by high school type implies that households’ subjective weight

among diﬀerent types of schools aﬀecting their WTP for better school performances

does not reﬂect the actual high school entrance rate.

There could be a few reasons for the discrepancy. First, households can over-

estimate the probability of attending the private high school. The willingness to pay

for better school performance is revealed when people buy a house.

In most prac-

tices, household movements are usually made long before children go to high school.56

Accordingly, what they anticipated at the time of moving and actual high school en-

rollment a few years later can be somewhat diﬀerent. Second, psychological factors

such as the bandwagon eﬀect can be another reason. Private high schools are evenly

distributed in Seoul, and each high school represents the school performance of the

region because they are, in general, the best performing school in each region. Thus, if

people consider the best performing high school representing the area when evaluating

the school performances of a certain region, the performance of a private high school

representing each area can have a larger eﬀect than a public school. To summarize, the

ﬁrst method – which uses the high school entrance rate for three types of high schools

as the weight for combining the high school performances from three diﬀerent types –

can be inappropriate for matching the high school performances to a house.

Based on the above problems and ﬁndings, this paper used the following method.

First, I excluded the performance statistics of top high schools for the matching pro-
cess based on the fact that only 2∼3% of top-ranked students in a middle school enter

56 According to a survey, most households with students move before their child enters a secondary school

and they tend to stay while the child goes to school.

55

top high schools. Also, the estimation result showing no signiﬁcant implicit price of

performance of top high schools supports the exclusion of top high schools. Second,

this paper did not distinguish between private and public high schools for the match-

ing process; however, it considered the school zone system. For example, suppose that

20 public schools are located in school zone “A” , and there are 22 private schools in

Seoul. Again note that public high schools adopt a school zone system while private

high schools use open enrollment. For the matching between the school performances

and a house located in school zone “A”, the distance weight matrix calculates distances

from the house to both 20 public high schools in the same school zone and 22 private

schools in Seoul. That is, the 22 private high schools are involved in the matching

process of all houses in Seoul while considering distances, whereas public schools are

included in the distance matrix according to the school zone and the location of a house.

School Zone and Student Allocation System

To analyze the housing market with the hedonic price regression, we need to match the

education-related statistics with a house. If a one-to-one mapping is possible between

a school and a house according to a clearly deﬁned school zone, the matching is simple.

However, the school zone system is rather complex in Seoul. The ﬁrst problem is

that there are multiple public high schools in the same school zone. Also, the student

allocation system considers the ease of commute when assigning students to a high

school, but there is no guarantee that students are allocated to the nearest school from

their houses. Moreover, the education authority of Seoul uses a three-step application

system when allocating middle school students to high schools.57

However, even though the multi-step student allocation process includes an open

enrollment system during the ﬁrst step, according to the recent publication data,58 more

than 96% of students have attended a school in the school zone where they reside.

57 At the ﬁrst stage, 20% of students among the total capacity of each high school are enrolled by student
applications, and the application is not restricted by the school zone and the location of house. That is, in
the ﬁrst stage, 20% of student enrollment is the same as the open enrollment system. Next, 40% of the total
capacities are made up of the applications of students who live in the school zone. Lastly, the remaining 40%
of students are assigned from the wide school zone (two or three adjacent school zones consist of a wide school
zone). During the process, if there are more applicants than the number of capacity of a school at each step,
the lottery system is used to select students. However, since even narrow school zones include more than 20
high schools on average, we cannot match a house with a high school by one-to-one.

58 The Seoul Metropolitan Oﬃce of Education oﬀers this data (http://statistics.sen.go.kr/)

56

There can be several reasons for this. First of all, daily commute costs to attend high

schools in another school zone could be substantial. Secondly, even though students

have the choice to apply for a school located in the other school zone, the possibility of

actually attending the school can be pretty low. In general, the schools that a parent

and a student want to attend usually coincide with other people’s choices. The similar

preferences among people can lower the allocation possibility of students from other

school zones. For instance, students can apply to a popular high school in their school

zone at every application step, whereas a student from other school zones can only be

an applicant during the ﬁrst step, which accounts for only 20% of the total allocations.

Therefore, this study assumes that all students are allocated to a public high school

within their school zone, ignoring the less than 4% of students who attend a public

high school in other school zones.

Neighborhood Statistics
This paper primarily uses the portion of BA or above degree holders in the 40∼49
age group in a neighborhood as a proxy for neighborhood education level in a region.

The data is drawn from the Population Census which contains details of the education

level of people and is published every ﬁve years. However, for the panel VAR analysis,

we need every year of information about neighborhood education levels by age group.

This paper estimates the neighborhood characteristic between the ﬁve years by using

interval regressions. For this process, this paper adopts the method used by Bayer,

Ferreira, and McMillan (2007). They use the Census data that represents housing

prices as 26 categorical bands.59 From the categorical value, they estimate the housing

price – the dependent variable of their analysis – by using interval regressions, which

restrict the estimated housing price so it lies in the interval. In this paper, I use interval

regressions based on the generalized Tobit and estimate the neighborhood education

level of a region in a year when the Population Census is not available. This paper uses

the value of two adjacent Population Census dataset as lower and upper boundaries.
For example, if we estimate the portion of BA or above degree holders in the 40∼49 age
group in a census tract in 2013, this paper uses the portions of that in 2010 and 2015 as

59 For example, the Housing Census shows the interval in which the housing price is included such as $

100,000 ∼ 200,000.

57

the lower boundary and upper boundary respectively. The neighborhood information is

estimated separately for each of the 25 districts in Seoul. For the estimation, this paper

uses the demographic information that is published every year by census tracts such
as the changes of a population by age group (the changes of population in the 35∼39,
40∼44, 45∼49, and 50∼54 age group in a region). This paper also uses neighborhood
education levels of diﬀerent age groups (the portion of BA or above degree holders in
the 35∼39 and 50∼54 age groups) and the changes in the number of employees (the
number of full-time and part-time workers and the number of self-employed) by census

tracts.

A.1.3 Panel Dataset Generating Process

This section explains how the two consecutive years panel dataset was generated for

the analysis in this paper. Because of the same group deﬁnition as described in section

2.2, it is likely that multiple transactions of the same housing unit happen within a

year, and in some cases, even in a month. From the multiple transactions within a

year, this study generated the price of the year of a group of houses as follows. First,

this paper calculated the average monthly price of a group of apartments from the

multiple transactions.

In the case of a month, it is not common to have multiple

transactions within the same group. Second, from the monthly average price data, this

paper constructed monthly panel dataset (unbalanced) within a year. Third, by using

the monthly panel dataset, the monthly ﬁxed eﬀects by 25 district regions are estimated

to eliminate irrelevant macro ﬂuctuations from the analysis. Fourth, I converted the

average monthly price into the price of the base month using the estimated monthly

price index.

As a robustness check for the same ﬂoor group deﬁnition, this paper also tried

other group deﬁnitions. To generate more observations in the panel dataset, this paper

put a group of apartments in the same group if their address, size, and construction

year are all the same after eliminating the diﬀerent ﬂoor eﬀect. That is, this paper

converted the housing prices located on the various ﬂoors in the same building into

the virtual prices assuming all houses are located on the same ﬂoor. To estimate

and eliminate the ﬂoor eﬀect, I constructed the monthly panel dataset with the same

58

method that is used to estimate the monthly price index by region. After that, the

ﬂoor group ﬁxed eﬀect is estimated from the panel dataset and used to change the

housing price into the counterfactual price of the same ﬂoor. Next, this paper also

converted the housing price – with the ﬂoor eﬀect eliminated – into the same monthly

price, again using the monthly price index by region.

59

A.2 Table & Figure Appendix

D. Amenities

Manufacturing
Supermarket
Grocery
Cloth
Culture
Restaurant
Bank
Research
Law and Tax
Company HQ
Gov oﬃce
Police, Fire Oﬃce
Elementary School
Middle and High School
Private Academy
Clinic
Library
Sport
Sewage
Trash Collection
Trash Disposal
Hotel, Motel
Alchols
Welfare Accommodate

PM10

Obs.

Obs.

Table A1 Summary Statistic

All

Average Median

SD

Gangnam
Average Median

147
43.9
40.8
77.3
14.3
194.3
7.9
4.9
24.1
16.2
1.7
1.4
3.5
1.7
31.1
37.3
4.2
15.4
0.4
0.4
0.1
8.4
88.8
2.2

66
38.0
31.0
37.0
11.0
150.0
5.0
1.0
4.0
3.0
1.0
1.0
3.0
1.0
22.0
29.0
3.0
13.0
-
-
-
3.0
63.0
1.0

327
28.6
62.6
147.3
13.1
172.0
9.6
14.1
99.0
49.2
1.5
1.3
2.1
1.8
37.3
37.9
5.7
12.6
0.7
0.8
0.4
15.4
98.3
2.6

72
45.7
34.4
93.6
18.9
259.4
13.0
9.4
82.3
54.5
1.6
1.2
3.2
1.8
55.1
69.7
6.1
26.3
0.3
0.4
0.2
5.5
110.2
1.6

68
37.0
32.0
45.0
15.0
197.0
9.0
4.5
19.5
19.0
1.0
1.0
3.0
1.0
36.0
43.0
5.0
25.0
-
-
-
1.0
78.0
1.0

SD

55
32.5
20.4
141.5
12.8
235.5
11.7
12.2
232.6
94.5
1.0
1.4
1.9
1.6
61.0
75.7
4.9
20.3
0.7
0.7
0.5
11.2
117.0
1.9

The Other
Average Median

162
43.8
42.0
73.4
13.5
183.8
7.0
4.1
13.3
10.0
1.7
1.4
3.5
1.7
27.1
31.7
3.9
13.6
0.4
0.5
0.1
9.0
85.7
2.3

67
38.0
31.0
36.0
11.0
144.0
5.0
1.0
3.0
3.0
1.0
1.0
3.0
1.0
21.0
28.0
3.0
12.0
-
-
-
3.0
63.0
2.0

SD

353
28.1
67.5
147.1
13.0
156.8
8.9
14.3
29.3
33.2
1.6
1.2
2.2
1.8
30.3
23.0
5.8
9.7
0.7
0.8
0.4
16.0
94.8
2.6

423

66

357

47.5

47.4

2.1

47.8

48.0

1.4

47.5

47.3

2.2

40

5

35

60

(a) School Zone Map

(b) Top University Entrance Rate (Above 2%)

Figure A1 School Zone, Top Performing High School

61

(a) Portion of University Graduates (Age 40∼49, Above 62%)

(b) Housing Price per Size (Top 5%)

Figure A2 Neighborhood Segregation and Housing Price

62

Table A2 School Reputation Eﬀect

– 25 District Fixed Eﬀect –

Panel A : Including all previous (10 years) results of schools

Dependent Variable = Log(Price)

Variable

Cross Section

(1)

(2)

House Fixed

(3)

(4)

Time-varying

(5)

(6)

Top Univ.
Entrance Rate (t)

2.769***
(0.618)

3.938***
(0.729)

1.083***
(0.320)

1.381***
(0.373)

1.343***
(0.322)

1.489***
(0.358)

(t − 1)

(t − 2)

(t − 3)

(t − 4)

(t − 5)

(t − 6)

(t − 7)

(t − 8)

(t − 9)

4.140***
(0.550)

6.309***
(0.704)

2.716***
(0.305)

1.834***
(0.335)

2.976***
(0.328)

1.595***
(0.366)

-0.865
(0.634)

-1.393
(0.748)

2.048***
(0.313)

1.505***
(0.373)

2.119***
(0.334)

1.237***
(0.284)

2.088**
(0.689)

4.815***
(0.633)

1.702***
(0.259)

0.982***
(0.263)

2.009***
(0.268)

1.191***
(0.277)

0.0687
(0.630)

3.277***
(0.641)

-0.950
(0.718)

-1.162
(0.740)

-1.797**
(0.563)

-1.699**
(0.622)

-4.271***

-5.491***

(0.578)

(0.646)

-2.653***

-3.257***

(0.451)

(0.520)

4.672***
(0.526)

3.763***
(0.564)

1.393***
(0.257)

1.075***
(0.300)

1.758***
(0.289)

1.077***
(0.324)

1.329***
(0.279)

0.720***
(0.213)

0.728**
(0.235)

0.822***
(0.180)

0.393
(0.226)

0.925**
(0.343)

0.904**
(0.344)

0.786*
(0.387)

-0.648*
(0.303)

0.153
(0.238)

1.876***
(0.269)

0.607***
(0.238)

0.946***
(0.215)

0.840***
(0.251)

0.763**
(0.312)

0.465
(0.358)

0.751***
(0.187)

-0.576**
(0.286)

0.120
(0.240)

0.543**
(0.226)

Sum(t ∼ t − 7)

Sum(t ∼ t − 9)

W/ All lags
Obs.
Adj R2

5.408***

4.366***

11.718***

9.392***

13.866***

8.425***

(1.408)

(1.432)

(1.514)

(1.834)

(1.619)

(1.993)

7.428***

4.872***

12.933***

8.897***

14.737***

8.392***

(1.207)

(1.239)

(1.695)

(2.088)

(1.832)

(2.019)

144,492

0.887

Yes

144,492

0.890

72,246
0.197

Yes

72,246
0.237

72,246
0.991

Yes

72,246
0.993

Panel B : Including previous 3-year averaged results of schools

Variable

Cross Section

(7)

(8)

Top Univ. Entrance
3-Year Average

8.510***

5.697***

(1.094)

(0.378)

W/ All lags
Obs.
Adj R2

144,492

0.878

Yes

144,492

0.889

House Fixed

Time-varying

(9)

2.369***

(0.536)

72,246
0.188

(10)

0.537
(0.702)

Yes

72,246
0.232

(11)

2.108***

(0.756)

72,246
0.989

(12)

0.842
(0.609)

Yes

72,246
0.993

Standard errors clustered at IDs are in parentheses

*** p < 0.01, ** p < 0.05, * p < 0.1

Note: All coeﬃcients are reported as β*100. All models include ﬁxed eﬀects (25-district
by monthly level) and the other variables listed in Table 1.1 and A1 in the model.

63

Table A3 Reputation Eﬀect (House Fixed-Eﬀect)

– 11 School Zone –

Variable

Dependent Variable = Log(Price)

Distance-Weight

Radius : 1km

Nearest 3 Schools

Radius : 1km

Nearest 3 schools

(1)

(2)

(3)

(4)

(5)

(6)

Top Univ.
Entrance Rate (t)

1.991***
(0.328)

1.942***
(0.359)

1.175***
(0.214)

0.902***
(0.233)

1.038***
(0.227)

1.149***
(0.261)

(t − 1)

(t − 2)

(t − 3)

(t − 4)

(t − 5)

(t − 6)

(t − 7)

(t − 8)

(t − 9)

3.117***
(0.307)

2.602***
(0.310)

2.113***
(0.191)

1.622***
(0.226)

1.932***
(0.196)

1.732***
(0.251)

2.049***
(0.273)

2.457***
(0.361)

1.511***
(0.173)

1.361***
(0.240)

1.485***
(0.179)

1.754***
(0.218)

1.936***
(0.241)

1.612***
(0.391)

1.688***
(0.169)

1.093***
(0.252)

1.607***
(0.170)

0.983***
(0.216)

1.104***
(0.217)

1.595***
(0.378)

0.697***
(0.151)

1.094***
(0.198)

0.585***
(0.150)

1.044***
(0.265)

1.218***
(0.324)

1.372***
(0.260)

0.632**
(0.215)

0.710***
(0.194)

0.538*
(0.225)

0.779***
(0.172)

1.147***
(0.242)

0.448*
(0.197)

0.619**
(0.200)

0.563**
(0.177)

1.052**
(0.392)

0.626
(0.407)

0.780**
(0.289)

0.377
(0.239)

0.427**
(0.130)

0.380*
(0.150)

0.188
(0.143)

0.749**
(0.266)

0.499**
(0.167)

-0.0967
(0.193)

0.512***
(0.117)

0.473**
(0.152)

0.396**
(0.135)

0.324*
(0.149)

0.215
(0.138)

0.338**
(0.118)

0.536*
(0.271)

0.311
(0.279)

0.580**
(0.206)

0.404*
(0.167)

Sum(t ∼ t − 7)

Sum(t ∼ t − 9)

13.010***

13.259***

8.623***

8.030***

7.904***

8.289***

(1.505)

(1.936)

(0.955)

(1.165)

(0.995)

(1.276)

14.192***

14.416***

9.323***

8.407***

8.456***

9.273***

(1.654)

(2.161)

(1.061)

(1.306)

(1.096)

(1.435)

W/ All lags

Obs.
Adj R2

72,246
0.194

Yes

72,246
0.258

72,246
0.185

Yes

72,246
0.235

72,246
0.194

Yes

72,246
0.256

Standard errors clustered at IDs are in parentheses

*** p < 0.01, ** p < 0.05, * p < 0.1

Note: All coeﬃcients are reported as β*100. All models include ﬁxed eﬀects (11 school
zone by monthly level) and the other variables listed in Table 1.1 and A1 in the model.

64

Table A4 Reputation Eﬀect (Time-varying)

– 11 School Zone –

Dependent Variable = Log(Price)

Variable

Distance-Weight

Radius : 1km

Nearest 3 Schools

Radius : 1km

Nearest 3 schools

(1)

(2)

(3)

(4)

(5)

(6)

Top Univ.
Entrance Rate (t)

2.254***
(0.346)

1.964***
(0.361)

1.176***
(0.227)

1.324***
(0.238)

1.095***
(0.244)

1.207***

(0.21)

(t − 1)

(t − 2)

(t − 3)

(t − 4)

(t − 5)

(t − 6)

(t − 7)

(t − 8)

(t − 9)

3.787***
(0.325)

2.74***
(0.28)

2.421***
(0.202)

1.653***
(0.229)

2.282***
(0.209)

1.926***

(0.19)

2.800***
(0.283)

1.588***
(0.321)

1.975***
(0.182)

1.287***
(0.213)

1.954***
(0.190)

1.325***
(0.247)

2.481***
(0.255)

1.793***
(0.298)

2.092***
(0.179)

1.731***
(0.201)

1.950***
(0.180)

1.06***
(0.222)

1.884***
(0.254)

1.313***
(0.307)

0.965***
(0.162)

1.47***
(0.178)

0.704***
(0.164)

0.799***
(0.155)

1.693***
(0.325)

0.871***
(0.231)

0.844***
(0.216)

1.077***
(0.228)

0.689***
(0.202)

1.244***
(0.237)

0.831**
(0.388)

0.409
(0.359)

0.643***
(0.132)

0.791***

(0.16)

0.702***
(0.153)

0.726***
(0.152)

0.452*
(0.219)

0.736***
(0.273)

0.220
(0.152)

0.056
(0.178)

0.677**
(0.226)

0.544***
(0.137)

0.618***
(0.153)

0.108
(0.149)

0.9***
(0.175)

0.552**
(0.24)

0.575**
(0.266)

0.454**
(0.188)

0.656***
(0.182)

0.847***
(0.226)

0.656***
(0.122)

0.754***
(0.143)

0.452***
(0.121)

0.584***
(0.153)

Sum(t ∼ t − 7)

Sum(t ∼ t − 9)

16.831***

11.509***

10.818***

10.058***

9.822***

8.345***

(1.615)

(1.725)

(1.018)

(1.045)

(1.075)

(1.166)

17.939***

13.091***

11.694***

10.868***

10.383***

9.383***

(1.935)

(1.955)

(1.144)

(1.183)

(1.196)

(1.255)

W/ All lags

Obs.
Adj R2

72,246
0.991

Yes

72,246
0.993

72,246
0.991

Yes

72,246
0.993

72,246
0.991

Yes

72,246
0.993

Standard errors clustered at IDs are in parentheses

*** p < 0.01, ** p < 0.05, * p < 0.1

Note: All coeﬃcients are reported as β*100. All models include ﬁxed eﬀects (11 school
zone by monthly level) and the other variables listed in Table 1.1 and A1 in the model.

65

Table A5 Reputation Eﬀect (House Fixed-Eﬀect)

– 25 District –

Variable

Dependent Variable = Log(Price)

Distance-Weight

Radius : 1km

Nearest 3 Schools

Radius : 1km

Nearest 3 schools

(1)

(2)

(3)

(4)

(5)

(6)

Top Univ.
Entrance Rate (t)

1.618***
(0.270)

1.315***
(0.290)

1.149***
(0.217)

1.112***
(0.258)

1.124***
(0.231)

0.727***
(0.191)

(t − 1)

(t − 2)

(t − 3)

(t − 4)

(t − 5)

(t − 6)

(t − 7)

(t − 8)

(t − 9)

2.598***
(0.312)

2.322***
(0.392)

1.716***
(0.196)

1.121***
(0.236)

1.544***
(0.199)

1.117***
(0.231)

2.142***
(0.335)

1.640***
(0.336)

1.224***
(0.178)

1.023***
(0.245)

1.168***
(0.178)

1.104***
(0.258)

1.303***
(0.285)

1.313***
(0.259)

1.109***
(0.332)

0.663**
(0.223)

0.668**
(0.235)

0.636***
(0.181)

0.348
(0.223)

1.093**
(0.402)

1.186**
(0.381)

1.078**
(0.377)

0.941*
(0.425)

0.797
(0.429)

0.0130
(0.326)

0.105
(0.256)

0.875***
(0.183)

0.638**
(0.202)

0.830***
(0.186)

0.853***
(0.227)

0.557***
(0.141)

0.668**
(0.207)

0.544***
(0.147)

0.801**
(0.264)

0.571***
(0.162)

0.546**
(0.177)

0.385
(0.217)

0.424**
(0.160)

0.524***
(0.119)

0.0602
(0.155)

0.529*
(0.237)

0.495
(0.274)

-0.531*
(0.208)

0.191
(0.159)

0.527**
(0.164)

0.384
(0.225)

0.380*
(0.161)

0.0728
(0.151)

0.376**
(0.120)

0.689*
(0.268)

0.350
(0.286)

0.452
(0.291)

0.139
(0.224)

0.230
(0.176)

Sum(t ∼ t − 7)

Sum(t ∼ t − 9)

11.413***

10.373***

6.903***

6.133***

6.501***

6.093***

(1.577)

(2.087)

(0.992)

(1.214)

(1.022)

(1.332)

12.397***

10.491***

7.487***

5.793***

6.950***

6.462***

(1.749)

(2.367)

(1.110)

(1.367)

(1.131)

(1.505)

W/ All lags

Obs.
Adj R2

72,246
0.208

Yes

72,246
0.264

72,246
0.197

Yes

72,246
0.241

72,246
0.207

Yes

72,246
0.262

Standard errors clustered at IDs are in parentheses

*** p < 0.01, ** p < 0.05, * p < 0.1

Note: All coeﬃcients are reported as β*100. All models include ﬁxed eﬀects (25-district
by monthly level) and the other variables listed in Table 1.1 and A1 in the model.

66

Table A6 Reputation Eﬀect (Time-varying)

– 25 District –

Dependent Variable = Log(Price)

Variable

Distance-Weight

Radius : 1km

Nearest 3 Schools

Radius : 1km

Nearest 3 schools

(1)

(2)

(3)

(4)

(5)

(6)

Top Univ.
Entrance Rate (t)

2.262***
(0.362)

1.306***
(0.371)

1.203***
(0.234)

1.184***
(0.235)

1.406***
(0.189)

0.850***
(0.254)

(t − 1)

(t − 2)

(t − 3)

(t − 4)

(t − 5)

(t − 6)

(t − 7)

(t − 8)

(t − 9)

2.953***
(0.338)

1.829***
(0.313)

1.972***
(0.212)

1.349***
(0.242)

1.765***
(0.219)

1.302***
(0.203)

1.869***
(0.286)

1.736***
(0.301)

1.627***
(0.187)

1.156***
(0.213)

1.168***
(0.254)

0.820***
(0.181)

1.758***
(0.299)

1.098***
(0.398)

1.272***
(0.193)

0.986***
(0.188)

1.154***
(0.200)

0.721***
(0.273)

1.653***
(0.274)

1.021***
(0.337)

0.691***
(0.141)

0.671***
(0.235)

0.557***
(0.149)

0.648***
(0.164)

1.279***
(0.335)

0.676***
(0.244)

0.735***
(0.165)

0.698***
(0.222)

0.602***
(0.170)

0.669***
(0.224)

0.713**
(0.229)

0.536*
(0.256)

-0.192
(0.241)

0.461
(0.378)

0.544
(0.332)

0.163
(0.298)

0.599**
(0.217)

0.607***
(0.165)

0.624***
(0.169)

0.606***
(0.160)

-0.0367
(0.161)

-0.446**
(0.191)

0.428
(0.227)

0.355*
(0.172)

-0.253
(0.160)

0.593**
(0.236)

0.410*
(0.249)

0.102
(0.200)

0.509**
(0.186)

0.660***
(0.237)

0.591***
(0.123)

0.454***

(0.15)

0.360**
(0.123)

0.485***
(0.159)

Sum(t ∼ t − 7)

Sum(t ∼ t − 9)

13.023***

8.671***

8.724***

7.256***

7.435***

6.013***

(1.729)

(1.825)

(1.064)

(1.090)

(1.125)

(1.166)

13.340***

9.494***

9.278***

9.278***

(1.935)

(2.089)

(1.196)

(1.244)

7.542***

(1.252)

6.6***
(1.315)

W/ All lags

Obs.
Adj R2

72,246
0.992

Yes

72,246
0.993

72,246
0.991

Yes

72,246
0.993

72,246
0.992

Yes

72,246
0.994

Standard errors clustered at IDs are in parentheses

*** p < 0.01, ** p < 0.05, * p < 0.1

Note: All coeﬃcients are reported as β*100. All models include ﬁxed eﬀects (25-district
by monthly level) and the other variables listed in Table 1.1 and A1 in the model.

67

Table A7 WTP for Educational Environments

– 25 District –

Dependent Variable = log(price)

House Fixed-Eﬀect Model

Time Varying Unobservable

VARIABLE

Coeﬃcient

WTP ($)

(1 unit)

(1 SD)

Coeﬃcient

WTP ($)

(1 unit)

(1 SD)

Top Univ
Entrance Rate

N-SAT
Score

Top High School
Entrance Rate

Private High school
Entrance Rate

Neighborhood
% BA or Above

Private
Academy

Student
per Class

Student
per Teacher

% Regual
Teacher

Distance to
Secondary School

Distance
Elementary School

PM10
Air Pollution

9.392***

3,024

2,155

8.425***

2,699

1,923

0.044

14

131

0.076***

23

226

-0.116

- 36

- 32

-0.085

- 26

- 23

0.096

0.003

0.146*

29

1

45

140

10

967

0.108***

0.308*

0.121**

33

95

37

157

1,091

808

-0.529***

- 162

- 375

-0.584***

- 179

- 414

-0.774***

- 237

- 317

-1.274***

- 389

- 520

-0.041

- 13

- 50

0.018

6

22

-3.80**

- 1,190

- 523

-9.59**

- 2,925

- 463

-0.729***

- 223

- 255

-0.882***

- 270

- 309

All coeﬃcients are report as β*100.

Willingness to pay is annualized at rate of 7% for mean house price of $438,656

*** p < 0.01, ** p < 0.05, * p < 0.1

68

Table A8 Housing Price Gap Decomposition

– 25 District –

Variable

Gangnam The other

Coeﬃcient

Decomposition Diﬀerence(Mean)

Mean value of Variable

($) Price Gap

% of

Top Univ
Entrance Rate

N-SAT
Score

Top High School
Entrance Rate

Private High school
Entrance Rate

Neighborhood
% BA or Above

Private
Academy

Student
per Class

Student
per Teacher

% Full-time
Teacher

Distance to
Secondary School

Distance
Elementary School

Mean Housing Price
(900 f t2)

2.02

0.60

8.425***

68,028

19.75%

109.45

100.07

0.076***

4,041

1.17%

2.71

2.09

-0.085

- 303

-0.09%

12.83

8.53

0.108***

2,626

0.76%

70.01

43.44

0.308*

46,470

13.49%

50.61

32.74

0.121**

12,340

3.58%

33.76

31.40

-0.584***

- 7,817

-2.27%

15.85

14.47

-1.274***

- 10,028

-2.91%

84.44

86.36

0.018

- 200

-0.06%

1.01

1.13

-3.80**

2,427

0.70%

0.33

0.32

-9.59**

- 799

-0.23%

$ 757,076

$ 412,610

SUM

116,785

33.90%

All coeﬃcients are report as β*100.
*** p < 0.01, ** p < 0.05, * p < 0.1

69

Table A9 WTP for the School Zone Adjustment without Weight

Area

All

Lower Score

Higher Score

WTP %

WTP($) / Year

Time

Median Mean

Median

Mean

Short-term
Long-term

Short-term
Long-term

Short-term
Long-term

0.32
2.87

0.48
3.43

0.32
- 0.05

0.69
3.27

83
548

119
775

65
- 660

187
807

- 0.27
- 5.85

- 0.33
- 5.98

- 132
- 2,846

- 152
- 3,278

Obs

34,818

22,315

12,503

Table A10 The Changes in Experimental Area without Weight

Housing Price ($1,000)

Top Univ. Entrance (%)

BA or Above (%)

Area

Time

Median Mean

SD

Median Mean

All

Lower Score

Higher Score

Initial
Short-term
Long-term

Initial
Short-term
Long-term

Initial
Short-term
Long-term

427.0
429.0
434.2

345.2
347.8
356.4

736.8
734.3
690.6

510.1
511.0
500.6

365.4
368.1
376.9

768.3
766.1
721.4

280.2
278.9
255.5

135.7
137.1
138.7

285.8
285.5
267.1

0.61
0.74
0.87

0.45
0.51
0.71

2.40
2.29
1.73

1.18
1.20
1.18

0.54
0.63
0.79

2.34
2.22
1.87

SD

0.95
0.88
0.66

0.31
0.37
0.26

0.51
0.55
0.59

Median Mean

SD

52.94

56.71

12.95

53.66

56.26

10.09

49.43

48.38

6.18

50.87

49.81

5.07

72.35

71.59

7.34

68.31

67.77

5.55

Obs

34,818

22,315

71,046

70

(a) Student/Class

(b) Student/Teacher

(c) Bank

(d) Culture

(e) Sports

(f ) Restaurant

(g) Population (%)

(h) Housing Price
(%)

Figure A3 (Cumulative) IRF

– Response = Housing Price (%) –

71

(a) BA (%p)

(b) Top Univ (%p) (c) Private Academy (d) Population (%)

(e) Employment (%)

(f ) Bank

(g) Culture

(h) Library

Figure A4 (Cumulative) IRF

– Impulse = Housing Price (1%) –

72

A.3 Structural Panel VAR

T(cid:88)

ΦYit =

ΓpYit−p + ΘXit + Djt + ξi + εit

(6)

p=1

This paper also adopts a structural panel VAR model, which is presented in equation

(6) to consider the simultaneous relationships among variables. Unlike the panel VAR

model (equation (5)) in Section 1.5, the model includes the parameter matrix Φ deﬁn-

ing the contemporaneous relations among dependent variables in the system. Since the

structural model requires restrictions for the parameter speciﬁcation, to apply restric-

tions such as variable ordering, the model reduces the number of endogenous variables

and includes the key variables of this research. Speciﬁcally, the structural panel VAR

model includes the neighborhood composition (percent of BA or above degree holders
in the 40∼49 age group), top university entrance rate, log of housing price, popula-
tion growth, and log of employment (full- and part-time employment) as endogenous

variables and controls for all the other variables as exogenous variables.60

Note that in the structural model that has simultaneous relationships among

endogenous variables, the variance-covariance matrix of the shocks is not likely to be

diagonal. In other words, a shock is correlated with shocks of other variables and they

are not independent of each other. Thereby, we need to make them orthogonal to isolate

shocks uncorrelated with other shocks. In this process, this paper uses the following

restrictions. The variable ordering this paper assumes is: top university entrance rate

- neighborhood composition - population growth - employment - housing price. That

is, we assume that the variable on the left side contemporaneously aﬀects variables

on its right side, however, the opposite is not true. The increase in the top university

entrance rate of a school can attract more educated people to the neighborhood around

the school at the same time. Given that, however, it is rare for households to move

during the academic year when their children take the university entrance exam, this

paper assumes the neighborhood composition cannot contemporaneously aﬀect school

performance. Also, based on the impulse-response analysis in Section 1.5, the housing

60 That is, the number of students per class and per teacher, the number of private academies, libraries,
supermarkets, culture related stores, shopping stores, clinics, restaurants, sport facilities, banks, and grocery
stores around a house. Also, the model controls distances from a house to the nearest subway station,
department store, supermarket, elementary school, middle school, and high school as exogenous variation.

73

price is treated as the most endogenous variable among the ﬁve variables. For the

estimation, the ﬁrst diﬀerenced variables are used and IV-GMM is adopted as in Section

1.5.

(a) BA : Top Univ (%p)

(b) BA : BA (%p)

(c) BA : House Price (%)

(d) Top Univ : BA (%p)

(e) Top Univ : Top Univ

(f ) Top Univ : House Price

Figure A5 Structural Panel VAR, (Cumulative) IRF

Figure A5 shows the results from the structural panel VAR adopting Cholesky

decomposition with the ordering, and the impulse response functions are generated

from the orthogonalized shocks. According to the results, compared to the results in

Section 1.5, the duration and the size of the eﬀect from a shock becomes shorter and

smaller compared to the results presented in Figure 1.5. The diﬀerences between these

two results can be caused by two sources. First, the diminished number of endogenous

variables in the system VAR could simplify and eliminate relationships among variables.

Second, considering the contemporaneous relation among endogenous variables, which

is represented by the parameter matrix Φ, can lead to the diﬀerences.

To ﬁgure out the factor driving the diﬀerences, this paper also estimates the

equation (5) while using the same variables as with the structural panel VAR model.

74

That is, for the estimation, ﬁve variables – in the structural panel VAR model – are

used as endogenous variables, and all the other variables are controlled as exogenous

variables. However, the model does not consider the contemporaneous relations among

endogenous variables. Thus, the model is a non-structural panel VAR model which

uses ﬁve endogenous variables.

(a) BA : Top Univ (%p)

(b) BA : BA (%p)

(c) BA : House Price (%)

(d) Top Univ : BA (%p)

(e) Top Univ : Top Univ (%p) (f ) Top Univ : House Price (%)

Figure A6 Panel VAR, (Cumulative) IRF

– (Impulse) : (Response) –

Figure A6 shows the results of the model. Note that the estimated impulse-

response functions are quite similar to the results in Figure A5. This result implies

that the diﬀerences – the shortened durations and the smaller size of eﬀect from a shock

– are mainly caused by the diminished number of endogenous variables that eliminate

the relationships among amenities and attributes of a house.

75

Chapter 2. Impact of Resale Price Maintenance (RPM) Reg-

ulation on the Book Market of Korea

2.1 Introduction

Resale Price Maintenance (RPM) regulations in book markets are now in force in many

countries in Asia and Europe, while it was abolished a long time ago in the United

States and the United Kingdom.61 Although RPM is not allowed generally, many Asian

and European countries made an exception for their book markets, considering a book

to be unique goods, which is diﬀerent from other goods in the market. They allow

RPM in a book market under the policy purpose of protecting a variety of cultures

and knowledge. The proponents of RPM regulation for books argue a proper price

should be guaranteed for a book – especially if the book is not popular – to survive in

the market, and it is justiﬁed to protect the value of cultural variety and knowledge.

Even though “A Book is Diﬀerent” in terms of its cultural and knowledge per-

spective, “A Book is the Same” in that it cannot be an exception of the law of demand

and supply in the market. First of all, a decent margin of a book by ﬁxing the retail

price does not guarantee a higher proﬁt for the author and publishing company. Be-

cause the higher price will reduce the demand for the book, ﬁxed retail price can help

the publishing company (the price setter under the RPM regulation) only if they can

set the optimal price by accurately forecasting the market demand when they publish

a book. However, it is generally known that forecasting the demand for a new book

is hard. Unexpected new publications of similar books and social trends make the

demand for a book highly volatile.

Second, a higher retail price also does not mean a higher wholesale price for

the publishing company and higher royalty for the author of a book. If some retail-

ers have overwhelming bargaining power over the publishing companies, resale price

maintenance regulation can be used as a method to help increase the proﬁt of market-

dominant retail ﬁrms at the cost of diminishing consumer surplus, proﬁt of publishing

61 Fifteen OECD countries, ten of which EU members, have regulation for ﬁxing the price of books. In
the Unites States, the resale price maintenance was the per se illegal under 1 of the Sherman Act, 15 U.S.C.
1. However, a decade ago, the U.S. Supreme Court overturned a nearly century-old precedent when it issued
its Leegin opinion. After that, the resale price maintenance is considered to be restrictedly allowed if there is
a reasonable reason. In the United Kingdom, the RPM for a book took eﬀect in 1900 and abolished in 1997
by the Restrictive Practices Court

76

companies and authors. De los Santos and Wildenbeest (2017) analyze the ﬁxing

e-book price case of the United States. In 2010, ﬁve of the six largest publishing com-

panies concurrently adopted the agency model where publishers set retail prices. They

empirically analyze the eﬀect of vertical price restraints on the price of e-books. They

show that e-book prices were 18 percent higher at Amazon and 8 percent higher at

Barnes & Noble when publishing companies adopt the agency model which controls

the retail price. However, it is important to note that the higher retail price during

the agency period did not enhance the proﬁt margin of publishing companies. On the

contrary, publishers’ proﬁt margins from e-books decrease under the agency period

mainly due to higher commission fee for the retailers.

In most cases, online companies sell books at lower prices than local oﬄine stores

because they have more eﬃcient cost structures compared to the small oﬄine book-

stores. Online ﬁrms operate Internet servers and large-scale distribution warehouses

instead of oﬄine stores, and therefore decrease the average cost by the economy of

scale. On the other hand, oﬄine stores should pay rent for the store and a substantial

cost for displaying the books and operating the store. Also, note that major online

ﬁrms can buy books from publishing companies at lower prices than oﬄine book-stores

because big online ﬁrms have higher bargaining power based on their dominant market

shares.62 According to a recent survey (The Survey on the Book Publishing Industry)

which is investigated by the Statistics Korea, online companies – of which sales are

over 10 billion () – buy a book from publishing companies on average at 63.6% of

the list price. However, the small oﬄine stores (sales below one hundred million ())

buy books from publishing companies on average at 70.1% of list prices. In Korea,

There were 3,429 oﬄine book-stores in 2005; however, the number decreased to 2,116

(– 38.3%) in 2015 while top online ﬁrms have continuously expanded their market

shares.63 The number of small oﬄine book stores was expected to decrease more given

the price gaps between online and oﬄine stores based on the cost eﬃciency and the

irreversible trend of e-commerce.

In these circumstances, on the purpose of protecting cultural variety and helping

small publishing companies and oﬄine local bookstores, the National Assembly of

62 For example, the Kyobo Book Centre (No.1 book company in Korea) alone accounts for 25% of the

total book market share in terms of sales.

63 The top 7 online ﬁrm’s market share was over 50% in 2012 and have been continuously increasing.

77

Korea legislated the new RPM regulation (reinforcing the price regulation on book

markets), which took eﬀect in November of 2014, despite oppositions from consumers.

Table 2.1 shows what was changed in the RPM regulation from November 2014.

Table 2.1 Amendment of RPM Regulation

RPM regulation
(Before 11.2014)
Newly-published book64

(published within the last 18 months)

Up to 19% from list price
• Direct discount up to 10%
• Indirect discount up to 9%
(coupon, points, other beneﬁts)

Regulation
Coverage

Discount

Changing
List Prices

New RPM regulation

(After 11.2014)

All books

Up to 15% from list price
• Direct discount up to 10%
• Indirect discount up to 5%
(coupon, points, other beneﬁts)

Publishing companies can

change the list price only after

18 months pass from the

new-release

Before the New RPM Regulation

Before the new regulation was in eﬀect, in most cases, a local oﬄine store had sold its

book at its list price while an online ﬁrm had usually sold it at a 10% discounted price

(maximum discount rate by the regulation) for newly-published books. However, the

price gap between online and oﬄine stores was more prominent for books published

more than 18 months ago. No price regulation for the non-newly-published books65 and

price competition among online ﬁrms make a prices fall to its marginal cost. Moreover,

online users can easily compare the online ﬁrm’s price because search engines such as

NAVER.com and DAUM.net oﬀer the price comparison among all online book-stores66

and it makes search-cost almost zero. If users enter the title of a book, they show all

prices of each online stores for comparison. Because of the above reasons, prices of

non-newly-published books were usually much lower in online stores than local oﬄine

stores. That is, the price gap was just 10% for a newly-published book; however, in

case of a non-newly-published book which is published more than 18 months ago, the

64 According to the regulation, a book is categorized as a newly-published book if the book was published
in recent 18 months. However, if a book is published more than 18 months ago, it had not been under the
RPM regulation until 11. 2014

65

66

In this paper, a non-newly-published book means a book which is published more than 18months ago.
In Korea, NAVER and DAUM have more market share than Google. Each portion was 80% (NAVER),

13% (DAUM) and 5% (Google) in 2015.

78

price gap between online and oﬄine store was larger than 10% until the new RPM

regulation took eﬀect.

After the New-RPM Regulation

The most critical change in the reinforced new RPM regulation is that the books which

were published more than 18 months ago are newly included in the new regulation

coverages. Many countries are enforcing resale price maintenance regulation on books.

However, no country has price regulation for books published a long time ago67 because

price regulation for all kinds of books can cause some problems.

First of all, regulating the price of all books can make serious ineﬃciency in the

inventory disposal process. It was possible under past RPM regulation to lower the

price of unpopular books if a book store wanted to sell them. However, this is not

the case under the new regulation system. Now, the bookstore should wait until the

publishing companies change the list price of a book. A worse problem is that to change

the list price, publishing companies should recall all books from the bookstores and

re-distributes again after changing the list prices. Before the new RPM was in eﬀect,

the inventory disposal was processed eﬃciently through online book stores. The cost

for changing the price (menu cost) of a book at online stores is almost zero because

changing the price on an Internet server is not costly and this process can be done

systematically by a program accoroding to the situation of inventory stock. However,

under the new RPM system, the market no more can use the eﬃcient inventory disposal

process of online ﬁrms.

That is, the new RPM regulation restricting prices of all books increase inventory

disposal cost. Because of the increased burden of inventory controls, it is also possible

that bookstores become reluctant to buy and retain unpopular books from publishing

companies. Then, it is against the purpose of the new policy to support the survival

of unpopular books. Moreover, if the new regulation increases menu cost, inventory

management cost and logistical cost, retailers (online ﬁrms and small oﬄine stores) and

publishing companies should share and divide the extra burdens, and it is likely that

they divide the extra burden depending on the bargaining power between a publishing

67 Even in the countries enforcing more restrict resale price maintenance on book market such as France
and Germany, the RPM regulation is applied for a newly-published books for a limited time (two years in
France and 18 months in Germany)

79

company and an online store or an oﬄine store. Thus, one might expect that publishing

companies and small oﬄine stores might be not in a favorable position in this process

when considering the dominant market share of large online ﬁrms.

The proponents of new RPM regulation anticipate that some portions of online

ﬁrm’s customers move to local oﬄine stores because of a diminished price gap between

online and oﬄine stores and it would help the business of small oﬄine local stores.

However, converted demands from online stores to oﬄine stores could be smaller than

expected for some reasons.

First, even though the regulation restricts the indirect discount up to 5% of

a list price, online ﬁrms can legally avoid the law by ﬁnding the blind spot of the

regulation. According to the report of the Publication Industry Promotion Agency of

Korea (KPIPA, 2016), most online ﬁrms are using various alliance credit card discounts

which are not under the price restriction. For example, the YES24 adopts more than

30 alliance credit cards, and if customers use one of these credit cards, they can get

more discounts over the regulation limit. Moreover, some indirect discounts are hard to

be measured accurately; thus it is diﬃcult for the authorities to regulate it. Also, lots

of other beneﬁts such as gifts and coupons for other goods and services are not under

the price restriction.68 Because of these problems, according to a recent publication

industry survey administered by KPIPA, more than 60% oﬄine stores responded that

there should be more restrictions about indirect discounts of online stores including

alliance credit card discounts.

Second, most online ﬁrms are also operating oﬄine stores in addition to their

online stores. They are usually large-scale stores and also located in a good place such

as the central business district of a city. Thus, it is possible that a large portion of

converting demands from online to oﬄine can be absorbed by the large oﬄine stores

operated by online ﬁrms. After the new regulation, the proﬁt of online ﬁrms signif-

icantly increases69 and they have expanded their business aggressively opening many

large-scale oﬄine stores based on their increased proﬁts. Kyobo Book Centre (No.1

68 According to a recent news report (Large Online Book Stores Out of RPM Regulation, 4.14.2017, Seoul
Economy, http://www.sedaily.com/NewsView/1OEN7SNHXB/GK01) a person obtained various coupons of
which total value reach 50% of a book price after he bought a book. However, surprisingly, all gifts and
discount coupons he obtained on the day did not violate the indirect discount regulations of the new RPM
regulation.

69 The empirical analysis about this point is in Section 2.3.

80

book store) has opened 17 more oﬄine stores (among their 39 oﬄine stores) after the

new RPM regulation took eﬀect.70 Aggressive oﬄine entry of large online companies

could make the converted demand from online to oﬄine small local stores more diﬃcult.

The bottom line is that various indirect discounts out of law’s restriction, large-

scale oﬄine stores operated by online ﬁrms of which numbers are sharply increasing

recently make it harder to achieve the policy purpose of new RPM regulation. However,

paradoxically, all the above things are caused by the new RPM regulation itself. Note

that online companies having grip on the book market can avoid price regulations by

using various methods explained above. Due to the increased price margin of a book,

large online ﬁrms’ proﬁt ratio and the amount of proﬁts are enhanced signiﬁcantly after

the new regulation took eﬀect as I show the empirical evidences in the following Section

2.3. As a result, the market share of top 7 online ﬁrms even increase rapidly after the

new regulation took eﬀect even though the new regulation intended the opposite results.

At the same time, the amount of sales of publishing companies and small oﬄine

stores have decreased signiﬁcantly after the new regulation. In addition, after the new

regulation, the number of newly-published book which can be used as the index for

cultural variety is decreased by 7% annually after de-trending. Even though the RPM

regulation took eﬀect under the purpose to preserve cultural variety and knowledge by

supporting small publishing companies and book stores to survive in the market, many

empirical evidence imply the opposite results.

Also, the new RPM regulation could aﬀect the average price of books, transaction

quantity, and consequently the consumer surplus. Cho (2015) shows the new regulation

increases the price of best-seller books by 16.2% and steady-seller books by more than

40%. The study also estimates that the new regulation results in the decrease of book

transactions by 12.1%. In terms of welfare change, he estimates the consumer surplus

decreases more than 13 billion (). Among them, 9 billion () is transferred to online

ﬁrms and oﬄine stores lose 3.5 billion () owing to the new regulation.

This paper proceeds as follows: Section 2.2 describes and analyzes the data

illustrating the changes in the whole book market situation before and after the new

70 The Aladin has opened 17 out of 37, the Yongpung has opened 9 out of total 30, and the Bandi & Lunis
have opened 4 out of total 14 oﬄine stores during the same short period. The YES24 (the second largest
online book store) opened their ﬁrst oﬄine store last April. 2016 and have opened four more oﬄine stores
until now.

81

RPM regulation. More speciﬁcally, Section 2.3 investigates the eﬀect of new regulation

on each ﬁrms based on ﬁrm-level dataset and various surveys. Section 2.4 concludes

and summarize the empirical analysis of this paper.

2.2 Impact of RPM Regulation on the Book Market

2.2.1 Market Size and Online Share

Table 2.2 The Book Market of Korea

(1)

Market
Size 1

(2)

(3)

Expediture
/ Household

Expenditure

/ Person

(4)

Market
Size 2

(5)

On-line

(6)
(7)
Online Share

(8)=(1)-(5)

Oﬄine

Market Size

(5)/(1)

(5)/(4)

Market Size

26,288
27,331
29,471
29,465
29,498
31,963
33,530
32,600
34,738
33,377
33,959

-4% 12,066
-7% 13,108
0% 14,614
0% 15,147
-8% 15,502
-5% 16,937
3% 17,939
-6% 17,484
4% 18,586
-2% 17,710
18,607

-8% 51,299
-10% 53,575
-4%
58,074
58,428
-2%
58,761
-8%
64,006
-6%
67,664
3%
-6%
66,116
70,815
5%
68,559
-5%
70,107

-4%
-8% 25,385
-1% 29,438
-1% 25,397
-8% 24,133
-5% 28,504
2% 27,258
-7% 27,244
3% 25,810
-2% 31,461
25,810

13,406
-14% 11,512
16%
12,804
5%
11,961
-15% 12,728
12,743
5%
11,691
0%
6%
10,298
8,752
-18%
7,442
22%
6,277

16%
51%
-10% 42%
43%
41%
43%
40%
35%
32%
25%
22%
18%

7%
-6%
0%
9%
14%
18%
18%
19%

12,882
45% 15,819
43%
16,667
17,503
47%
16,770
53%
19,220
45%
21,839
43%
38%
22,302
25,986
34%
25,935
24%
24%
27,682

-19%
-5%
-5%
4%
-13%
-12%
-2%
-14%
0%
-6%

Year

2016
2015
2014
2013
2012
2011
2010
2009
2008
2007
2006

Table 2.2 shows the estimated size of the book market and online ﬁrms’ market

share by year. The amount of market size in column (4) is estimated and published by

the Korean Publishers Association (KPA). According to KPA, they use the following
formula to calculate the total market size: the average price of books × the number
of calculation for a year × 2, by every 12 categories and the summation of the total
amounts becomes the estimated market size. Thus, it is multiplication of Table B2

and Table B3.71 On the other hand, I estimate the total market size presented in the

column (1) by using the two statistics published by Statistics Korea. It publishes the

average expenditure on books by the household size and also publishes the number

of household by the number of household member. Thus, we can estimate the total

amount of expenditure on books by using the two statistics. The column (2) shows

the average amount of expenditure on books per household for a year, which is also

published by Statistics Korea. The column (3) comes from the calculation that the

estimated market size in the column (1) is divided by the total population of Korea.

71 However, the calculation results are diﬀerent from column (4), because the ﬁnal amount is decided after
an expert meeting about the total market size based on the calculation. The ground for the multiplication of
2 in the KPA’s formula is based on the rough estimation that the total amount of transaction for the books
published more than one year ago is similar with that for a newly-published book during the year.

82

Thus, it means the average expenditure for books per person. The column (5) shows

the total amount of transactions for books at online stores which comes from the Online

Shopping Survey by Statistics Korea. The column (6) and (7) show the market share

of online stores by using the ratio (5)/(1) and (5)/(4).

According to Table 2.2, the total market size, the average expenditure on book

per household and per person are all decreased after the new RPM regulation. Although

all of them were on the decreasing trends, the new regulation could not change the trend

and the decreasing trend even speeds up. The online sales decreased just after the new

regulation in 2015, however, it rapidly rebounded in 2016. As a result, the market

share of online exceeded 50% for the ﬁrst time while it stayed around 40% before the

new regulation. On the other hand, the amount of oﬄine sales continued to decrease.

It even decreases by 19% in 2016. Based on the above statistics, it is hard to say some

demands are converted to oﬄine stores as the new RPM regulation intended. Even

after the new regulation took eﬀect, both increasing trend of online and decreasing

trend of oﬄine have accelerated. Figure 2.2 shows the trends of total book market size,

the number of online sales and the online market share.

Table 2.3 The 7 Biggest Online Firms

(1)

(2)

(3)

(4)

(5)

Top 7 Firms

Sales72

Top 7 Market Share
Maket 1 Market 2

Top 7 Firms

Proﬁt73

All the Other
Firms Sales

16,979
15,791
15,954
15,934
15,728
15,685
15,401
14,326

8%
-1%
0%
1%
0%
2%
8%

65%
58%
54%
54%
53%
49%
46%
44%

62%
54%
63%
65%
55%
57%
53%

4,826
4,302
3,694
3,520
3,492
3,362
3,377
3,387

12% 9,310
16% 11,541
13,517
5%
13,531
1%
13,770
4%
16,278
0%
0%
18,129
18,274

-19%
-15%
0%
-2%
-15%
-10%
-1%

Year

2016
2015
2014
2013
2012
2011
2010
2009

Table 2.3 shows the top 7 online ﬁrms’ amount of sales, proﬁt (= sales - cost)

and their market share by year. In 2015, the sales of top 7 ﬁrms decrease a little (-1%),

however, the amount of proﬁt increase by 16% owing to the increased price margin.74

72 The top 7 (in terms of sales) ﬁrms in the book market is 1) KYOBO 2) YES24 3) ALADIN 4) INTER-

PARK 5) YOUNGPUNG 6) BANDI & LUNIS 7) LIBRO.

73 This amount of proﬁt is sum of 6 ﬁrms except INTERPARK. Because INTERPARK is operating 4
business parts and they have only published the sales of each business, I could not ﬁgure out the cost of
INTERPARK book business part.

74 Table 2.11 shows the enhanced proﬁt ratio of top 7 online ﬁrms. Before the new regulation, there was
no price regulation for the books which were published more than 18 months ago. Price competition among

83

Figure 2.1 Market Size and On-line Share

84

In 2016, their amount of sales rebounded and increased by 8% while the amount of

proﬁt increase 12%. If we check the top 7 ﬁrms’ market share and all the other ﬁrms’

sales, the Top 7 online ﬁrms’ market share rapidly increases after 2015 and it was over

60% in 2016. On the contrary, all the other ﬁrms’ sales decrease by 15% and 19% after

2015. In summary, the degree of market concentration is aggravated after 2015.

2.2.2 Variety of New Book Publications

In addition to supporting the business of small publishing companies and oﬄine book

stores, protecting the cultural variety and the value of knowledge was another impor-

tant aim of the new RPM regulation. The variety of newly-published books can be an

index to evaluate the eﬀect of the reinforced RPM policy on a cultural variety. Table

2.4 shows the number of newly-published books by 11 categories by year. This paper

uses the book registration dataset from the National Library of Korea (the copyright

library of Korea)75 in which all new publications are registered.

Table 2.5 shows various regression results by using the data of Table 2.4. For

the analysis, the equation (7) is used as the baseline model to estimate the eﬀect of

new RPM regulation on the number of newly-published books.

log(N ew P ublication)it = αt + β1t RP M 15t + β2t RP M 16t + δt log(P riceit) + ξit (7)

The dependent variable is the log of the number of newly-published books of book cat-

egory i at time (year) t. RPM15 and RPM16 represent time dummy variable for year

2015 and 2016 respectively. The model also uses the average price of books of the cat-

egory i at time (year) t as a control variable, because the average price of books could

online book stores and zero search cost drove the price to its marginal cost. However, resale price maintenance
regulation restricting price competitions among them enhanced their price margin signiﬁcantly.

75 According to the library law, all newly-published books should be registered on the National Library
of Korea within 30 days from the new publication. There is another dataset about statistics of the number
of new books published by the Korean Publishers Association (KPA). However, the statistics from KPA only
included the books which were registered on the National Library of Korea through KPA’s system. Because
of direct registration to the National Library of Korea, KPA dataset does not include the all-new publication
book. Table B1 in the appendix shows the data from KPA. Publication Industry Promotion Agency of Korea
(KPIPA) also compiles and publishes the statistics about newly-published books. However, it extended its
data source in 2015 and 2016.
In 2015, books recorded in Yongpung and Bandi & Lunis are included as
the data source, and they again added Aladin in 2016 to make the new publication statistics and did not
retroactively compile the previous data with the same standard. Thus, we cannot compare the statistics
comes from the broadened data source with the previous statistics. KPA also broadened their data source to
make new publication statistics from 2016 and also did not retroactively compile the previous statistics based
on the same data source.

85

Table 2.4 The Number of Newly-Published Books by Categories

Category

2009

2010

2011

2012

2013

2014

General
Philosophy
Religion
Social Science
Science
Engineering
Art
Language
Literature
History
Children

1,992
1,262
3,906
13,248

964
6,966
2,471
2,122
10,042
2,085
12,417

1,829
1,390
3,595
13,158

981
7,043
2,521
2,420
10,662
2,202
12,154

1,870
1,395
3,788
13,542

968
7,666
2,556
2,536
10,501
2,145
12,722

1,610
1,665
3,639
12,717

855
7,441
2,638
2,422
10,738
2,128
12,062

1,628
1,892
3,503
13,307

960
8,068
2,669
2,582
11,240
2,221
10,133

1,994
2,074
3,459
14,559
1,064
8,149
3,095
2,685
13,502
2,459
9,032

2015

1,946
2,084
3,169
13,612
1,117
9,242
3,200
2,216
12,591
2,364
7,872

2016

2,545
1,505
3,132
13,141

900
8,871
2,898
2,114
12,162
2,208
7,715

Total

57,475

57,955

59,689

57,915

58,203

62,072

59,413

57,191

Data : National Library of Korea

aﬀect the number of new publication. Lastly, ξit denotes the omitted or unobservable

attributes aﬀecting the number of newly-published books of category i at time t. By

using two consecutive years of data and additionally assuming that the unobservable

attributes of a house and the implicit prices for time-varying variables do not change
over two years (βt(cid:48) = βt = β, δt(cid:48) = δt = δ, ξit(cid:48) = ξit, (t(cid:48) > t)), the equation (7) can
be re-written as the following model by ﬁrst diﬀerencing:

First-diﬀerencing Model

∆log(N ew P ublication)i = ∆α + β1 ∆RP M 15 + β2 ∆RP M 16 + δ ∆log(P rice)i (8)

This study also ﬂexibly allows the unobservables to vary within the two consec-

utive years given the assumption that it follows the ﬁrst-order Markov process:

ξit(cid:48) = γ ξit + ηit(cid:48)

(t(cid:48) > t)

(9)

here, γξit is the expected value of the unobservable attributes at time t(cid:48), and ηit(cid:48) is
the stochastic innovation in the unobservable attributes. Note that this study assumes
E[ηit(cid:48)|It] = 0, where It denotes the available information at time t. Then, equation (7)
can be re-written as follows by using the Markov process assumption.

86

Time-varying Unobservable Model

log(N ew Books)it(cid:48) = αt(cid:48) + β1t(cid:48) RP M 15t(cid:48) + β2t(cid:48) RP M 16t(cid:48) + δt(cid:48) log(P riceit(cid:48)) + ξit(cid:48)

= αt(cid:48) + β1t(cid:48) RP M 15t(cid:48) + β2t(cid:48) RP M 16t(cid:48) + δt(cid:48) log(P riceit(cid:48)) + γξit + ηit(cid:48)

= αt(cid:48) + β1t(cid:48) RP M 15t(cid:48) + β2t(cid:48) RP M 16t(cid:48) + δt(cid:48) log(P riceit(cid:48))

+ γ[log(N ew Books)it − αt − β1t RP M 15t − β2t RP M 16t
− δt log(P ricet)] + ηit(cid:48)

= (αt(cid:48) − γαt) + γlog(N ew Books)it + β1[RP M 15t(cid:48) − γRP M 15t]
+ β2[RP M 16t(cid:48) − γRP M 16t] + δ[log(P riceit(cid:48)) − γlog(P riceit)] + ηit(cid:48)
(10)

Table 2.5 Eﬀect of RPM on the Number of New Publications

VARIABLES

Constant

RPM15

Dependent = Log(Number of Newly-Published Books)

(1)
FD

.0241*
(.0125)

(2)
FD

.0307**
(.0131)

-.0610**

-.0691**

(.0305)

(.0308)

(3)

FD-IV

.0277
(.0183)
-.0649*
(.0353)

(4)
TV

.1144
(.1013)
-.0599*
(.0312)

(5)
TV

.1180
(.1354)

-.0675**

(.0309)

(6)

TV-IV

.1047
(.1113)
-.0600*
(.0336)

RPM16

-.1425***

-.1574***

-.1501***

-.1401***

-.1541***

-.1406***

Log(Price)

γ

(.0466)

(.0474)
-.2636
(.1787)

(.0564)
-.0767
(.4404)

Observation
Adj. R2

77

.1149

77

.1407

66

.1281

(.0466)

.9889***
(.0122)

77

.9889

(.0474)
-.2425
(.1792)
.9916***
(.0136)

77

.9892

(.0521)
.0514
(.2004)
.9896***
(.0139)

66

.9879

Standard errors clustered at IDs are parentheses

*** p < 0.01, ** p < 0.05, * p < 0.1

Table 2.5 shows the various estimation results by using above models. Column

(1), (2) and (3) show estimation results from the ﬁrst diﬀerencing model (equation (8)),

and column (4), (5) and (6) come from the time-varying unobservable model (equation

(10)). The column (3) and (6) adopt two-stage least square (2SLS) method to control

a simultaneous relation between price and new publication of books. I use the average

price of books in the previous year and all other predetermined variables as instruments

for the average price of books. It is reasonable that higher average price leads to more

publication. In column (6), the coeﬃcient changes to positive but is not signiﬁcant.

87

According to various panel regression results, the number of newly-published books
decreases by 6∼7% in 2015 after de-trending.
In 2016, the accumulated amount of
decrease reaches 14∼15% implying the amount of decrease in 2016 is about 8∼9%.
Based on the above results, it is hard to insist that the new RPM regulation help to

enhance the number of new publication as an index for a variety of knowledge and

culture. According to the result of the column (6), the time-varying unobservable

model using instrument variables, the number of newly-published books decreased by

6% in 2015 and 8% more in 2016.

2.2.3 Price of Books

Another controversial issue of new RPM regulation is its eﬀect on the price of books.

The proponents of new RPM regulation insist that the new RPM regulation help to

stabilize the rapid increase in the average price of a book. The ground for this argument

is that because of high discount rate of online ﬁrms (the result from price competition

among online ﬁrms and zero search cost) especially for a book out of RPM regulation

(a book which is published more than 18 months ago), the publishing company could

increase the list price to oﬀset the discount rate of online ﬁrms.

If the publishing

companies sell their product, for example, at 50% of the list price, they can earn more

by increasing the list price of a book.

However, there are some problems with this argument. First, the actual selling

price of a book is tied to its list price for 18 months after the initial publication. If

a publishing company increase its list price, it could increase its earnings per book;

however, the selling quantity can be decreased by the increased list price and the RPM

regulation applied to a newly-published book. Accordingly, increasing the list price

does not guarantee a higher proﬁt for publishing companies. Second, the important

thing for a customer is not a list price but the actual selling price of a book. Even

though the new regulation can help to stabilize the list price, it does not mean the

stabilization of the actual selling price.

In the short-run, we can guess the new RPM regulation may increase the actual

price of a book at online stores especially for a book which is published more than 18

months ago because the book is newly included in the new RPM regulation coverage.76

76 According to Table 2.1, there was no price regulation for a book published more than 18 months ago

88

According to Cho (2015) that analyzes the change of actual price at online stores, the

introduction of the new RPM regulation increases the price of best-selling books by

16.2% and steady-selling books by more than 40% on average.77

However, it is not simple to show the eﬀect of new regulation on the actual price

of books in the long-run. First, the new RPM regulation can slow down the increasing

rate of list price of books, thereby the actual price ﬁxed with the list price can be lower

in the long-run compared to the counter-factual situation that the new regulation was

not introduced and a book is sold at, for example, more than 20% discounted price from

the list price. Second, according to the new regulation, only the publishing companies

can change the list price after 18 months from the initial publication.78 Thus, we need

to compare the newly-set-price by a publishing company with the counter-factual-price

under no price regulation. In other words, the new RPM regulation can aﬀect both

the list price of a new book and the newly-set-price of a book published more than

18 months ago by publishing companies, accordingly makes it hard to evaluate the

eﬀect of the new regulation on the overall price level of books. Moreover, I could not

obtain information about the actual selling price of books – which were published more

than 18 months ago – at online stores before the new regulation took eﬀect, which is

essential to analyze the eﬀect of new regulation on the price of non-newly-published

books.

Thus, in this paper, I focus on the list price of newly-published books. Because

this kind of books are already under the price regulation, the amount of change in

list price mean the amount of actual price change. I use the book registration dataset

published by the National Library of Korea. This dataset include the information of

before the new RPM regulation took eﬀect. However, the maximum discount rate of those books becomes
15% under the new RPM regulation.

77 The diﬀerence in the increasing rate between best-sellers and steady-sellers comes from the diﬀerence
in their compositions. Best-sellers are mostly composed of newly-published books while non-newly-published
books – which are published more than 18 months ago – consist of steady-sellers. Because there was no price
regulation for a non-newly-published book before the new regulation took eﬀect, those books had room for a
further price increase.

78

It is clear that this system is not eﬃcient in that it will increase menu cost and logistical cost. Before
the new regulation took eﬀect, an online store can change the price of a book ﬂexibly according to the market
situation and their stocks. Because they can quickly change the price of a book by changing the price on the
screen, the menu cost was almost zero. However, after the new regulation, the publishing company should
print again the cover showing the new price and collect all books in the market and re-distribute the book with
a new price. Moreover, online stores could frequently change the actual price to cope with a volatile market
situation (e.g., the new publication of similar books and their amount of stocks. However, the publishing
company usually have less information about the market, and because the newly-set list price should be ﬁxed
for 18 months again, they cannot deal with the unexpected situations by ﬂexibly changing the price.

89

each book’s title, category, author, publishing company, edition, publication year, page,

size, the physical characteristics of a book such as (include color picture, partial color

and so on), the list price, ISBN, the number of books and supplements information

(CD-ROM, DVD, MP3, map, picture and so on) from 2009 to 2017.

Table 2.6 The Average List Price of Books by Category

Category

2009

2010

2011

2012

2013

2014

2015

2016

2017

General
Philosophy
Religion
Social Science
Natural Science
Engineering
Art
Language
Literature
History
Children

17,514
17,320
12,019
18,750
18,523
23,346
19,698
14,519
10,142
17,861
6,546

19,142
16,324
12,566
19,095
19,699
24,716
19,108
14,146
10,074
17,982
7,389

18,597
17,607
12,840
20,276
20,494
28,013
19,941
14,597
10,594
18,627
6,995

16,833
18,277
13,149
20,300
21,796
26,649
19,342
15,236
10,765
18,436
7,451

18,676
17,310
13,840
20,550
20,548
28,032
26,655
14,588
10,794
18,499
7,579

19,586
17,059
14,289
21,818
21,342
27,002
23,610
16,021
11,537
18,888
8,003

19,435
16,715
14,367
20,878
21,345
26,309
24,941
16,756
10,847
18,667
8,042

18,656
17,518
14,056
20,865
20,773
26,517
26,395
15,445
11,028
19,761
7,846

20,041
17,049
14,632
20,886
20,268
23,560
26,725
15,790
10,065
17,300
8,274

Total

13,525

13,662

14,165

14,461

15,072

15,873

15,599

15,766

15,741

Data : National Library of Korea

Table 2.6 shows the average list price of newly-published books which use the

book registration data published by the National Library of Korea. Based on the data

represented in Table 2.6, this paper analyzes the eﬀect of new RPM regulation on

the list price of newly-published books. Table 2.7 shows the various regression result

according to category classiﬁcation of the National Library of Korea. The full table

including all the other explanatory variables and ﬁxed eﬀects is represented in Table B4

in the appendix. The column (1) presents the regression result for all books. After the

new regulation took eﬀect, the list price decrease by 4% (=11.83%/3) annually after

eliminating the eﬀect of time trend. The eﬀects of new RPM regulation are somewhat

diﬀerent depending on the book category, however, in general, the new regulation

decreased the list prices.

Table 2.8 shows the results when we consider the case that there was a disconti-

nuity or a jump in the list price when the new regulation took eﬀect and the time trend

of the list price is changed after the new regulation. “Time Trend” variable shows the

time trend of the list price before the new RPM regulation by each book category.

The coeﬃcient of “RPM15” represents how much the list price jumped in 2015 when

the new regulation took eﬀect. The coeﬃcient of “RPM Trend” means the amount

90

Table 2.7 Eﬀect of RPM on List Price

Dependent Variable = Log(Price per Book)

(2)

(3)

(4)

(5)

(6)

(7)

Non-Children

Children

General

Religion

Philosophy

Literature

2.795***
(0.0708)
-2.877***

(0.383)

-2.709***

(0.453)

-8.686***

(0.576)

65,382
0.709

2.575***
(0.0376)
-2.232***

(0.213)

-3.191***

(0.245)

-8.318***

(0.317)

327,034

0.696

(9)

3.086***
(0.131)

-2.941***

(0.777)

4.077***
(0.207)
-0.885
(1.072)

2.951***
(0.151)
-0.925
(0.893)

-8.585***

-6.611***

-4.709***

(0.948)

(1.147)

(1.075)

-5.023***

-8.976***

-3.605***

(1.293)

(1.682)

(1.298)

74,232
0.287
(10)

13,382
0.670
(11)

26,109
0.739
(12)

Natural

Engineering

Language

History

1.228***
(0.225)
2.041
(1.272)
2.173
(1.529)
-2.280
(1.788)

1.852***
(0.0990)
-1.753***

(0.529)

2.712***
(0.148)

-3.982***

(0.914)

-3.636***

-8.741***

(0.597)

(0.982)

1.878***
(0.158)
-0.895
(0.890)
1.309
(1.023)

-7.002***

-7.865***

-6.772***

(0.773)

(1.261)

(1.223)

1.656***
(0.169)
-1.439
(0.878)
1.628
(1.092)
-1.139
(1.215)

12,993
0.662
(13)
Art

1.713***
(0.206)
-0.603
(1.095)
4.085***
(1.427)
-1.367
(1.751)

VARIABLES

Time Trend

RPM15

RPM16

RPM17

Obs.
Adj. R2

VARIABLES

Time Trend

RPM15

RPM16

RPM17

(1)
ALL

2.968***
(0.0546)
-4.366***

(0.280)

-6.391***

(0.382)

-11.83***

(0.463)

401,266

0.619

(8)

Social

2.832***
(0.0596)
-2.755***

(0.354)

-4.254***

(0.383)

-8.593***

(0.487)

Obs.
Adj. R2

95,740
0.703

6,901
0.721

53,317
0.589

16,895
0.696

15,126
0.662

18,916
0.495

Robust standard errors in parentheses
*** p < 0.01, ** p < 0.05, * p < 0.1

NOTE: All coeﬃcient are reported as β*100. Fixed eﬀects of each book category and various
characteristics of a book are included in the all above model. Full table including all the
other explanatory variables is presented in B4 in the appendix.

91

Table 2.8 Eﬀect of RPM on List Prices (RPM Trend)

(2)

(3)

(4)

(5)

(6)

(7)

Non-Children

Children

General

Religion

Philosophy

Literature

3.086***
(0.131)

-4.082***

(0.762)

4.076***
(0.207)
-1.401
(1.032)

2.951***
(0.151)
-1.562*
(0.884)

-2.188***

-4.445***

-1.894***

(0.569)

(0.722)

(0.548)

1.656***
(0.169)
-0.778
(0.852)
0.588
(0.535)

1.057**
(.520)

2.244***

(.507)

.8972
(.551)

74,232
0.287
(10)

-.3690
(.692)

13,382
0.670
(11)

26,109
0.739
(12)

2.795***
(0.0708)
-2.029***

(0.370)

-2.376***

(0.247)

.4187*
(.235)

65,382
0.709

VARIABLES

(A) Time Trend

(B) RPM15

(1)
ALL

2.969***
(0.0546)
-3.913***

(0.280)

(C) RPM Trend

-3.405***

(A)+(C)

Obs.
Adj. R2

VARIABLES

(A) Time Trend

(B) RPM15

(0.176)

-.4368***

(.159)

401,266

0.619

(8)

Social

2.832***
(0.0596)
-2.370***

(0.337)

(C) RPM Trend

-2.661***

(A)+(C)

Obs.
Adj. R2

(0.225)

.1714
(.217)

95,740
0.703

2.575***
(0.0376)
-1.673***

(0.205)

-2.663***

(0.138)

-.0877
(.132)

327,034

0.696

(9)

1.229***
(0.225)
2.678**
(1.229)
-1.809**
(0.808)

-.5796
(.776)

Natural

Engineering

Language

History

1.852***
(0.0990)
-1.551***

(0.509)

2.711***
(0.148)

-4.715***

(0.871)

1.877***
(0.158)
0.426
(0.855)

-2.506***

-2.512***

-1.869***

(0.344)

(0.604)

(0.572)

-.6545**

(.330)

.1993
(.585)

.0077
(.549)

15,126
0.661

53,317
0.589

6,901
0.720
Robust standard errors in parentheses
*** p < 0.01, ** p < 0.05, * p < 0.1

16,895
0.696

12,993
0.662
(13)
Art

1.711***
(0.206)
0.631
(1.072)
0.772
(0.792)

2.483***

(.763)

18,916
0.494

NOTE: All coeﬃcient are reported as β*100. Fixed eﬀects of each book category and various
characteristics of a book are included in the all above model. Full table including all the
other explanatory variables is presented in B5 in the appendix.

92

of change in time trend after the new RPM regulation, thus the sum of (A) and (C)

becomes the new time trend and (C) becomes the diﬀerence between new and past

time-trend. Overall, the new regulation immediately decreases the list price by 3.9%

and diminishes the time trend by 3.4%, thereby it changes the time trend from 2.9%

to -0.43%. In most cases, except books included in natural science, history, and art,

new regulation decreases the list price immediately and also slows down the increasing

time trend except philosophy and art related books. However, the cross-sectional anal-

ysis has problems in that there are lots of unobservable omitted variables which were

not included in explanatory variables aﬀecting the list price of a book such as an au-

thor’s reputation, the popularity of the book’s topic and advertising cost expenditure of

publishing companies and book stores. Thus, the omitted variables and unobservable

attributes of a book can cause bias problems in regression estimation.

To overcome this problem, I construct panel dataset by using the same books

with diﬀerent editions from the National Library of Korea database. The diﬀerent

editions of the same books could have diﬀerent characteristics such as the change

of pages, diﬀerent supplement and so on. To control the possible diﬀerence within

the diﬀerent editions, I additionally control the various attributes of a book in panel

regression analysis. The full table including all control variables is Table B6 and Table

B7 in appendix. According to results in Table 2.9, the main results are similar to

cross-sectional results in Table 2.7, however, the amount of decrease in list price is

diminished in panel regression results. The estimated time trend is 2.9% in cross

section while it is estimated as 2.4% in panel regression.

In panel regression, after

the new regulation took eﬀect, the list price decrease by 1.36% (=4.14%/3) annually

while it is 4% (=11.83%/3) in cross-section analysis. The list price of books included

in general, natural science and language category shows more rapid decrease whereas

that of literature books increases signiﬁcantly in 2017.

If we consider the case that there was a discontinuity in list prices of books

when the new regulation took eﬀect and it followed diﬀerent time trend after that, the

regression results about the case are presented in Table 2.10. The results are similar

to the cross-sectional analysis in Table 2.8, however, the eﬀects of new regulation on

list prices are estimated as smaller than the result of the cross-sectional analysis. The

introduction of the new RPM regulation immediately decreased the list price by 1.6%

93

Table 2.9 Eﬀect of RPM on List Prices (Panel)

VARIABLES

Time Trend

RPM15

RPM16

RPM17

Obs.
Adj. R2
# of IDs

VARIABLES

Time Trend

RPM15

RPM16

RPM17

Obs.
Adj. R2
# of IDs

(1)
ALL

2.396***
(0.116)

-1.740***

(0.373)

-2.636***

(0.427)

-4.144***

(0.588)

10,855
0.588
4,334

(8)

Social

2.265***
(0.137)

-1.569***

(0.481)

-2.161***

(0.492)

-3.318***

(0.726)

6,173
0.568
2,456

(2)

(3)

(4)

(5)

(6)

(7)

Non-Children

Children

General

Religion

Philosophy

Literature

2.393***
(0.117)

-1.813***

(0.374)

-2.760***

(0.430)

-4.198***

(0.591)

10,764
0.586
4,296

(9)

2.913***
(1.059)
4.797
(4.055)
2.223
(3.796)
-6.261
(5.002)

91

0.845

39
(10)

4.773***
(1.183)
-3.283
(2.745)
-9.437*
(5.154)
-13.20**
(5.454)

118
0.801

63
(11)

2.030
(2.659)
-1.921
(16.02)
8.410
(20.34)
-9.138
(13.08)

104
0.712

54
(12)

Natural

Engineering

Language

History

3.614***
(0.577)
-3.502*
(2.029)
-5.170*
(2.872)

-9.659***

(2.973)

2.006***
(0.176)

-1.662***

(0.457)

-2.585***

(0.610)

-3.963***

(0.819)

3.832**
(1.569)

-7.109***

(2.580)
-11.64**
(4.595)
-15.34**
(6.966)

3.345***
(1.064)
1.086
(3.840)
5.058
(4.321)
-3.176
(4.102)

144
0.779

57

232
0.576
109
Standard errors clustered at IDs are parentheses

3,512
0.683
1,411

232
0.751
114

-0.408
(4.293)

2.91
(2.55)
2.32
(2.84)
4.82***
(1.22)

43

0.722

23

4.106***
(1.446)
-8.348
(5.608)
-5.589
(7.483)

91

0.890

54
(13)
Art

1.355
(1.311)
-1.690
(3.306)
-4.772
(4.109)
-3.773
(6.369)

70

0.632

31

*** p < 0.01, ** p < 0.05, * p < 0.1

NOTE: All coeﬃcient are reported as β*100. Fixed eﬀects of each book category and various
characteristics of a book are included in the all above model. Full table including all the
other explanatory variables is presented in B6 in the appendix.

94

Table 2.10 Eﬀect of RPM on List Prices (Panel, RPM Trend)

(2)

(3)

(4)

(5)

(6)

(7)

Non-Children

Children

General

Religion

Philosophy

Literature

VARIABLES

(A) Time Trend

(B) RPM

(1)
ALL

2.395***
(0.116)

-1.666***

(0.353)

(C) RPM Trend

-1.141***

(0.229)

2.392***
(0.117)

-1.755***

(0.354)

-1.144***

(0.230)

2.866***
(1.045)
5.235
(4.029)
-4.268*
(2.476)

4.759***
(1.182)
-3.426
(2.700)
-5.261**
(2.441)

1.254***

(.201)

1.248***

(.202)

-1.402
(2.076)

-.5018
(1.954)

(A)+(C)

Obs.
Adj. R2
# of IDs

VARIABLES

(A) Time Trend

(B) RPM

10,855
0.588
4,334

(8)

Social

2.264***
(0.137)

-1.501***

(0.453)

(C) RPM Trend

-0.822***

(A)+(C)

Obs.
Adj. R2
# of IDs

(0.298)

1.442***

(.276)

6,173
0.568
2,456

10,764
0.586
4,296

(9)

91

0.843

39
(10)

118
0.800

63
(11)

Natural

Engineering

Language

History

3.640***
(0.579)
-3.177
(2.005)

-3.078***

(1.109)

.5619
(.914)

144
0.777

57

2.006***
(0.176)

-1.603***

(0.442)

-1.105***

(0.325)

.9003***

(.278)

3,512
0.683
1,411

3.830**
(1.564)

-7.175***

(2.575)
-4.265
(2.746)

-.4347
(1.733)

232
0.751
114

3.409***
(1.041)
2.049
(3.650)
-0.728
(2.801)

2.681
(2.414)

232
0.566
109

Standard errors clustered at IDs are parentheses

*** p < 0.01, ** p < 0.05, * p < 0.1

2.229
(2.597)
-3.617
(15.39)
4.944
(6.197)

7.1728
(5.796)

104
0.707

54
(12)

-0.960
(3.568)
21.12
(24.60)
7.227
(6.256)

6.267
(8.619)

43

0.719

23

4.106***
(1.446)
-8.348
(5.608)
2.759
(4.188)

6.864*
(4.143)

91

0.890

54
(13)
Art

1.349
(1.309)
-2.361
(3.410)
-1.803
(2.085)

-.4541
(2.596)

70

0.630

31

NOTE: All coeﬃcient are reported as β*100. Fixed eﬀects of each book category and various
characteristics of a book are included in the all above model. Full table including all the
other explanatory variables is presented in B7 in the appendix.

95

(in cross section, – 3.9%) and decrease the time trend by 1.14% (in cross section, -

3.4%). Thus, the time trend after the new regulation is changed to 1.25% from 2.4%

(in cross section, from 2.9% to -0.44%). If we consider the change of regulation for

a newly-published book (the maximum discount limit is changed to 15% from 19%,

thus 4% increase in actual price at the maximum discount), in short run, it is hard to

say that the new regulation decreases the actual price of a new book. According to

the panel regression results, it is estimated that the new regulation can oﬀset the 4%

decreased discount limit after 3 years, that is from 2018.79

2.3 Eﬀect of RPM on Firms

2.3.1 Online Firm

One of the main aims of reinforced RPM regulation is to support small-scale oﬄine

bookstores and publishing companies in the context of expanding market dominance of

large online ﬁrms. In order to assess the eﬀectiveness of the reinforced RPM regulation,

this section analyzes the impact of the RPM on the proﬁt ratio and sales of top 6 online

ﬁrms by using ﬁrm-level data. I analyze the proﬁt and loss account of the 6 largest

online book store companies from 2009 to 2016.80 Figure 2.2 shows the total amount

of sales, costs, and proﬁt of the top 6 online ﬁrms. According to the graph, the amount

of proﬁt had stayed almost the same from 2009 to 2014, however, it increased in both

2015 and 2016 in the result of enhanced price margin, that is the increased gap between

sales and costs. Table 2.3 shows the top 6 ﬁrms’ proﬁt rapidly increases by 16% and

12% in 2015 and 2016 compared to the previous year and their market share reaches

65% in 2016 while it was 54% in 2014.81

Proﬁt Ratio

First, to show the change of price margin of top 6 online ﬁrms, I use the proﬁt ratio (=

79

1.666%(immediate jump) + 2*1.141%(time trend) = 3.948% compared to the counter-factual situation

that there is no change in RPM regulation and also assume that there is no change in time trend.

80 Because the new RPM regulation took eﬀect from Nov. 2014, I consider the year 2015 and 2016 as the

period under the new regulation.

81 The top 7 (in terms of sales) ﬁrms in the book market is 1) KYOBO, 2) YES24, 3) ALADIN, 4)
INTERPARK, 5) YOUNGPUNG, 6) BANDI & LUNIS, and 7) LIBRO. However, sales, cost, and proﬁt in
the ﬁgure is sum of six ﬁrms except INTERPARK. Because INTERPARK is operating 4 business parts and
they have only published the sales of each business. Thus, I could not ﬁgure out the cost and the proﬁt of
INTERPARK book business part.

96

Figure 2.2 Top 6 Online Firm’s Sales, Cost and Proﬁt

1 – the rate of cost to selling price) variable82 which can capture the pure eﬀect of new

RPM regulation on ﬁrm’s proﬁt structure. we can think of using operating surplus as

a proxy value for proﬁt condition of a ﬁrm under the new regulation. However, it is

aﬀected by lots of other factors such as the change of wage or cost for other input for

operating the ﬁrm (e.g.: rent, investment, and other management costs.).83 Figure 2.3

shows the changes in proﬁt ratio by each ﬁrm.84 From Figure 2.3 we can see that the

proﬁt ratios of all ﬁrms were enhanced after the new regulation took eﬀect.

This paper also analyzes the eﬀect of new RPM regulation on top 6 online ﬁrms’

proﬁt ratio by using regressions and Table 2.11 shows the results. The column (1)

and (2) are the results from pooled OLS, and the columns from (3) to (6) are results

from panel regression models.

In column (3) and (4), I use the ﬁxed-eﬀect model,

and time-varying unobservable model is adopted in column (5) and (6). In all three
models (column (1), (3) and (5)), the proﬁt ratio is signiﬁcantly increased by 5.2∼5.5%.

82 For example, if an online ﬁrm sells a book at 100 and buy the book from a publishing company at 70,

then the proﬁt ratio is 0.3.

83 Actually, top 6 online ﬁrms have been expanding their business aggressively after 2015 based on the

increased earnings. Refer to the footnote 70.

84

INTERPARK, the 4th biggest online book store have many business areas besides the online book store.
The total amount of sales is published by each business sector, however, the cost of sales is not published by
each sector. Thus, INTERPARK is not included in the analysis of the proﬁt ratio.

97

Figure 2.3 The Proﬁt Ratio of online Firms

According to model (2), (4) and (6) showing the eﬀect separately by each year in 2015

and 2016, the proﬁt ratio is increased more in 2016 compared to 2015. Time trend

shows negative values in all models; however, it is not signiﬁcant. All models also

control the sales amount to consider a plausible size eﬀect of proﬁt ratio; however, it

is not signiﬁcant in all models. Another interesting thing is the coeﬃcient of “Only

Online” which is -4.4% (OLS) and -5.16% (Time-varying).

It means that if a ﬁrm

only operates the online store without an oﬄine store, its proﬁt ratio is lower roughly

by 5% than a company operating both online and oﬄine stores. This results can be

caused by the fact that the price margin of the oﬄine store is larger than the online

store. Another possible reason for this is that top online ﬁrms have been opening new

oﬄine stores aggressively after the new RPM regulation took eﬀect.85 Thus, it is highly

possible that the big oﬄine stores operated by top online ﬁrms absorb a large portion

of converting demand from online to oﬄine. Consequently, online ﬁrms simultaneously

operating an oﬄine store could have obtained two gains: enhanced price margin at the

online store and more customer at their oﬄine stores.

85 Refer to the footnote 70.

98

Table 2.11 Eﬀect of RPM on Online Firm’s Proﬁt Ratio

VARIABLES

RPM

RPM15

RPM16

Time Trend

Sales

Only Online

γ

Dependent, Proﬁt Ratio = (Sales - Cost) / Sales

OLS

Fixed Eﬀect

Time-varying

(1)

(2)

(3)

(4)

(5)

(6)

5.243**
(2.615)

5.553**
(2.616)

5.240***

(1.931)

4.526
(2.886)
6.370**
(3.219)
-.1537
(.506)
-.000
(.000)

-4.455**
(2.127)

-.1028
(.495)
-.000
(.000)

-4.461**
(2.112)

4.976*
(2.905)
6.431**
(3.217)
-.4538
(.557)
.000
(.000)

-.425
(.548)
.000
(.000)

4.348*
(2.384)
6.206**
(2.468)

-.000
(.000)
-5.166*
(2.763)
.1502
(.168)

.000
(.000)
-5.169*
(2.703)
.1381
(.166)

Constant

27.481***

27.667***

18.554***

18.997***

22.872***

22.539***

(2.113)

(2.150)

(6.436)

(6.566)

(4.681)

(4.746)

Obs.
Adj. R2

48

0.2353

48

0.2420

48

0.1999

48

0.2047

42

0.2846

42

0.2928

*** p < 0.01, ** p < 0.05, * p < 0.1

Note: Robust standard errors (OLS) and standard errors clustered at IDs (FE and Time-
varying) are parentheses. “RPM” is dummy variable which is equal to one if the year of data
is 2015 or 2016. “RPM15” and “RPM16” are also dummy variables representing 2015 and
2016 respectively. “Only Online” is dummy variable which is one if the ﬁrm only operates
at online without oﬄine stores. γ is the coeﬃcient of lagged proﬁt ratio in the Time-varying
Unobservable model.

Proﬁts and Sales of the Top 6 Online Firms

The Figure 2.4a shows the amount of proﬁt by each online ﬁrm. In general, the amount

of proﬁt rapidly increase after the new regulation took eﬀect in 2015 and 2016. Because

of plausible zero-sum competition among online ﬁrms, the picture becomes more clear

if we see the Figure 2.2. Until 2014, the total amount of proﬁt had stayed at almost

same level, however, it started increase after the new regulation took eﬀect. The upper

panel of Table 2.12 shows the eﬀect of new regulation on the top 6 online ﬁrms’ proﬁt

by using various regression models.

If we focus on the result of the Time-varying

Unobservable model in the column (3), the new regulation increases the online ﬁrm’s

proﬁt by 24% after controlling the ﬁrms’ sales. However, the eﬀect is not signiﬁcant in

OLS and ﬁxed eﬀect model in the column (1) and (2).

Figure 2.4b shows the changes in sales by each online ﬁrm. After the new

99

(a) Proﬁt

(b) Sales

Figure 2.4 Proﬁts and Sales of Top 6 Online Firm

100

Table 2.12 Eﬀect of the New RPM on Online Firms

Dependent = Log(Proﬁt)

(2)

(3)

Fixed Eﬀect Time-varying

VARIABLES

RPM

Time Trend

(1)
OLS

.206
(.210)
0.001
(.039)

Log(Sales)

1.042***

(.073)
-.168
(.174)

.238
(.203)
-0.023
(.039)

1.955***

(.380)

Only Online

γ

Constant

Obs.
Adj. R2

VARIABLES

RPM

Time Trend

Only Online

γ

Constant

Obs.
Adj. R2

-1.864**

(.871)

-12.74***

(4.517)

48

0.8399

48

0.4498

Dependent = Log(Sales)

(4)
OLS

-.037
(.434)
.0260
(.082)
.928***
(.330)

-.037
(.085)
.026
(.015)

11.72***

(.325)

11.87***

(.062)

48

0.1548

48

0.0933

.241*
(.145)

1.048***

(.077)
-.186
(.186)
-.0548
(.163)
-2.053*
(1.005)

42

0.8329

-.059
(.052)

.087
(2.23)
.999***
(.023)
-0.252
(.271)

42

0.9842

(5)

(6)

Fixed Eﬀect Time-varying

*** p < 0.01, ** p < 0.05, * p < 0.1

Note: Robust standard errors (OLS) and standard errors clustered at IDs (Fixed eﬀect
and Time-varying) are parentheses. “RPM” is dummy variable which is equal to one if
the year of data is 2015 or 2016. “Only Online” is dummy variable which is one if the
ﬁrm only operates at online without oﬄine stores. γ is the coeﬃcient of lagged proﬁt
ratio in the Time-varying Unobservable model.

101

regulation took eﬀect, the sales of KYOBO, YOUNGPUNG, and LIBRO have stayed

almost the same level while those of YES24 and ALADIN increase and BANDI &

LUNIS’ sales have decreased. The growth of YES24 and ALADIN is remarkable in

that both companies are leading online stores in used book market.86 The lower panel

of Table 2.12 shows the eﬀect of new regulation on online ﬁrm’s sales. The eﬀect of

new regulation on the ﬁrm’s sales shows negative values (3.7% in the column (4) and

(5), and 5.9% in the Time-varying Unobservable model in the column (6)); however,

it is not signiﬁcant.

If we sum up the above results, we can conclude that online ﬁrms’ sales are not

signiﬁcantly aﬀected by the RPM regulation. However, both top 6 online ﬁrms’ proﬁts

and proﬁt ratios increase signiﬁcantly by the enhanced price margin. One might expect

that the prices of books are increased by the new RPM regulation. However, the cost

of buying books for online ﬁrms are not increased relatively. As a result, ironically, the

market share of top online ﬁrms even increase rapidly after the new regulation took

eﬀect even though the new regulation intended the opposite results.

2.3.2 Publishing Company

In this section, this paper analyzes the eﬀect of reinforced RPM regulation on business

circumstances of publishing companies. Figure 2.5 shows the Publishing Industry Pro-

duction Index published by Statistics Korea in the Survey of Service Industry. This

statistics is compiled by using publishing companies’ sales data. This paper also uses

ﬁrm-level data including more than 70 ﬁrms from KIS database to show the eﬀect of

new RPM regulation on the publishing industry in the following regression analysis.

However, the ﬁrms in the KIS database are included in the biggest ﬁrms among all

publishing companies. A company is required to be externally audited only if the ﬁrm’s

sale is above a certain level. Because the KIS database is generated from companies

which take external audit, the dataset does not include smaller publishing companies.

On the other hand, the Publishing Industry Production Index uses data of all publish-

ing companies including various size of companies which is not opened to public. Thus,

86 ALADIN has opened 17 more used book oﬄine stores (total number of used book stores be-
comes 37 in 2016) after the new RPM regulation. YES24 also opened the ﬁrst oﬄine used book
store last April.
2016 and have opened four more oﬄine used book stores. According to a news
report, YES24’s revenue from used book part has been growing on average 30% for each month.
http://news.heraldcorp.com/view.php?ud=20160425000116

102

Figure 2.5 The Publishing Industry Production Index

the index shows the overall situation of sales in the publishing industry. According to

the nominal index, it tends to increase until 2014. However, it decreases rapidly in 2015
and 2016. The nominal index changed from 100 to 102.5 during 2010∼2014; however,
it decreases 8% in 2015 and again 4.3% in 2016. If we check the real index eliminating

the eﬀect of an inﬂation, it shows a downward trend but rebounded in 2013 and have

been decreasing more rapid speed in 2015 and 2016. The real index decreases 10% in

2015 and again 5.6% in 2016.

That is, after the new regulation took eﬀect, sales of publishing companies did

not increase. Rather, the Publishing Industry Production Index shows the opposite

result. Now, this paper investigates each ﬁrm-level data from KIS database including

more than 70 publishing companies.87

Sales, Proﬁt Ratio and Proﬁt of Publishing Companies

Table 2.13 shows the panel regression results. I use time-varying unobservable mod-

els for various categories.88 According to the result in the column (1) including all

87 The number of companies required to get outside audit diﬀers each year. Some companies disappeared

from the sample, and some companies newly appear on the list. Thus, the dataset is the unbalanced panel.

88

“Study” category includes companies that mainly publish textbooks and study related books such

103

Table 2.13 Eﬀect of RPM on Publishing Firm’s Sales

Dependent = Log(Sales)

(1)

(2)

(3)

(4)

(5)

Time-varying Unobservable Model

.0235
(.040)
.0426
(.067)
1.246**
(.494)
.952***
(.085)
-.185
(.119)

VARIABLE

RPM15

-.086***

(.024)

-.0343
(.043)

-.1712***

-.1076**

(.051)

(.044)

RPM16

-.122***

-.0934*

-.2268***

-.1098*

(.031)

(.053)

(.068)

(.056)

Log(Mean Sales)

1.033***

1.024***

1.056***

.9951***

(.025)

(.037)

(.023)

.5379***

.6342***

.4356***

γ

Constant

Obs.
Category
Adj. R2

(.016)
.549***
(.033)
-.167**
(.076)

471
ALL
0.9768

(.046)
-.119
(.124)

198

Study
0.9764

(.104)
-.229*
(.135)

78

(.082)
.002***
(.128)

142

61

Children General English
0.9937

0.9891

0.9769

Standard errors clustered at IDs are parentheses

*** p < 0.01, ** p < 0.05, * p < 0.1

Note: “RPM15” and “RPM16” are dummy variables which are equal to one if the year
of data is 2015 and 2016 respectively. γ is the coeﬃcient of lagged proﬁt ratio in the
Time-varying Unobservable model.

companies, publishing companies’ sales decrease by 8.6% and 12.2% in 2015 and 2016

respectively. To control a plausible size eﬀect on dependent variable,89 I include the

log of the mean sales of a ﬁrm. The companies publishing books for children show the

largest decrease in their sales among all categories. In 2015, their sales decreased by

17% and more than 22% in 2016. On the other hand, the sales of companies publish-

ing English study books did not decrease after the RPM regulation was strengthened.

The results show that, after the new regulation took eﬀect, the top 70 ﬁrms’ sales

are decreased signiﬁcantly. However, the amount of decrease is somewhat diﬀerent

depending on the category of companies’ books. The companies related to publish-

ing study related material (“Study” and “English”) comparatively retained their sales.

One could expect that the diﬀerent price elasticity of demand for a book results in

diﬀerent outcomes. In general, a book related to study and exam preparation could

have lower price elasticity compared to other books such as literature work; thereby

its sale could decrease less or even increase when the price of book increase.

Table 2.14 shows the eﬀect of new regulation on publishing ﬁrms’ proﬁt ratio.

as books for exam preparation. “Children” includes companies publishing books for children. Companies
publishing a single book (such as literature, philosophy, history and so on) are included in the “General”
category and “English” includes the companies publishing English study materials.

89

It is possible that the eﬀect of the new RPM regulation can be diﬀerent according to a ﬁrm’s size.

104

Table 2.14 Eﬀect of RPM on Publishing Firm’s Proﬁt Ratio

Dependent = Proﬁt Ratio = (Sales - Cost) / Sales

VARIABLE

(1)

(2)

(3)

(4)

(5)

Time-varying Unobservable Model

RPM15

RPM16

Log(Sale)

γ

Constant

Obs.
Category
Adj. R2

1.203
(.838)

2.939**
(1.262)
.5490
(1.263)
.9266***

(.015)
2.885**
(1.171)

471
ALL
0.8902

1.544
(1.262)
4.712**
(1.897)
-.3474
(1.677)
.9273***

(.020)
3.260**
(1.834)

198

Study
0.9174

1.240
(1.674)
-2.014
(2.479)
2.568
(3.044)
.9371***

(.041)
1.311
(1.973)

1.949
(1.707)
3.579
(2.517)
1.296
(1.949)
.7726***

(.053)
6.759
(1.949)

1.930
(1.836)
5.699**
(2.699)
-.4448
(.4.660)
.9186***

(.052)
5.1036
(4.808)

78

142

61

Children General English
0.8529

0.6586

0.9073

Standard errors clustered at IDs are parentheses

*** p < 0.01, ** p < 0.05, * p < 0.1

Note: “RPM15” and “RPM16” are dummy variables which are equal to one if the year
of data is 2015 and 2016 respectively. γ is the coeﬃcient of lagged proﬁt ratio in the
Time-varying Unobservable model.

Overall, after the new regulation was enacted, the proﬁt ratio of publishing companies

increase by 1.2% in 2015 (not signiﬁcant at 10% signiﬁcance level) and 2.9% (its p-value

is 0.02) in 2016. On the other hand, the enhanced proﬁt ratio of “Study” and “English”

category is remarkable. There were 4.7% and 5.7% improvement in the proﬁt ratio for

each category. Similar to the eﬀect of new RPM regulation on ﬁrms’ sales, the low

price elasticity of these books can be the reason for the enhanced price margin.

Table 2.15 shows the panel regression result about the eﬀects of new regulation

on ﬁrms’ amount of proﬁt. In all columns, the coeﬃcients of (RPM15) and (RPM16)

shows positive sign except the children’s book publishing companies in 2016. However,

the only the coeﬃcient of (RPM15) in column (2) – which shows the eﬀect of new

regulation on proﬁts of companies which are included in (Study) category in 2015 – is

signiﬁcant under 10% signiﬁcance level. Also, the coeﬃcient of (RPM16) in the column

(5) implies the signiﬁcant eﬀect of the RPM regulation on the proﬁts of ﬁrms which

publish English study books.

To sum up, according to the above results analyzing the eﬀect of new regulation

on publishing companies’ sales, proﬁt ratio, and the amount of proﬁt, the reinforced

RPM regulation signiﬁcant decreased ﬁrms’ sales, increased in proﬁt ratio and made no

105

Table 2.15 Eﬀect of RPM on Publishing Firm’s Proﬁt

Dependent = Log(Proﬁt = Sale - Cost)

VARIABLE

(1)

(2)

(3)

(4)

(5)

Time-varying Unobservable Model

RPM15

RPM16

Log(Sale)

γ

Constant

Obs.
Category
Adj. R2

.0255
(.0313)
.0321
(.0443)
.9630***
(.0346)
.7994***
(.0211)
-.0921**
(.0737)

471
ALL
0.9681

.0614*
(.0367)
.0581
(.0554)
.9133***
(.0514)
.9383***
(.0240)
.0031
(.0373)

198

Study
0.9834

.0474
(.0476)
-.0600
(.0703)
1.033***
(.0847)
.9317***
(.0452)
-.0791
(.0943)

.0356
(.0515)
.0478
(.0624)
1.094***
(.0212)
.2701***
(.0430)

-1.291***

(.1765)

78

142

Children General

0.9930

0.9745

.0332
(.0322)
.0945**
(.0468)
.9805***
(.0721)
.8960***
(.0592)
-.0363
(.0843)

61
Eng
0.9950

Standard errors clustered at IDs are parentheses

*** p < 0.01, ** p < 0.05, * p < 0.1

Note: “RPM15” and “RPM16” are dummy variables which are equal to one if the year
of data is 2015 and 2016 respectively. γ is the coeﬃcient of lagged proﬁt ratio in the
Time-varying Unobservable model.

signiﬁcant changes in the proﬁt of publishing companies. That is, given the decreasing

sales, publishing companies kept their proﬁts by enhancing the proﬁt ratio. Enhanced

proﬁt ratio could be attained by some changes in ﬁrms’ proﬁt structure and strategy.

First, if publishing companies sell their books to a retail book store at a higher price

than before, it could increase the proﬁt ratio of publishing companies. Increased re-

tail price owing to the reinforced RPM regulation can raise the wholesale price of a

book, and this is the eﬀect that the government intended from the introduction of

the new regulation. However, second, it is also possible that publishing companies

decrease their costs against the decrease in sales by reducing the variety of publication

and concentrating on publishing some popular books. The diminished variety of new

publication as we can see in (Table 2.4) implies that the enhanced proﬁt ratio could

be the result of the second reason.

2.3.3 Oﬄine Book Stores

As I discuss in the introduction part, one of the important purposes of reinforced RPM

regulation is to support small oﬄine bookstores. In this section, this paper introduces

some survey results to investigate the eﬀect of the new RPM regulation on small oﬄine

106

bookstores, because there is no available ﬁrm-level data in case of the small-scale oﬄine

bookstores.

Table 2.16 Survey Result of Oﬄine Stores

Year

Response

Increase Almost Same Decrease

Total

(%)

Less Than 2

3∼4 Employees

5∼9 Employees

More Than 10

2014
2015
2016

2014
2015
2016

2014
2015
2016

2014
2015
2016

2014
2015
2016

839
813
774

604
552
665

95
152
113

59
74
46

16
35
15

4.1
8.7
5.0

3.0
8.2
5.0

4.2
9.2
2.7

10.2
9.5
6.5

25.0
14.3
20.0

37.1
42.3
28.8

34.4
36.6
28.4

48.4
44.7
33.6

44.1
68.9
28.3

43.8
65.7
13.3

58.8
49
66.2

62.6
55.3
66.6

47.4
46.1
63.7

45.8
21.6
65.2

31.3
20.0
66.7

DATA: KPIPA (2017)

Table 2.16 shows the results of the Publishing Industry Survey by Publication

Industry Promotion Agency of Korea (KPIPA). The survey asked oﬄine bookstores

whether sales increased, decreased, or remained unchanged compared to the previous

year. The table presents the number of respondents and the portion of each response

into three categories by the number of employees. According to the results, 66.2%

of oﬄine book store responded that sales decreased compared to the previous year in

2016. The proportion of respondents who reported a decrease in sales overall declined

temporarily in 2015; however, increased again in 2016. The results by the number of

employees show a similar pattern; however, when the number of respondents is small,

the volatility of the result becomes large.

It is necessary to note that analyses based on the survey data are highly limited

because the survey is not panel data and also there is no quantitive information such

as the number of changes in sales of oﬄine stores. However, according to the results

of the survey, it is hard to insist that the introduction of new RPM regulation helps to

107

support the business of small oﬄine book stores.

2.4 Conclusion

The RPM regulation on Korean book market was enacted and reinforced to protect a

variety of culture and knowledge by helping unpopular books to survive in the market

and also to support small publishing companies and oﬄine book stores under the

situation that some large online bookstores have been dominating the book market.

Based on the purposes of the RPM regulation, this paper investigates and evaluates

whether the regulation achieves its goal.

Various empirical analyses imply that the RPM regulation causes various adverse

eﬀects opposite to the original intention of the regulation. First of all, the market share

of the seven online ﬁrms, which stay around at 54%, increased to 65% and the total

sales of all the remaining bookstores decreased by 34 and consequently, the degree

of market concentration became worse. Second, this paper evaluates the eﬀect of

RPM regulation on the number of new book publications as a proxy for the variety of

culture and knowledge by using the data of the National Library of Korea. According

to various panel regression results, the number of newly-published books decreased

by 6% (in 2015) and 8% more (in 2016) after the regulation. Third, this paper also

analyzes the eﬀects of RPM regulation on the management conditions of large online

ﬁrms, publishing companies, and small oﬄine bookstores, respectively. Both proﬁt

ratios and proﬁts of large online ﬁrms signiﬁcantly enhanced after the regulation was

reinforced. On the other hand, the Publishing Industry Production Index based on the

sales of publishing companies decreased 16% after the new regulation. Also, regression

results using the data of the top 70 publishing companies from KIS database show

similar results. Overall, sales of the publishing companies decrease 8% in 2015 and

12% in 2016. As there is no data available for small oﬄine bookstores, this paper

refers to the KPIPA’s survey results to examine the impact of the new regulation on

oﬄine bookstores. From the survey results, we might estimate that the introduction

of the new RPM regulation has not helped to improve the management condition of

small bookstores.

Also, the reinforced RPM regulation decreased the list price of the newly-

published book by 4.1% for three years based on the panel regression results in Table

108

2.9. Nonetheless, we could not insist that the new RPM regulation decreases the retail

price of a book because the regulation diminishes the discount rate. When we con-

sider the signiﬁcant discount rate of online book stores for a non-newly-published book

which has no price regulation before the new RPM regulation took eﬀect, 4.1% (from

the panel regression in Table 2.9) or 11.8% (from the cross-sectional result in Table

2.7) decrease for three years might be insuﬃcient to oﬀset the diminished discount rate

by the new RPM regulations.

In conclusion, the results of various empirical studies show that new RPM regula-

tion is not accomplishing the purpose of its introduction. I suggest that the government

of Korea should review other methods such as the direct subsidies for small publishing

companies and oﬄine book stores instead of price regulation such as the resale price

maintenance.

109

APPENDIX

110

APPENDIX B

Table B1 The Number of Newly-Published Books

Category

General
Philosophy
Religion
Social Science
Science
Engineering
Art
Language
Literature
History
Study
Children

2008

822
946
2,009
6,335
593
2,754
1,451
1,823
8,482
1,139
1,787
8,417

2009

805
915
2,177
6,483
542
3,054
1,407
1,660
8,718
1,008
1,803
7,884

2010

703
1,055
1,899
6,017
541
3,206
1,382
1,625
8,192
1,031
2,512
7,352

2011

715
1,152
1,925
5,919
647
3,628
1,354
1,385
8,184
989
2,159
9,546

2012

613
1,237
1,889
6,089
521
3,552
1,329
1,192
7,963
1,083
1,379
7,495

2013

822
1,335
1,899
7,097
645
3,880
1,402
1,399
9,296
1,283
1,356
7,424

2014

2015

1,078
1,367
2,078
8,015
718
4,417
1,605
1,408
10,671
1,291
1,462
7,269

930
1,303
1,668
7,561
665
4,508
1,729
1,392
10,899
1,403
1,393
5,572

Sub-Total

36,558

36,456

35,515

37,603

34,342

37,838

41,379

39,023

Comic
Total

6,541
43,099

5,735
42,191

4,776
40,291

6,433
44,036

5,425
39,767

5,308
43,146

6,210
47,589

6,190
45,213

Data : Korean Publisheres Association

Table B2 The Average Price of Newly-Published Books

Category

2008

2009

2010

2011

2012

2013

2014

2015

2016

General
Philosophy
Religion
Social Science
Science
Engineering
Art
Language
Literature
History
Study
Children

18,782
16,835
13,292
17,587
22,960
21,142
18,465
15,668
9,845
19,963
10,373
8,536

19,927
18,231
12,677
18,795
19,162
21,776
23,731
15,319
10,227
19,088
10,328
8,992

18,456
16,061
13,497
19,108
19,963
21,459
18,600
14,263
10,352
18,733
9,365
9,427

18,932
16,883
13,778
19,855
20,916
21,647
19,928
16,710
10,887
19,666
10,314
9,813

19,999
17,012
14,158
19,821
21,569
22,585
19,438
18,704
11,297
19,760
11,113
10,617

20,365
17,405
13,952
21,955
21,679
24,579
21,262
16,636
11,485
20,398
11,875
9,932

24,050
17,796
14,762
22,347
19,771
24,400
28,590
17,554
13,229
19,713
11,845
10,527

20,046
17,628
14,929
20,574
21,625
23,169
27,332
18,081
11,852
22,043
10,889
9,955

21,170
18,581
15,486
22,453
22,879
25,586
29,684
16,644
11,805
20,704
11,495
10,545

Sub-Total

13,494

14,148

13,965

14,459

15,333

16,055

17,232

16,505

18,607

Comic
Total

4,413
12,116

4,441
12,829

4,309
12,820

4,541
13,010

4,720
13,885

4,865
14,678

4,959
15,631

4,998
14,929

5,186
17,356

Data : Korean Publisheres Association

111

Category

General
Philosophy
Religion
Social Science
Science
Engineering
Art
Language
Literature
History
Study
Children

Sub-Total

Comic
Total

Table B3 The Number of Circulation

(Unit = 1000)

2008

1,616
1,729
4,393
10,853
761
3,641
2,265
4,048
17,641
2,151
13,621
26,885

89,605

2009

1,514
1,488
4,368
10,937
907
3,902
2,202
3,591
18,644
1,826
14,297
29,275

92,951

2010

1,405
1,980
3,790
10,765
727
4,397
2,116
4,338
17,280
1,829
22,007
26,200

2011

2012

2013

2014

2015

1,336
2,153
3,997
9,364
1,113
4,977
2,151
2,720
15,837
1,816
17,217
37,705

1,190
2,162
3,328
9,774
675
4,634
2,007
1,871
14,796
1,866
10,547
26,537

1,448
2,240
3,384
9,618
783
4,872
1,936
2,185
15,945
2,064
10,630
24,863

1,574
1,892
3,496
9,651
796
4,986
1,941
2,000
15,176
2,029
16,713
26,167

1,675
2,036
2,986
9,490
798
4,798
2,415
2,062
15,610
2,197
16,524
16,837

96,833

100,387

79,388

79,968

86,420

77,429

16,911
106,516

13,263
106,215

9,477
106,310

9,163
109,550

7,518
86,907

6,546
86,513

7,746
94,166

7,589
85,018

Data : Korean Publisheres Association

112

VARIABLE

Time Trend

RPM15

RPM16

RPM17

log(page)

size

color picture

partial color

picture

graph

etc

2.968***
(0.0546)
-4.366***

(0.280)

2.575***
(0.0376)
-2.232***

(0.213)

3.086***
(0.131)

-2.941***

(0.777)

4.077***
(0.207)
-0.885
(1.072)

2.951***
(0.151)
-0.925
(0.893)

-6.391***

-3.191***

-8.585***

-6.611***

-4.709***

(0.382)

(0.245)

(0.948)

(1.147)

(1.075)

-11.83***

-8.318***

-5.023***

-8.976***

-3.605***

(0.463)
42.14***
(0.211)
1.346*
(0.700)
14.48***
(0.251)
9.297***
(0.489)
2.328***
(0.174)
1.726***
(0.510)

-2.611***

(0.356)

(0.317)
53.15***
(0.180)
6.145***
(0.0353)
14.88***
(0.214)
8.162***
(0.348)
1.018***
(0.140)

-2.829***

(0.153)
-0.0144
(0.165)

(1.293)
22.26***
(0.363)
0.242
(0.158)

-5.250***

(0.820)
9.185***
(1.831)

-6.487***

(0.871)
11.49***
(1.026)

(1.682)
57.14***
(0.701)
4.716***
(0.145)
0.897
(0.950)

-7.935***

(1.414)
3.772***
(0.792)
-0.273
(0.658)

(1.298)
72.95***
(0.836)
6.519***
(0.188)
22.13***
(1.330)
13.35***
(1.762)
-0.322
(0.580)
-0.0563
(1.394)

-4.239***

-3.542***

-1.640***

(0.467)

(0.819)

(0.617)

1.656***
(0.169)
-1.439
(0.878)
1.628
(1.092)
-1.139
(1.215)
56.55***
(0.905)
6.260***
(0.147)
10.82***
(0.902)
9.718***
(1.840)
1.334**
(0.540)

-5.237***

(0.736)
-1.561**
(0.696)

number of books

-28.39***

-67.26***

-15.75***

-65.18***

-57.85***

-61.62***

DVD, CD, MP3

# of Appendix

(2.461)
12.12***
(0.696)

(1.439)
1.267***
(0.428)

(1.765)
-2.933
(2.005)
23.29***
(1.170)

(0.872)
0.764
(0.971)

(9.942)
7.200**
(3.118)

(1.799)
14.75***
(3.449)

Constant

-5,246***

-4,590***

-5,405***

-7,609***

-5,495***

-2,770***

(105.2)

(75.87)

(265.6)

(416.1)

(309.8)

(341.9)

2.795***
(0.0708)
-2.877***

(0.383)

-2.709***

(0.453)

-8.686***

(0.576)
39.37***
(0.337)
9.097***
(0.151)
16.17***
(0.330)

-3.086***

(0.660)
3.805***
(0.302)
2.291
(1.410)
2.906***
(0.390)

-74.63***

(0.551)
10.53***
(2.884)

-5,026***

(141.9)

65,382
0.709

Table B4 Eﬀect of RPM on List Price

(1)
ALL

(2)

Non-Child

(3)

Child

(4)

(5)

(6)

(7)

General

Religion

Philosophy

Literature

Obs.
Adj. R2

VARIABLE

Time Trend

RPM15

RPM16

RPM17

log(page)

size

color picture

partial color

picture

graph

etc

401,266

0.619

(8)

Social

2.832***
(0.0596)
-2.755***

(0.354)

-4.254***

(0.383)

-8.593***

(0.487)
60.74***
(0.353)
5.039***
(0.0613)
9.457***
(0.518)
5.681***
(0.740)
0.870***
(0.201)

327,034

0.696

(9)

Natural

1.228***
(0.225)
2.041
(1.272)
2.173
(1.529)
-2.280
(1.788)
60.59***
(0.942)
4.635***
(0.162)
21.91***
(1.138)
14.96***
(1.755)
0.996
(0.945)

-2.993***

-4.830***

(0.211)
0.388
(0.262)

(0.812)
0.843
(0.791)

74,232
0.287

(10)
Tech

13,382
0.670

(11)

26,109
0.739

(12)

Language

History

1.852***
(0.0990)
-1.753***

(0.529)

2.712***
(0.148)

-3.982***

(0.914)

-3.636***

-8.741***

(0.597)

(0.982)

1.878***
(0.158)
-0.895
(0.890)
1.309
(1.023)

-7.002***

-7.865***

-6.772***

(0.773)
50.23***
(0.431)
5.250***
(0.0777)
21.95***
(0.488)
21.25***
(0.776)
3.076***
(0.425)
1.005***
(0.334)
1.147*
(0.618)

(1.261)
55.32***
(0.623)
4.558***
(0.0872)
6.236***
(0.703)
4.513**
(1.944)
1.161**
(0.519)
3.807***
(0.921)
4.901***
(0.903)

(1.223)
55.75***
(1.251)
5.790***
(0.134)
-1.201
(0.789)
3.498***
(1.150)

-4.493***

(0.723)

-2.129***

(0.803)

12,993
0.662

(13)
Art

1.713***
(0.206)
-0.603
(1.095)
4.085***
(1.427)
-1.367
(1.751)
35.46***
(0.689)
5.008***
(0.102)
25.29***
(0.905)
20.39***
(1.092)
2.192**
(0.894)
-2.067**
(0.943)

-1.652***

-4.348***

(0.380)

(0.620)

# of Books

-61.80***

-66.17***

-68.36***

-66.66***

-63.32***

-74.06***

DVD, CD, MP3

(2.258)
3.297***
(1.096)

(1.326)
20.49***
(5.702)

(0.856)
5.142***
(0.884)

(0.496)
3.937***
(0.581)

(0.774)
15.41***
(3.939)

(0.981)
1.770
(1.489)

Constant

-5,134***

-1,890***

-3,100***

-4,852***

-3,192***

-2,719***

(119.7)

(452.4)

(199.0)

(297.8)

(317.7)

(413.7)

Obs.
Adj. R2

95,740
0.703

6,901
0.721

16,895
0.696
Robust standard errors in parentheses

53,317
0.589

*** p<0.01, ** p<0.05, * p<0.1

113

15,126
0.662

18,916
0.495

Table B5 Eﬀect of RPM on List Price (RPM Trend)

(1)
ALL

(2)

Non-Child

(3)

Child

(4)

(5)

(6)

(7)

General

Religion

Philosophy

Literature

VARIABLES

Time Trend

RPM

2.969***
(0.0546)
-3.913***

(0.280)

2.575***
(0.0376)
-1.673***

(0.205)

3.086***
(0.131)

-4.082***

(0.762)

4.076***
(0.207)
-1.401
(1.032)

2.951***
(0.151)
-1.562*
(0.884)

RPM Trend

-3.405***

-2.663***

-2.188***

-4.445***

-1.894***

log(page)

size

color picture

partial color

picture

graph

etc

(0.176)
42.13***
(0.211)
1.345*
(0.700)
14.48***
(0.251)
9.308***
(0.489)
2.327***
(0.174)
1.727***
(0.510)

-2.604***

(0.356)

(0.138)
53.15***
(0.180)
6.145***
(0.0353)
14.87***
(0.214)
8.177***
(0.348)
1.016***
(0.140)

-2.829***

(0.153)
-0.00103
(0.165)

(0.569)
22.28***
(0.363)
0.242
(0.158)

-5.286***

(0.820)
9.232***
(1.834)

-6.504***

(0.871)
11.37***
(1.023)

(0.722)
57.12***
(0.700)
4.726***
(0.145)
0.971
(0.948)

-7.905***

(1.415)
3.837***
(0.791)
-0.260
(0.658)

(0.548)
72.96***
(0.837)
6.520***
(0.188)
22.13***
(1.330)
13.34***
(1.762)
-0.337
(0.579)
-0.0336
(1.393)

-4.228***

-3.574***

-1.641***

(0.468)

(0.821)

(0.617)

1.656***
(0.169)
-0.778
(0.852)
0.588
(0.535)
56.56***
(0.905)
6.252***
(0.148)
10.85***
(0.902)
9.734***
(1.844)
1.342**
(0.540)

-5.234***

(0.737)
-1.555**
(0.698)

# of Books

-28.39***

-67.28***

-15.77***

-65.15***

-57.87***

-61.58***

2.795***
(0.0708)
-2.029***

(0.370)

-2.376***

(0.247)
39.34***
(0.338)
9.100***
(0.151)
16.17***
(0.330)

-3.093***

(0.661)
3.818***
(0.302)
2.324*
(1.408)
2.935***
(0.391)

-74.69***

(0.551)
10.47***
(2.882)

-1.607***

-4.264***

(0.380)

(0.622)

# of Books

-61.80***

-66.16***

-68.36***

-66.66***

-63.36***

-74.21***

DVD, CD, MP3

Constant

Obs.
Adj. R2

(2.258)
3.282***
(1.096)
553.5***
(3.229)

95,740
0.703

(1.329)
20.40***
(5.705)
577.2***
(6.557)

(0.856)
5.143***
(0.884)
618.7***
(3.079)

(0.495)
3.920***
(0.581)
592.7***
(4.327)

(0.775)
15.55***
(3.967)
579.1***
(7.248)

6,901
0.720
Robust standard errors in parentheses

53,317
0.589

16,895
0.696

15,126
0.661

*** p<0.01, ** p<0.05, * p<0.1

114

(0.983)
1.764
(1.488)
721.2***
(5.620)

18,916
0.494

DVD, CD, MP3

# of Appendix

Constant

Obs.
Adj. R2

VARIABLES

Time Trend

RPM

RPM Trend

log(page)

size

color picture

partial color

picture

graph

etc

(2.461)
12.12***
(0.696)

(1.439)
1.263***
(0.428)

714.5***
(16.95)

580.2***
(2.207)

401,266

0.619

(8)

Social

2.832***
(0.0596)
-2.370***

(0.337)

-2.661***

(0.225)
60.74***
(0.353)
5.039***
(0.0613)
9.456***
(0.518)
5.700***
(0.741)
0.869***
(0.201)

327,034

0.696

(9)

Natural

1.229***
(0.225)
2.678**
(1.229)
-1.809**
(0.808)
60.58***
(0.941)
4.634***
(0.162)
21.90***
(1.139)
15.08***
(1.754)
0.997
(0.945)

-2.993***

-4.836***

(0.211)
0.387
(0.262)

(0.812)
0.846
(0.792)

(0.871)
0.749
(0.970)

(9.944)
7.186**
(3.119)

(1.795)
14.75***
(3.441)

577.8***
(4.808)

431.4***
(12.64)

556.0***
(6.063)

587.3***
(3.600)

(1.766)
-2.989
(2.000)
23.32***
(1.166)
792.1***
(5.559)

74,232
0.287

(10)
Tech

13,382
0.670

(11)

26,109
0.739

(12)

Language

History

1.852***
(0.0990)
-1.551***

(0.509)

2.711***
(0.148)

-4.715***

(0.871)

1.877***
(0.158)
0.426
(0.855)

-2.506***

-2.512***

-1.869***

(0.344)
50.23***
(0.431)
5.249***
(0.0777)
21.96***
(0.488)
21.27***
(0.776)
3.083***
(0.425)
1.004***
(0.334)
1.145*
(0.618)

(0.604)
55.32***
(0.623)
4.562***
(0.0872)
6.233***
(0.703)
4.577**
(1.949)
1.153**
(0.519)
3.790***
(0.923)
4.903***
(0.902)

(0.572)
55.75***
(1.251)
5.793***
(0.134)
-1.198
(0.790)
3.558***
(1.149)

-4.529***

(0.724)

-2.116***

(0.803)

65,382
0.709

12,993
0.662

(13)
Art

1.711***
(0.206)
0.631
(1.072)
0.772
(0.792)
35.41***
(0.693)
5.003***
(0.102)
25.15***
(0.909)
20.32***
(1.095)
2.070**
(0.898)
-2.116**
(0.942)

Table B6 Eﬀect of RPM on List Price (Panel)

VARIABLES

Time Trend

RPM15

RPM16

RPM17

log(page)

size

color picture

graph

etc

(1)
ALL

(2)

Non-Child

(3)

Child

(4)

(5)

(6)

(7)

General

Religion

Philosophy

Literature

2.396***
(0.116)

-1.740***

(0.373)

2.393***
(0.117)

-1.813***

(0.374)

-2.636***

-2.760***

(0.427)

(0.430)

-4.144***

-4.198***

(0.588)
7.008***
(1.561)
3.146***
(0.736)
1.730
(1.893)
-0.163
(0.408)
1.186
(0.997)

(0.591)
7.026***
(1.567)
3.133***
(0.736)
1.674
(1.933)
-0.162
(0.408)
1.166
(1.023)

2.913***
(1.059)
4.797
(4.055)
2.223
(3.796)
-6.261
(5.002)
12.19
(11.82)
16.16*
(8.006)
-0.685
(6.448)
2.431**
(1.170)
-0.396
(3.619)

4.773***
(1.183)
-3.283
(2.745)
-9.437*
(5.154)
-13.20**
(5.454)
17.14
(15.66)
0.0672
(1.684)
-8.736
(9.582)
-4.416
(3.932)
14.19*
(7.590)

2.030
(2.659)
-1.921
(16.02)
8.410
(20.34)
-9.138
(13.08)
61.31***
(13.95)
6.318***
(1.444)
2.487
(6.807)

-45.42***

(16.42)
1.855
(3.534)

4.106***
(1.446)
-8.348
(5.608)
-5.589
(7.483)

45.34***
(4.371)
6.308***
(1.074)
6.337
(6.251)
4.913
(4.174)
0.237
(3.973)

-0.408
(4.293)
20.91
(24.55)
21.32
(28.84)
40.82***
(12.12)
27.38
(23.02)
12.91***
(3.555)
20.20
(12.97)

# of Books

-65.03***

-64.81***

-98.12***

-60.93***

-72.14***

# of Appendix

partial color

picture

DVD, CD, MP3

(1.406)

(1.408)

0.325
(1.460)
1.263**
(0.510)
5.920**
(2.431)

0.313
(1.464)
1.267**
(0.509)
5.897**
(2.431)

(15.29)

(10.58)

(13.47)
-2.456*
(1.240)

95.07***
(18.04)
-12.91
(14.41)

4.469
(6.915)
-1.200
(2.996)

-52.71**
(19.98)
-18.85
(18.40)

-32.69
(20.36)
-3.890
(4.371)
5.410
(9.015)

-3,866***

-3,860***

(232.2)

(233.2)

-5,286**
(2,101)

-8,652***

(2,340)

-3,575
(5,315)

-7,712**
(2,899)

1,331
(8,669)

10,855
0.588
4,334

(1)

Social

2.265***
(0.137)

-1.569***

(0.481)

-2.161***

(0.492)

10,764
0.586
4,296

(2)

Natural

3.614***
(0.577)
-3.502*
(2.029)
-5.170*
(2.872)

-3.318***

-9.659***

-3.963***

(0.726)
8.720***
(3.223)
2.063**
(0.948)
-3.018
(3.233)
-0.189
(1.964)
0.0698
(0.550)
-0.204
(0.625)
-0.110
(1.235)

(2.973)
40.69***
(7.962)
-3.912
(2.411)
1.503
(3.958)
-7.672**
(3.244)
-3.415
(2.919)
-0.608
(4.516)
0.0710
(3.417)

(0.819)
3.427***
(1.220)
1.559**
(0.650)
3.781*
(2.165)
2.203
(1.771)
2.115***
(0.746)
-0.166
(0.488)
-0.181
(1.474)

91

0.845

39

(3)
Tech

118
0.801

63

(4)

104
0.712

54

(5)

Language

History

2.006***
(0.176)

3.832**
(1.569)

-1.662***

-7.109***

(0.457)

-2.585***

(0.610)

3.345***
(1.064)
1.086
(3.840)
5.058
(4.321)
-3.176
(4.102)
7.923
(5.458)
2.660
(1.989)
-8.622
(20.61)

10.81**
(4.917)
-0.133
(3.730)
1.867
(3.397)

(2.580)
-11.64**
(4.595)
-15.34**
(6.966)
54.23***
(14.03)
7.133**
(3.353)

4.118
(3.103)
7.309**
(2.950)
3.108
(2.786)
-14.47
(10.67)

43

0.722

23

91

0.890

54

(6)
Art

1.355
(1.311)
-1.690
(3.306)
-4.772
(4.109)
-3.773
(6.369)
96.25*
(50.20)
4.498***
(0.687)
10.08
(14.33)
-6.510
(9.525)
-9.275
(9.595)
-0.304
(2.210)
0.679
(5.408)

Constant

Obs.
Adj. R2
# of IDs

VARIABLES

Time Trend

RPM15

RPM16

RPM17

log(page)

size

color picture

partial color

picture

graph

etc

# of Books

-65.26***

-67.86***

-68.87***

-47.15***

DVD, CD, MP3

(2.267)
4.732*
(2.485)

(1.526)
9.269*
(5.627)

Constant

Obs.
Adj. R2
# of IDs

-3,578***

-6,403***

-3,019***

(267.4)

(1,125)

(353.1)

6,173
0.568
2,456

144
0.779

57

3,512
0.683
1,411

(3.488)
11.70**
(5.003)
-7,168**
(3,154)

232
0.751
114

(4.974)

-5,822***

(2,162)

-2,419
(2,393)

232
0.576
109

70

0.632

31

Standard errors clustered at IDs are parentheses

*** p<0.01, ** p<0.05, * p<0.1

115

Table B7 Eﬀect of RPM on List Price (Panel, RPM Trend)

(1)
ALL

(2)

Non-Child

(3)

Child

(4)

(5)

(6)

(7)

General

Religion

Philosophy

Literature

2.866***
(1.045)
5.235
(4.029)
-4.268*
(2.476)
12.58
(11.44)
16.27**
(7.852)
0.0603
(6.767)
2.369**
(1.149)
-0.335
(3.528)

4.759***
(1.182)
-3.426
(2.700)
-5.261**
(2.441)
16.49
(15.36)
-0.146
(1.338)
-8.174
(9.504)
-4.597
(3.993)
14.11*
(7.497)

2.229
(2.597)
-3.617
(15.39)
4.944
(6.197)
60.75***
(13.88)
6.345***
(1.561)
4.496
(7.141)

-47.56***

(15.69)
0.475
(3.354)

4.106***
(1.446)
-8.348
(5.608)
2.759
(4.188)
45.34***
(4.371)
6.308***
(1.074)
6.337
(6.251)
4.913
(4.174)
0.237
(3.973)

-0.960
(3.568)
21.12
(24.60)
7.227
(6.256)
28.13
(23.61)
13.46***
(2.900)
21.85*
(10.79)

-33.19
(20.66)
-3.574
(4.350)
5.404
(9.024)
927.6***
(101.0)

118
0.800

63

(11)

(2.575)
-4.265
(2.746)
54.26***
(14.04)
7.128**
(3.353)

4.300
(3.310)
7.292**
(2.932)
3.009
(2.654)
-14.17
(10.45)

95.97***
(18.47)
-11.02
(13.56)

4.469
(6.915)
-1.200
(2.996)

-46.87***

(14.47)
-19.41
(18.07)

497.2***
(72.18)

519.9***
(32.32)

499.3***
(117.6)

104
0.707

54

(12)

3.409***
(1.041)
2.049
(3.650)
-0.728
(2.801)
7.657
(5.580)
2.746
(1.943)
-7.701
(20.61)

10.29**
(5.010)
-1.866
(3.620)
2.353
(3.540)

43

0.719

23

91

0.890

54

(13)
Art

1.349
(1.309)
-2.361
(3.410)
-1.803
(2.085)
95.95*
(49.98)
4.477***
(0.701)
10.97
(13.83)
-6.357
(9.372)
-8.523
(9.085)
-0.431
(2.143)
0.700
(5.130)

551.7**
(230.3)

91

0.843

39

(10)
Tech

(0.442)

(0.325)
3.437***
(1.220)
1.562**
(0.653)
3.772*
(2.164)
2.206
(1.769)
2.111***
(0.747)
-0.157
(0.489)
-0.178
(1.475)

VARIABLES

Time Trend

RPM

2.395***
(0.116)

-1.666***

(0.353)

2.392***
(0.117)

-1.755***

(0.354)

RPM Trend

-1.141***

-1.144***

log(page)

size

color picture

graph

etc

(0.229)
7.016***
(1.562)
3.150***
(0.736)
1.717
(1.893)
-0.156
(0.408)
1.185
(0.998)

(0.230)
7.032***
(1.567)
3.136***
(0.736)
1.661
(1.933)
-0.156
(0.408)
1.165
(1.024)

# of Appendix

partial color

picture

DVD, CD, MP3

Constant

Obs.
Adj. R2
# of IDs

VARIABLES

Time Trend

RPM

0.329
(1.459)
1.259**
(0.510)
5.907**
(2.430)
937.1***
(22.16)

10,855
0.588
4,334

(8)

Social

2.264***
(0.137)

-1.501***

(0.453)

0.315
(1.463)
1.265**
(0.510)
5.886**
(2.430)
937.7***
(22.18)

10,764
0.586
4,296

(9)

Natural

3.640***
(0.579)
-3.177
(2.005)

log(page)

size

color picture

partial color

picture

graph

etc

# of Books

DVD, CD, MP3

Constant

Obs.
Adj. R2
# of IDs

(0.298)
8.723***
(3.224)
2.065**
(0.948)
-3.062
(3.228)
-0.234
(1.958)
0.0696
(0.551)
-0.198
(0.625)
-0.120
(1.235)

-65.26***

(2.267)
4.712*
(2.484)
963.4***
(33.94)

6,173
0.568
2,456

(1.109)
40.29***
(7.980)
-3.795
(2.501)
0.921
(3.945)
-7.647**
(3.413)
-3.692
(2.989)
-0.402
(4.530)
0.121
(3.430)

842.8***
(80.36)

# of Books

-65.04***

-64.81***

-99.77***

-60.94***

-72.31***

(1.406)

(1.408)

(15.37)

(10.36)

(12.97)
-3.158**
(1.226)

Language

History

2.006***
(0.176)

3.830**
(1.564)

-1.603***

-7.175***

RPM Trend

-0.822***

-3.078***

-1.105***

-67.86***

-68.71***

-47.56***

(1.526)
9.266*
(5.627)
1,003***
(19.07)

(3.412)
11.71**
(4.970)
513.9***
(63.21)

(4.950)

884.6***
(55.45)

299.5
(300.9)

144
0.777

57

232
0.566
109
Standard errors clustered at IDs are parentheses

3,512
0.683
1,411

232
0.751
114

70

0.630

31

*** p<0.01, ** p<0.05, * p<0.1

116

Chapter 3. Impact of Resale Price Maintenance (RPM) Reg-

ulation on the Used Book Market

3.1 Introduction

The most signiﬁcant diﬀerence from the existing resale price maintenance (RPM) reg-

ulation, implemented in November 2014, is a book which is published more than 18

months ago is newly included in the new regulation coverage. Before the new RPM

regulation, price competition among large online bookstores and low search cost (var-

ious search engines instantly compare prices of a book at all online bookstores if we

enter the title of the book) result in price cuts for books that are not subject to regula-

tion. Discount rates for those books diﬀered depending on the sales and popularity of

a book and inventory management of each online bookstore. However, in many cases,

a book published more than 18 months ago was sold at online bookstores at sometimes

more than 40% discounted price. In this situation, newly-established price regulation

for the non-newly published book,90 directly led to massive price increases and have

unintentionally led to increasing demand for used books that are substitution goods

for new books.

As a result of the increased price of books, the number of potential buyers in

the used book market platform increases, and it causes the matching cost (or searching

cost, or transaction cost) to decrease in the used book market platform. This kind of

network eﬀect is the property of a two-sided market. Subsequently, decreased searching

and matching cost in the used book market caused by more buyers on the platform

also could raise an incentive of used-book sellers to attend the platform even though

the price of a used book has not changed after the new RPM regulation. More sellers

in the platform decrease the matching or transaction cost of used book buyers again,

and it consecutively attracts customers of a new book and converts the demand to an

online used book market platform. Accordingly, the RPM regulation which signiﬁcantly

increases the price of a non-newly published book could trigger the positive network

eﬀect in the used book market platform and sequentially and repeatedly increase the

number of players in the platform and decrease the demand in the new book market.

The used book market has rapidly grown since the new RPM regulation was

90

In this study, a non-newly-published book refers to a book that is published more than 18 months ago.

117

enacted. For example, 523,401 kinds of used books were available at Aladin (No.1

online book company in the used book market (www.aladin.co.kr)) immediately before

the new RPM regulation was in eﬀect (at Oct.25.2014). However, this number increased

to 814,306 on Apr. 24. 2016.91 The Aladin also opened three more used book stores

interworking with their online platform in 2017. The YES24 (The second biggest

online book company (www.yes24.co.kr)) newly started their business for dealing used

books after Nov. 2014 and established a strategic alliance with the Yongpung (The

ﬁfth biggest ﬁrm operating both online and oﬄine stores) to buy used books for sales.

According to a recent news report, their revenue from the used book part has been

growing on average 30% for each month.92 As a result of the remarkable growth, the

YES24 also opened oﬄine used bookstores in Seoul.

In addition to aggressive entry into the used book market, large online ﬁrms

devised a new sales strategy to cope with the new RPM regulations. They call it the

“Buy-Back” service, which is not under the RPM restriction. By using this service,
customers can buy a new book at a largely discounted price (40∼50%) under the con-
tract that they will re-sell their books to online stores after a particular predetermined

time. The bottom line is that the rapid growth of the used book market – which is

also dominated by online ﬁrms – makes it harder to achieve the policy purpose of new

RPM regulation. However, paradoxically, all the above changes are triggered by the

new RPM regulation itself. Also, as we can see from the case of the “Buy-Back” ser-

vice, large online ﬁrms are capable of avoiding regulations by looking at the blind spot

of regulation.

Next, this paper examines the characteristics of the used book market and in-

vestigates how the new regulations led to the rapid growth of the used book market.

When trading a used book online, buyers want to check the quality of the used book,

unlike a new book. Thus, asymmetric information problems between the seller and

the buyer occur in the process. The asymmetric information can be an obstacle in the

online used book market to grow. However, an online platform can solve the asym-

metric information problem by using various methods. First, sellers in the platform

can upload the picture of their used books in addition to the speciﬁc descriptions to

91 This number was over one million in 2019, and some news reported the rapid growth of the used book

market. (http://news.heraldcorp.com/view.php?ud=20160425000116)

92 The growth rate becomes 2300% annually.

118

inform the condition of a book. Second, an online platform has system evaluating and

managing the sellers’ previous transactions. After a transaction, buyers can evaluate

their purchasing experience and check about the quality of the used book (e.g., same

as expected, above expected, and below expected). Based on the buyers’ evaluation,

other buyers can observe the seller’s reputation from previous transactions when they

buy a used book from a seller. That is, an online platform can manage seller reputation

to overcome asymmetric information exiting on online market for used goods. Third,

online platforms have a payment system suitable for dealing in second-hand goods.

Under the payment system, sellers can obtain the money from the platform ﬁrm after

a buyer receives the product and decides to buy it.

Furthermore, online ﬁrms operating used book market platforms have critical

advantages in used book tradings compared to small oﬄine bookstores. In the case

of a new book, an oﬄine store has almost all books for sale. However, it is not the

case with a used book. An oﬄine bookstore only has a little portion of used books

among all publications. Consequently, it is highly likely that there is no stock for a

book that a buyer wants to purchase in the oﬄine store. Also, one more condition is

needed to achieve a transaction for a used book. Even though an oﬄine store holds a

used book a buyer wants to buy, the quality of the used book should meet with the

quality that the buyer wants. Thus, the possibility that there exists a used book that

a consumer wants with the desired condition is quite low (double coincidence). On the

other hand, online used book stores have a lot more books with diverse quality than

an oﬄine bookstore.

Moreover, there is also a transaction cost to visit the local store to check the

condition of a used book. We should visit the store to check the quality of a used

book, and if it does not meet the buyer’s preference, the transaction will not occur

after wasting the buyer’s visiting cost. To summarize, much more books with a variety

of quality, various systems and methods to overcome asymmetric information, and no

physical visiting cost become reasons to make people prefer online bookstores to local

oﬄine stores when customers buy a used book.

According to the Publishing Industry Survey by Publication Industry Promotion

Agency of Korea (KPIPA) which asks oﬄine bookstore owners about the eﬀect of the

rapid growth of used book market on their business (Table 3.1), the proportion of

119

respondents who answered as positive, neutral and negative was 4.3%, 30%, and 60%,

respectively. The negative answer was higher for smaller bookstores based on sales and

the number of employees. The percentage of respondents who answered positively was

17% in the case of bookstores with 10 or more employees, which is much higher than

those of small bookstores.

Table 3.1 Eﬀect of Used Book Market Growth on Oﬄine Stores

Response

Total

By Sales (, Won)

Less Than 10 Million
10 ∼ 50 Million
50 ∼ 100 Million
Greater than 100 Million

By Employees

Less Than 2
3∼4 Employees
5∼9 Employees
More Than 10

814

48
525
143
90

552
153
74
35

(Unit: %)
Positive Neutral Negative

4.2

35

60.8

6.3
3.6
2.1
10

4.2
1.3
4.1
17.1

25
33.7
41.3
35.6

33.9
38.6
39.2
28.6

68.8
62.7
56.6
54.4

62
60.1
56.8
54.3

DATA: KPIPA (2017)

In the following Section 3.2, I construct the comprehensive theoretical model

including the used book market, which has properties of a two-sided market and is

linked with the new book market. From the model, this paper investigates the eﬀects of

new RPM regulations on the used book market in Section 3.3. Section 3.4 summarizes

the theoretical analysis and suggest another method instead of price regulation to

achieve the purpose of resale price maintenance regulation on the book market of

Korea.

3.2 Model

3.2.1 Structure of Used Book Market

In the case of a new book, if the books have the same ISBN, they are the exact same

products. However, even though used books have the same ISBN, the quality of used

books could be diﬀerent. Thus, we need to introduce the quality variable (q) of a

used book. As I discuss in the introduction, online used book market platforms have

120

some critical advantages compared to oﬄine stores. Based on the advantages – much

more variety of used books with various quality, and lower search costs compared to

oﬄine stores – this paper assumes that all used books are traded on an online platform.

Figure 3.1 shows the structure of the used book market that this paper assumes and

analyzes.

Note: Vi(q): Seller’s valuation for a used book with quality q, Xi: Seller’s location on a linear city.
Sellers are uniformly located on the linear city. Thus, Xi ∼ U[0,1]. τi(N d): Transaction cost of
the Type 1 seller, N d represent the number of buyers on the platform. γ: commission fee imposed
by an oﬄine book store. C: commission fee imposed by an online platform to Type 1 and Type
2 sellers. Vi is a buyer’s valuation for a new book. αi: an individual parameter representing the
preference for a used book. τi(N s): Buyer’s transaction cost where N s is the number of sellers on
the platform.

Figure 3.1 Used Book Market

Online Firm

Online ﬁrms operate platforms on which sellers and buyers transact each other. The

platform has properties of a two-sided market; that is, there exists a network eﬀect.

If the number of seller increases, buyer’s transaction cost or matching cost decrease

and sellers’ waiting cost or a matching cost also decreases as the number of buyers

121

increases. This paper assumes that online platform ﬁrms charge a commission fee (C)

only to the sellers when a transaction is completed, and it becomes their revenue from

operating a platform.

Large online ﬁrms can quickly expand their business to the used book market

by adding a menu or a category for used books on their web pages. When they expand

their business to used book markets, their market share in the new book market would

signiﬁcantly aﬀect the market share of the used book market. If a consumer usually

buys a new book in “A” online store, the person is highly likely to buy a used book on

“A” online platform because one-click converts the web page from a new book mar-

ket to used book platform. If a ﬁrm wants to enter the used book market, it should

pay a substantial initial investment cost, because the used book market is already an

oligopoly market by some large online ﬁrms. In Korea, only four major online ﬁrms

explain more than 90% of the total revenue of the online book market. Thus, this

paper additionally supposes that there is no entrant to the used book market platform.

Oﬄine Store

Based on the above the assumption that all used books are traded on an online plat-

form, oﬄine stores buy used books at the local market from individual sellers and sell

those used books at an online platform. Thus, this paper supposes that an oﬄine store
buys a used book from an individual living in the local at a price Pq − γ and sell it
at a price Pq at an online platform. Pq is the used book price of a certain quality (q),

and γ is commission fee charged by an oﬄine store for selling a used book instead of

an individual seller. Thus, oﬄine stores become agents of the Type 2 Seller in Figure

3.1.

Buyer

Buyers maximize their utility by comparing utilities from buying a used book of a

certain quality or a new book. This paper assumes that buyer’s valuation (Vi) for a
new book is uniformly distributed from 0 to 1. Thus, Vi ∼ U [0, 1]. Also, this paper
introduces the individual parameter αi representing an individual’s preference for a

used book and shows the degree of diminishing marginal utility from used book qual-

ity. People have diﬀerent preference over used-book quality and αi captures individual

122

heterogeneity over the quality of a used book. When a person buys a used book at an

online platform, there is a transaction cost or searching cost (τi(N s)). It decreases as
the number of sellers (N s) increase because of a network eﬀect.

Seller

Sellers decide whether they sell their used books directly at an online platform or sell

it to an oﬄine store located at the center of the linear city. This paper assumes that

sellers are uniformly distributed over the linear city. Xi shows the location of seller
i and Xi ∼ U [0, 1]. Each seller also has the valuation for an own used book with a
certain quality (q). Even though the quality of a used book is the same, the individual

value for the used book can be diﬀerent. Thus, this paper introduces another individual
parameter Vi(q) representing the individual i(cid:48)s heterogeneous valuation for a used book
of quality q. When they sell their used book directly via an online platform (Type 1

Seller), they should pay the commission fee (C) and a transaction cost (τi(N d), N d

means the number of buyers).

If they choose to sell it to an oﬄine store which is

located at the center of the linear city (Type 2 Seller), they pay the commission fee

(γ) imposed by the local oﬄine store and also a transaction cost depending on their

location (Xi).

3.2.2 Demand of Used Book Market

Discrete Used Book Quality Choice Model

The model assumes a used book has one of the two qualities value (qL (low-quality) or
qH (high-quality)) and they are given satisfying the condition that 0 (cid:28) qL (cid:28) qH (cid:28) 1
which mean the values are apart from each other enough to have interior solution.93 If
the value of quality is equal to 1, it is the same with a new book. A buyer i(cid:48)s utility
function for buying a used book of quality j ∈ [L, H] is the following:

ij = Vi · qj − 1
U used
2

αi · q2

j − τi(N s) − P (qj)

(11)

where Vi ∼ U [0, 1] is person i(cid:48)s value for a new book. αi is the degree of diminishing
marginal utility for used book quality which captures individual heterogeneity over the

93 To buy a used book, there is a ﬁxed transaction cost to buy it. Thus if the quality is too low, there is

no market for a low-quality used book.

123

quality of a used book. τi(N s) is transaction cost or searching cost that buyers need

to pay when they use the online platform, and it shows the network eﬀect of a two-

sided market; that is the online platform.94 P (qj) is price of used book with quality
j ∈ [L, H]. The price of a used book is endogenously decided by supply and demand
for the used book.

Conditions and Properties of Utility Function

This paper imposes some conditions that the above utility function should satisfy.

(Condition 1) Utility from getting a used book without any cost cannot be greater

than the utility from the new book. From this condition, the lower boundary of αi is

derived. That is, αi > 0.

Vi · qj − 1
2

αi · q2

j < Vi

(Condition 2) Marginal utility of increasing used book quality should be always

greater than 0 for all possible quality levels. The upper boundary of αi is derived from

this condition as αi < Vi

Vi − αi · qj > 0

From above utility function and parameter conditions, we can derive the following

properties.

(Property 1) Buyer i(cid:48)s utility from buying a used book is the increasing function of Vi

(Property 2) Buyer i’s marginal utility from increasing used book quality is also the

increasing function of Vi.

ij

d(U used
)
d(quality)

= Vi − αiqj

94

I assume that

d(τi(N s))

d(N s)

< 0 and

d2(τi(N s))
d(N s)d(N s)

smaller as the number of seller increase.

> 0 Thus, the changes in the network eﬀect become

124

(Property 3) Marginal utility is decreased as the quality increase (by the law of

diminishing marginal utility). However, this eﬀect is diﬀerent depending on αi.

If

αA > αB, then person A’s speed of decreasing marginal utility from increasing used

book quality is faster than person B’s. This individual characteristic makes a person

choose the diﬀerent quality of used book even though they have the same value for

the new book (e.g., VA = VB). Even though a group of people has the same value

for a new book, they could have diﬀerent preference over used-books according to the

quality of the used book. For instance, some people care much about the quality of a

used book while other people do not. αi captures this heterogenous property of people.

The speed of decreasing marginal utility can be independent of Vi. However, αi is not

independent of Vi. It has a diﬀerent boundary according to Vi.

(Property 4) αi has diﬀerent upper and lower boundary according to Vi. First, to

satisfy (Condition 1), αi should be greater than 0. That is the low boundary of αi.

Second, to satisfy the (Condition 2), αi should be less than Vi. An intuitive explanation

for diﬀerent α boundary according to Vi is the following: If VA is low, the marginal

utility of increasing used book quality at q (close to 0) should also be lower than that

of VB (> VA). However, even though the value of marginal utility is low, it should

stay greater than 0 for all quality levels. Thus, the marginal utility of increasing used

book quality should decrease with relatively slow speed for the person whose Vi is low.

Thus, low Vi restricts the maximum speed of decreasing marginal utility. On the other

hand, if Vi has high value, αi could have a wide range of value for Vi.

Based on the above utility function and individual parameters (Vi, αi), It is de-

cided whether to buy a new book, to buy a high-quality used book or a low-quality

used book.

Conditions for Buying High-Quality Used Book

(1) U used

iH > U used

iL

Utility from buying high-quality used book should be greater than that of buying

125

low-quality used book.

iL > 0

iH − U used
U used
H − q2
⇒ Vi · (qH − qL) − (1/2) · αi · (q2
⇒ 2[Vi · (qH − qL) − (P (qH) − P (qL))

H − q2
q2

L

L) − [P (qH) − P (qL)] > 0

> αi

(12)

This paper writes the above equation (12) for the condition as fHL(Vi) > αi. In Figure

3.2, it is area below the red line shows the combination of (Vi, αi) where the utility of

buying high-quality used book is the same with that of buying low-quality used book.

(2) U used

iH > M ax[U new

i

, 0] = M ax[Vi − Cnew, 0]

The second condition means that the utility of buying high-quality used book should

be greater than that of buying a new book. Cnew is cost of buying new book includ-

ing transaction or searching cost. Utility from buying high-quality used book should

greater than both buying a new-book and zero (buy nothing).

i

iH > M ax[U new

U used
, 0]
H − τi(N s) − P (qH) > M ax[Vi − Cnew, 0]
⇒ Vi · qH − (1/2)αi · q2
⇒ 2[Vi · qH − τi(N s) − P (qH) − M ax[Vi − Cnew, 0]]

> αi

q2
H

(13)

This paper writes the above equation (13) for the condition as fHN (Vi) > αi. In Figure

3.2, it is area below the green line. Thus, the area satisfying both fHL(Vi) > αi and

fHN (Vi) > αi buy high-quality used book. It is overlapped area where below the green

line and below the red line, that is yellow shaded area in Figure 3.2.

Area(Buy High Quality U sed Book) = (Vi, αi)

s.t

min[Vi, fHL(Vi), fHN (Vi)] > αi > 0

Conditions for Buying Low Quality Used Book

(1) U used

iH < U used

iL

This condition is the same with the equation (12) with the reverse inequality sign.

Thus, the condition can be written fHL(Vi) < αi. In Figure 3.2, it is area above the

126

This graph is drawn by using the parameter values as following:
Cnew = 0.7, qH = 0.7, qL = 0.4, P (qH ) = 0.3, P (qL) = 0.15 and
τ (N s) = 0.02.

Figure 3.2 Discrete Used Book Quality Model

red line.

(2) U used

iL > M ax[U new

i

, 0] = M ax[Vi − Cnew, 0]

⇒ ViqL − (1/2)αiq2
⇒ 2[ViqL − τi(N s) − P (qL) − M ax[Vi − Cnew, 0]]

L − τi(N s) − P (qL) > M ax[Vi − Cnew, 0]

> αi

(14)

q2
L

The second condition means that the utility of buying low-quality used book should

be greater than that of buying a new book. This paper also writes the equation (14)

as fLN (Vi) > αi. In Figure 3.2, it is area below the blue line. Thus, the area satisfying

both conditions – fHL(Vi) < αi and fLN (Vi) > αi – buy low-quality used book (gray

shaded area).

Area(Buy Low Quality U sed Book) = (Vi, αi)

s.t

min[Vi, fLN (Vi)] > αi > fHL(Vi)

127

Conditions for Buying a New Book

(1) U used

iH < M ax[U new

i

, 0] = M ax[Vi − Cnew, 0]

This condition can be written as fHN (Vi) < αi from the above equation (13) which

represents the area above the green line.

(2) U used

iL < M ax[U new

, 0] = M ax[Vi − Cnew, 0] This condition is the same with the
above equation (14) with reverse inequality, thus it can be written as fLN (Vi) < αi

i

which represents the area above the blue line. The area satisfying both conditions –

fLN (Vi) < αi and fHN (Vi) < αi – buy a new book (purple shaded area).

Eﬀects of Changing Prices and Parameters

Figure 3.3 and 3.4 show the variations in Figure 3.2 according to changes in prices and

transaction cost. Figure 3.3a show the situation when the price of high-quality used

book (P (qH)) increases from 0.3 to 0.33. Dashed lines show the eﬀect of increasing

P (qH) for each line. Both fHL(Vi) and fHN (Vi) shift downward, thereby the area for the

high-quality used book (yellow shaded area) shrink. The dark purple area is transferred

to the new book area, and the dark gray area is converted to the low-quality used book

area from the high-quality used book area. Figure 3.3b shows the situation when the

price of low-quality used book (P (qL)) increases from 0.15 to 0.18. Also, dashed lines

show the eﬀect of increasing P (qL) for each line – fHL(Vi) shifts upward and fHN (Vi)

shifts downward. As a result, the area for low-quality used book (gray shaded area)

shrink. The dark yellow area is changed to high-quality used book area, and the dark

purple area is changed to the new book area from the demand for a low-quality used

book.

Figure 3.4a shows the changes in functions when buyers’ transaction cost at on-

line platform (τi(N s)) increases from 0.02 to 0.05. Dashed lines show the eﬀect of the

increased transaction cost for each function. Both fLN (Vi) and fHN (Vi) shift down-

ward. As a result, both the low and high-quality used book areas shrink and the area

for buying new book expands. The dark purple area is changed to the new book area

from demands for the high and low-quality used book. On the other hand, Figure

3.4b shows the situation when the cost of buying a new book (Cnew) increases from

0.7 to 0.75. Again, dashed lines show the eﬀect of increasing Cnew for each function

128

(a) Eﬀect of Increasing P (qH )

(b) Eﬀect of Increasing P (qL)

Figure 3.3 The Eﬀect of Increasing Used Book Price

NOTE: Graphs are drawn by using the parameter values as following: Cnew = 0.7,
qH = 0.7, qL = 0.4, P (qH ) = 0.3, P (qL) = 0.15 and τ (N s) = 0.02. (a) shows the
eﬀect of increased P (qH ) from 0.3 to 0.33. (b) illustrates the eﬀect of increased
P (qL) from 0.15 to 0.18.

129

(a) Eﬀect of Increasing τ

(b) Eﬀect of Increasing Cnew

Figure 3.4 The Eﬀect of Changing Parameters

NOTE: Graphs are drawn by using the parameter values as following: Cnew = 0.7,
qH = 0.7, qL = 0.4, P (qH ) = 0.3, P (qL) = 0.15 and τ (N s) = 0.02. (a) shows the
eﬀect of increased τ from 0.02 to 0.05. (b) illustrates the eﬀect of increased cost
for buying a new book from 0.7 to 0.75

130

– both fHN (Vi) and fLN (Vi) shift upward. Consequently, dark yellow area converts to

high-quality used book demand, and the dark gray area turns to low-quality used book

area from the new book demand area.

Probability and Demand

Next, we can calculate demand for each product (high-quality used book, low-quality

used book, new book) from the above (Vi, αi) plane graph (Figure 3.2, 3.3 and 3.4).
given that αi ∼ U [0, Vi]

The probability density function (pdf) of (Vi, αi) plane is
and Vi ∼ U [0, 1].

1
Vi

Probability of Buying a High-Quality Used Book

The probability of buying a high quality used book for Vi is calculated by integrating

the area of high quality used book in Figure 3.2. This paper writes the probability

function as f qH

pr (Vi). The blue line in Figure 3.5 shows the probability of buying high

quality used book according to Vi.

P r(min[Vi, fHL(Vi), fHN (Vi)] > αi > 0) = f qH

pr (Vi)
= min[Vi, fHL(Vi), fHN (Vi)] · p.d.f
= min[Vi, fHL(Vi), fHN (Vi)] · 1/Vi

Probability of Buying a Low-Quality Used Book

At the same way, the probability of buying a low-quality used book is obtained by

integration of low-quality area which is as follows:

P r(min[Vi, fLN (Vi)] > αi > fHL(Vi)) = f qL

pr (Vi)
= (min[Vi, fLN (Vi)] − fHL(Vi)) · p.d.f
= (min[Vi, fLN (Vi)] − fHL(Vi)) · 1/Vi

The red line in Figure 3.5 shows the probability of buying low-quality used book ac-

cording to Vi.

131

NOTE: Graphs are drawn by using the parameter values as following: Cnew = 0.7,
qH = 0.7, qL = 0.4, P (qH ) = 0.3, P (qL) = 0.15 and τ (N s) = 0.02.

Figure 3.5 Probability of Buying Each Book for Vi

Probability of Buying a New Book

The integration of the new book area which means the probability of buying a new

book in Figure 3.2 is calculated as the following:

P r(Vi > αi > M ax[fHN (Vi), fLN (Vi)

= (Vi − M ax[fHN (Vi), fLN (Vi)

(cid:12)(cid:12)(cid:12)Vi > Cnew]) = f new
(cid:12)(cid:12)(cid:12)Vi > Cnew] · 1/Vi

pr (Vi)

The green line in Figure 3.5 represents the probability of buying a new book according

to Vi. According to Figure 3.5, if a group of people’s value for a new book is 0.7, then

50% of people in the group buy a low-quality used book, and the rest 50% of people

buy a high quality used book. In the case of people whose value for a new book is

0.9, 50% of people buy a high-quality used book and the rest 50% of people buy a new

book. Figure 3.6 shows the situation when the price of low-quality used book increases

132

from 0.15 to 0.18. The dashed line shows the eﬀect of increasing the low-quality book

price. As a result, the probability of buying a low-quality used book decrease and

probabilities for both a high quality used book and a new book increase for each Vi.

NOTE: Graphs are drawn by using the parameter values as following: Cnew = 0.7,
qH = 0.7, qL = 0.4, P (qH ) = 0.3, and τ (N s) = 0.02. P (qL) is increased from 0.15
to 0.18.

Figure 3.6 Changes in Probability Functions by Increased P (qL)

From Figure 3.5 and 3.6 which show the probability of buying each book accord-

ing to Vi, we can obtain the demand function for each book by integrating probability

function as following:

• Demand for a j quality used book

(cid:90) 1

f qj
pr (Vi)dVi

• Demand for a new book

0

(cid:90) 1

0

f new
pr (Vi)dVi

133

(15)

(16)

NOTE: Graphs are drawn by using the parameter values as following: Cnew = 0.65,
qH = 0.7, qL = 0.4, P (qL) = 0.15, and τ (N s) = 0.1. The dashed line shows the
eﬀect of decreased τ (N s) from 0.1 to 0.05.

Figure 3.7 Demand Curve of High Quality Used Book

Figure 3.7 shows the demand curve of a high-quality used book and the dashed

line describes the change of demand when the transaction cost (τ (N s)) decreases.

3.2.3 Supply of Used Book Market

Type of Sellers

As I discuss in the introduction, this paper assumes there are two types of sellers in a

linear city. Figure 3.8 describes the location of each type of seller.

(Type 1 Seller)

This type of sellers sells its own used book via online platform directly. In the result

of the transaction, they obtain the utility as follows:

i,T ype1 = P (qj) − C − Vi(qj) − τi(N d)
U seller

(17)

where P (qj) is the price j ∈ [L, H] quality used book, C is commission fee they need
to pay for using online platform and Vi(qj) is individual value for own used book of

quality j. Although the quality is the same, the individual value for the used book can

134

be diﬀerent among people. The assumption that Vi(qj) ∼ U [0, vj] considers individual’s
heterogeneous values on own used book with quality j. That is, the owner’s value is

distributed over uniformly between 0 and vj (vH > vL). τi(N d) is seller’s transaction

cost or a user cost (except the commission fee C) at online platforms such as a waiting

cost for selling a book or the cost for uploading used book information at online. The

transaction cost also has a network eﬀect like the same with τi(N s) in the demand side
and also has the same properties. It decreases as the number of buyers (N d) increase

because of network eﬀect

< 0

and the degree of network eﬀects become

(cid:16)d(τi(N d))

d(N d)

(cid:17)
(cid:16) d2(τi(N d))

d(N d)d(N d)

(cid:17)

> 0

.

smaller as the number of seller increase

(Type 2 Seller)

This type of sellers sell own used book with quality j to a local oﬄine store at a price
P (qj) − γ and the local oﬄine store sell again the used book at an online platform
at P (qj). Thus, the γ become the income of the oﬄine bookstore that acts as a sales
agent. For the Type 2 seller, there is a transaction cost t|Xi − 1/2| to visit the local
oﬄine store. The utility function of the Type 2 seller is given as follows:

i,T ype2 = P (qj) − γ − t|Xi − 1/2| − Vi(qj)
U seller

(18)

The Location and Individual Values of Each Type Seller

The seller with (Xi, Vi(qj)) becomes (Type 2) seller if

U seller

i,T ype2 > M ax[U seller

i,T ype1, 0]

From the condition U seller

i,T ype2 > U seller

i,T ype1, the location of (Type 2) seller is derived as

Xi ∈(cid:16)1

2

− τi(N d) + C − γ

t

,

1
2

+

τi(N d) + C − γ

t

(cid:17)

and the condition U seller

i,T ype2 > 0 restrict the individual value for own used book (vertical
values) as in Figure 3.8. At the same way, we can obtain the location and range of

135

Figure 3.8 Locations and Individual Values of Each Seller

individual value (Xi, Vi(qj)) of (Type 1) seller.

(Supply of Used Book)

The supply of used book with quality j is the area of both (Type 1) and (Type 2) seller

in Figure 3.8. Thus, we can derive the supply of quality j used book as follows:

Supply = [P (qj) − Cj − τi(N d) +

(τi(N d) + Cj − γ)2

t

] · vj

(19)

The online platform commission fee (C) and oﬄine store user cost (γ) are also

decided endogenously by the proﬁt maximization of online platform company and of-
ﬂine store. This paper derives the optimal C∗ and γ∗ in the following section.

Online Platform Firm

According to the model this paper analyzes, an online ﬁrm operates the platform for

the used book market. Additional cost for operating the used book market platform

136

is almost zero because the ﬁrm was already operating the webpage for selling a new

book. Thus, it only cost some ﬁxed costs at one time to set up a platform, and the

marginal cost of operating the platform for a used book is almost zero. This paper set

the proﬁt function of an online ﬁrm from operating the online platform as follows:

(Proﬁt function of online ﬁrm)

on = CL · Q(CL) + CH · Q(CH)
Πused

(20)

where Cj, is an online platform commission fee for selling a j quality used book on the

platform. Q(Cj) is used book quantities of quality j traded at the used book market

platform which is calculated from equation (19) as follows:

Q(Cj) =

1
2

· [P (qj) − C − τi(N d) +

(τi(N d) + C − γ)2

t

] · vj

When two online platform ﬁrms do Bertrand price competition, the commission fee

goes down to zero, that is, its marginal cost. On the other hand, if the collusion
between two platform ﬁrms is possible, they choose the optimal commission fee (C∗
j )
from the following equation which is the ﬁrst order condition of the proﬁt function.95

Q(Cj) + Cj · d(Q(Cj))
d(Cj)
Q(Cj)

Cj =

−d(Q(Cj))/d(Cj)

= 0

where

d(Q(Cj))

d(Cj)

(< 0) =

2(Cj + τi(N d) − γ)

− 1

·(cid:104)2(Cj + τi(N d) − γ)

t

1 − dτi(N d)
dQ(Cj)

t

(21)

(22)

(cid:105)

− 1

When the ﬁrm decides the optimal commission fee (C∗
j ), there is an additional network
eﬀect aﬀected by the fee. The commission fee changes the number of demand (N d),

and it will also aﬀect the transaction cost (τi(N d)). Thus the transaction cost is also

95 Equation(21) and (24) consist of reaction functions of an online ﬁrm and an oﬄine store. Thus, the

intersection of two curves on (Cj, γj) plane becomes an equilibrium.

137

the function of the commission fee. In equation (22),

captures the indirect

network eﬀect considered by the ﬁrms when they decide the optimal commission fee.

dτi(N d)
dQ(Cj)

If they do not internalize the indirect network eﬀect, equation (22), is simpliﬁed to
2(Cj + τi(N d) − γ)

is lower when they internalize the indirect

− 1. The optimal C∗

j

t

network eﬀect.

Oﬄine Store

Oﬄine book store purchases used books from individuals ((Type 2) seller), and sell it

via an online platform. It occurs transaction cost (τL(N d) or waiting and matching

cost) for oﬄine books store when they use an online platform to sell the used books.

I assume that the transaction cost of the oﬄine book store is lower than that of in-

dividuals. That is τL(N d) < τi(N d). The reasons for this assumption are: First, an

oﬄine book store sells used books frequently; thus they can have a higher reputation

and a longer selling history – which make selling easier and decrease a transaction

cost or a waiting cost – than an individual seller at an online. Moreover, oﬄine book

stores are more specialized in selling used books using the online platform because it is

their job. Second, because of the economy of scale, oﬄine book stores can spend less

transaction cost per used book. Importantly, this paper assumes that the transaction

cost of individual sellers (Type 1 Seller) changes more sensitively than that of oﬄine

(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) d(τL(N d)

d(N d)

(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) <

(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) d(τi(N d)

d(N d)

(cid:12)(cid:12)(cid:12)(cid:12)(cid:12). The

book stores as the number of buyers increase. That is,

Figure 3.9 illustrates this assumption.

(Proﬁt Function of Oﬄine Store)

The proﬁt function of an oﬄine store in the used book market is calculated as follows:

of f = [(P (qH) − CH) − (P (qH) − γH) − τL(N d)] · QH
Πused
+ [(P (qL) − CL) − (P (qL) − γL) − τL(N d)] · QL

of f (γH)

of f (γL)

= [γH − CH − τL(N d)] · QH

of f (γH) + [γL − CL − τL(N d)] · QL

of f (γL)

(23)

where Qj

of f (γj) is the used book quantity of quality j which is traded through the
oﬄine store (Type 2 Seller) at an online platform, and it is calculated from the area of

138

Figure 3.9 Network Eﬀect in Transaction Cost

(Type 2) seller in Figure 3.8 as follows:

of f (γj) = [(Cj + τi(N d) − γj)2 + 2(C + j + τi(N d) − γj)(P (qj) − Cj − τi(N d)] · vj
Qj
t

The oﬄine store decides the optimal commission fee for each quality used book γH and

γL to maximize the proﬁt from the equation (23).

From the ﬁrst order condition of proﬁt function, γ∗

j is calculated from the fol-

lowing equation:96

γj = Cj + τL(N d) − Qj
d(Qj

of f (γj)
of f (γj))
d(γj)

+

d(τL(N d))
d(Qj
of f )

· Qj

of f

(24)

96 This is the second order equation with respect to γj. However, γj has unique solution given the condition

which is Cj + τL < γj < Cj + τi

139

posite eﬀects.

d(τL(N d))
d(Qj
of f )

(< 0) lowers the optimal γ∗

j and

d(τi(N d))
d(Qj
of f )

(< 0) increases it

where

d(Qj

of f (γj))
d(γj)

=

2t · d(τi(N d))
of f )

d(Qj

2(P (qj) − γ)
· [P (qj) − Cj − τi(N d)] − t

(25)

When the oﬄine store internalize the indirect network eﬀect, there are two op-

compared to the situation when they ignore the indirect network eﬀect.97 Intuitively,
d(τL(N d))
d(Qj
of f )

j , because τL is the cost of an oﬄine store and decreas-

lowers the optimal γ∗

ing γ increases the demand (N d) and lowers its cost (τL). Thus, we can interpret this

eﬀect as the indirect cost saving eﬀect of decreasing γ. However, decreasing γ also

decreases the τi(N d), and it increases (Type 1 Seller) while decreasing (Type 2 seller).

Thus, we can interpret this eﬀect as the competitiveness eﬀect of rivals which increases
the optimal γ∗
j .

Figure 3.10 Reaction Curves of Online Firm and Oﬄine Store

Equation (21) and (24) consist of reaction functions of the online platform ﬁrm

and the oﬄine store. RCon(γ) is reaction curve of online ﬁrm according to oﬄine

store’s decision (γ) and RCof f (C) shows the optimal γ given online ﬁrm’s commission

97

In this case, equation (24) and (25) are simpliﬁed where

d(τL)
d(Qj
of f )

=

d(τi(N d))
d(Qj
of f )

= 0.

140

fee (C). The optimal commission fee imposed by the online platform decreases as the

γ increases. The reason is that the quantity of used book transacted via the online

platform is decreased if the γ increases, as we can check easily from Figure 3.8. That

is, increasing C or γ decreases the online ﬁrm’s quantity, thus C and γ have strategic

substitute relation for the online ﬁrm’s decision. On the other hand, oﬄine store’s

optimal price increases as the commission of the online ﬁrm increases. For oﬄine

stores, they compete with the online platform with price γ. Individual sellers become

(Type 1 Seller) or (Type 2 seller) by comparing C and γ. Thus, for the oﬄine stores,

C and γ have strategic complementary relation. Moreover, note that C also acts as

the cost of oﬄine stores. Increasing C means aggravated cost condition of the oﬄine

store and push the oﬄine store’s price upward.

3.3 Eﬀects of New RPM Regulation and the Equilibrium of Market

3.3.1 Proﬁt of Oﬄine Stores

Because of the new RPM regulation, a part of demand in new book market moves

to the used book market. As a result, the number of demand (N d) in the used book

market increases that lowers the transaction cost of the online platform by the network

eﬀect. Also, an increased number of buyers at the platform directly aﬀects the proﬁt

of an oﬄine store. I investigate two cases when the two platform ﬁrms do Bertrand

price competition with commission fee (Cj) and when the collusion for Cj is possible.

(Proposition 1)

If the following condition is satisﬁed, the proﬁt of oﬄine store is decreased as the

number of demand increases in used book market98 when two online platform ﬁrms do

(cid:12)(cid:12)(cid:12) d(τi(N d))
(cid:12)(cid:12)(cid:12)
(cid:12)(cid:12)(cid:12)
(cid:12)(cid:12)(cid:12)d(τL(N d))

d(N d)

d(N d)

Bertrand price competition.

(1 <)

P (qj) − γj

P (qj) − τi(N d) − Cj

<

(Proposition 2)

98 The proof is in appendix

141

(26)

If we consider the case when the online ﬁrms can collude, the condition that the proﬁt

of an oﬄine store is decreased as the number of demand increases in the used book

market is changed as the follows:

(1 <)

P (qj) − γj

P (qj) − τi(N d) − Cj

<

(cid:12)(cid:12)(cid:12) d(τi(N d))
(cid:12)(cid:12)(cid:12) d(τL(N d))

d(N d)

d(N d)

(cid:12)(cid:12)(cid:12)
(cid:12)(cid:12)(cid:12)

d(Cj)
d(N d)
d(Cj)
d(N d)

+

+

(27)

Figure 3.11 Eﬀect of Increasing Demand on Reaction Curves

If an individual’s amount of decreasing transaction cost caused by the network

eﬀect from increasing demand is greater than that of the oﬄine store and satisfy the

above condition, then the proﬁt of the oﬄine store is decreased. The intuitive reason is
that γj should be Cj + τL(N d) < γj < Cj + τi(N d) to have positive demand (Qj
of f > 0).
In this situation, if the number of demand (N d) increases, the transaction cost of

an individual (τi(N d)) decreases more than that of an oﬄine store (τL(N d)) because

of the reason explained in section 3.5.2. This eﬀect aggravates oﬄine ﬁrms’ proﬁt
condition. Lowered γj can increase the demand (Qj
of f ), but this is also restricted by
decreasing τi(N d). Figure 3.11 describes the eﬀects of increasing demand (N d) on both
reaction curves. As a result of increasing demand (N d), both curves shift to the right.
Consequently, the optimal commission fee C∗
j
oﬄine store γ∗
j can be increased or decreased.

increases and the optimal price of the

142

3.3.2 Equilibrium of the Used Book Market
To sum up, the equilibrium of the used book market of quality j {P (qj), Qj, Cj, γj} is
decided by the following equations.

(1) Demand for j quality used book

(cid:90) 1

0

f qj
pr (Vi)dVi

(2) Supply for j quality used book

[P (qj) − Cj − τi(N d) +

(τi(N d) + Cj − γj)2

t

] · vj

(3) Reaction function of online ﬁrm99

Cj =

Q(Cj)

−d(Q(Cj))/d(Cj)

(4) Reaction function of oﬄine store

γj = Cj + τL(N d) − Qj
d(Qj

of f (γj)
of f (γj))
d(γj)

+

d(τL(N d))
d(Qj
of f )

· Qj

of f

s.t Cj + τL(N d) < γj < Cj + +τi(N d)

(15)

(19)

(21)

(24)

Figure 3.12 shows the changes in equilibrium in high quality used book market

caused by the new RPM regulation. At ﬁrst, it increases the number of demand

(N d) for high quality used book which is illustrated in Figure 3.4b. After deriving

the probability functions from the (Vi, αi) plane, we can obtain the new demand curve

(dashed negative slope in Figure 3.12) by integrating the changed probability functions.

Also, more buyers in the online platform decrease the transaction costs (τi(N d)) of

used booksellers, thereby the number of seller increases. These changes also aﬀect the

optimal decisions of ﬁrms in the used book market described in the reaction curve

(Figure 3.10). As a result of the shock from new RPM regulation, sellers and online

platform companies change their decisions, which changes the supply curve (dashed

99

If online ﬁrms do Bertrand price competition, C∗

j = 0.

143

upward slope in Figure 3.12.

Figure 3.12 Eﬀect of Increasing Demand on Used Book Market

3.4 Conclusion

The reinforced resale price maintenance (RPM) regulation took eﬀect from November

2014 in Korea under the policy purpose to protect a variety in culture and knowledge by

supporting the business of small-scale publishing companies and oﬄine bookstores. The

biggest changes from the previous RPM regulation is that a non-newly-published book

– which is published more than 18 months ago – is also included in the new regulation

coverage. As a result, the new regulation naturally causes a massive price increase of

non-newly-published books. A non-newly-published book has substitution goods, that

is a used book. Thus, the new RPM regulation unintentionally sent customers for a

new book to a used book market and triggers the rapid growth of the used book market

in Korea.

The problem is that large online ﬁrms have also dominated the used book market.

After the new RPM regulation took eﬀect, online ﬁrms aggressively enter the used book

market. As I discuss in the introduction, online ﬁrms are in a better position in the

used book market than oﬄine bookstores because they have a much larger variety of

144

books with diverse qualities. Moreover, large online ﬁrms have various marketing tools

that can legally evade the new RPMregulation as we can see at the case of “Buy-Back”

service. According to the survey by KPIPA, 60% of oﬄine bookstore owners answered

that the rapid growth of the used book market negatively aﬀected their business.

This paper shows that the proﬁts of oﬄine stores from the used book market

can be diminished based on the theoretical model including the used book market,

which has properties of a two-sided market and is linked with the new book market.

If the diminishing transaction cost of an oﬄine store is relatively smaller than that

of an individual seller, when the increasing number of buyers decrease the transaction

cost by the network eﬀect of the two-sided market, then the proﬁt of an oﬄine store

decrease at the used book market.

Based on the problems of new RPM regulation, I suggest considering an alterna-

tive method to accomplish the policy purpose of RPM regulation on the book market.

We can think of direct subsidy instead of existing the price regulation.

Nowadays, many people buy books online after browsing the book at local oﬄine

stores. From the local oﬄine store, they get some beneﬁt, however, they do not pay

anything for that. Oﬄine stores have cost for displaying books and operating the oﬄine

store. However, online ﬁrms get some beneﬁts indirectly from the service oﬀered by

oﬄine local stores. Oﬄine local stores act like a real displaying place for online ﬁrms

which do not have an actual place for display. Thus, there is an externality in the used

book market and the existence of this externality justiﬁes the transfer of proﬁt or direct

subsidy from online ﬁrms to oﬄine bookstores. The direct subsidy as a corrective tax

to ﬁx the externality between online ﬁrms and oﬄine bookstores can more eﬀectively

achieve the policy goal, less distorting the book market than price regulation such as

RPM. In the following research, we can consider the market-friendly incentive design

considering the externality that oﬄine stores serve as display places for online ﬁrms.

145

APPENDIX

146

APPENDIX C

Proof of Proposition 1 and 2

From the proﬁt function of an oﬀ-line store in used book market(equation (23))

of f = [γH − CH − τL(N d)] · QH
Πused

of f (γH) + [γL − CL − τL(N d)] · QL

of f (γL)

(23)

dΠused
of f
dN d =

=

dN d

(cid:105) · Qj
(cid:104) dγj
dN d − dτL
dN d − dCj
(cid:105) · Qj
(cid:104) dγj
dN d − dτL
dN d − dCj
dN d · ∂Qj
of f (γj)
(cid:105) · Qj
(cid:104) − dCj
of f
dN d
∂γj
of f (γj) −
dN d − dτL

dQj

dN d

dN d

dγj

=

of f (γj) +
of f (γj) −

where

=

(cid:104)

γj − Cj − τL(N d)

(cid:105) · dQj

of f (γj)
dN d
of f (γj)
dN d

dτi

· dQj
dN d · ∂Qj
·(cid:104) dCj
dN d · ∂Qj

of f
∂Cj

∂τi

+

of f

of f (γj)

Qj
of f (γj)(cid:30)dγj

dQj

+

dCj

dN d · ∂Qj

of f
∂Cj

of f (γj)

Qj
of f (γj)(cid:30)dγj

dQj

(cid:105)

+

dτi

dN d · ∂Qj

∂τi

of f

where

dQj

of f (γj)
dγj

= P (qj) − γj

and

of f

∂Qj
∂τi

=

∂Qj
of f
∂Cj

= P (qj) − τi − Cj

(28)

Thus, the condition that the proﬁt of an oﬀ-line store decrease as the number of demand

for used book is following :

dN d − dτL

dN d

of f (γj) −

dΠused
of f
dN d < 0

⇒(cid:104) − dCj
⇒(cid:12)(cid:12)(cid:12) dCj

dN d +

(cid:105) · Qj
(cid:12)(cid:12)(cid:12) dCj

(cid:12)(cid:12)(cid:12) <

dN d +

dτL
dN d
P (qj) − γj

⇒ (1 <)

P (qj) − τi − Cj

<

of f (γj)

·(cid:104) dCj
dN d · ∂Qj
(cid:12)(cid:12)(cid:12) · P (qj) − τi − Cj

dQj

of f
∂Cj

Qj
of f (γj)(cid:30)dγj
P (qj) − γj
dτi
dN d
dτL
dN d

(cid:12)(cid:12)(cid:12)
(cid:12)(cid:12)(cid:12)

dτi
dN d

(cid:12)(cid:12)(cid:12) dCj
(cid:12)(cid:12)(cid:12) dCj

dN d +
dN d +

+

dτi

dN d · ∂Qj

∂τi

of f

(cid:105)

< 0

(29)

If we set

dCj
dN d = 0, the case when the on-line ﬁrms do Bertrand price competition
and there is no market power to set the optimal C j, above equation is changed to the

proposition 1.

147

BIBLIOGRAPHY

148

BIBLIOGRAPHY

Armstrong, M. (2006), “Competition in two?sided markets.” The RAND Journal of Economics

37(3): 668–691.

Bajari, P., C. L. Benkard, and J. Krainer. (2005), “House Prices and Consumer Welfare.”

Journal of Urban Economics 58: 474–487.

Bajari, P., and C. L. Benkard. (2005), “Demand Estimation with Heterogeneous Consumers and
Unobserved Product Characteristics: A Hedonic Approach.” Journal of Political Economy, 113
(6): 1239–76.

Bajari, P., Fruehwirth, J. C., Kim, K. I., and Timmins, C. (2012), “A rational expectations
approach to hedonic price regressions with time-varying unobserved product attributes: The price
of pollution.” The American Economic Review, 102(5): 1898–1926.

Bajari, P., and M. Khan. (2005), “Estimating Housing Demand With An Application to Explain-

ing Racial Segregation in Cities.” Journal of Business and Economic Statistics, 23(1): 20–33

Baum-Snow, N. and Lutz, B.F. (2011), “School desegregation, school choice, and changes in

residential location patterns by race.” American Economic Review, 101(7): 3019–46.

Bayer, P., Fang, H. and McMillan, R. (2014), “Separate when equal? Racial inequality and

residential segregation.” Journal of Urban Economics, 82: 32–48.

Bayer, P., Ferreira, F. and McMillan, R. (2007), “A Uniﬁed Framework for Measuring Prefer-

ences for Schools and Neighborhoods.” Journal of Political Economy, 115(4): 588–638.

Bayer, P., Keohane, N. and Timmins, C. (2009) “Migration and hedonic valuation: The case

of air quality.” Journal of Environmental Economics and Management, 58(1): 1–14.

Bayer, P. and McMillan, R. (2012), “Tiebout sorting and neighborhood stratiﬁcation.” Journal

of Public Economics, 96(11-12): 1129–1143.

Bayer, P., McMillan, R., Murphy, A., and Timmins, C. (2016). “A dynamic model of demand

for houses and neighborhoods.” Econometrica, 84(3): 893–942.

Bayer, P., N. Keohane, and C. Timmins. (2009) “Migration and Hedonic Valuation: The Case

of Air Quality.” Journal of Environmental Economics and Management, 58(1) 1–14.

Bento, A., Freedman, M. and Lang, C. (2015) “Who beneﬁts from environmental regulation?
Evidence from the Clean Air Act Amendments.” Review of Economics and Statistics, 97(3): 610–
622.

Bifulco, R. and Ladd, H. F. (2007), “School choice, racial segregation, and test-score gaps: Evi-
dence from North Carolina’s charter school program.” Journal of Policy Analysis and Management,
26(1): 31–56.

Black, S. E. (1999), “Do Better Schools Matter? Parental Valuation of Elementary Education.”

Quarterly Journal of Economics, 114(2): 577–99.

Black, S., and Stephen M. (2011), “Housing valuations of school performance.” Handbook of the

Economics of Education, 3: 485–519

Boustan, L.P. (2010), “Was postwar suburbanization “white ﬂight”? Evidence from the black

migration.” The Quarterly Journal of Economics, 125(1): 417–443.

149

Boustan, L.P. (2012), “School desegregation and urban change: Evidence from city boundaries.”

American Economic Journal: Applied Economics, 4(1): 85–108.

Boustan, L.P. (2013) “Racial Residential Segregation in American Cities.” The Oxford Handbook

of Urban Economics and Planning

Caetano, G. and Maheshri, V. (2017) “School segregation and the identiﬁcation of tipping

behavior.” Journal of Public Economics, 148: 115–135.

Card, D., Mas, A. and Rothstein, J. (2008), “Tipping and the Dynamics of Segregation.” The

Quarterly Journal of Economics, 123(1): 177–218.

Card, D. and Rothstein, J. (2007), “Racial segregation and the black-white test score gap.”

Journal of Public Economics, 91(11): 2158–2184.

Case, K.E. and Shiller, R.J. (1989), “The Eﬃciency of the Market for Single-Family Homes.”

American Economic Review 79 (1): 125–37.

Chay, K.Y. and Greenstone, M. (2005), “Does air quality matter? Evidence from the housing

market.” Journal of Political Economy, 113(2): 376–424.

Cho, S.I. (2014), “Pricing Capability of Retailers and the Eﬀect of e-Commerce: the Case of the

Book Retail Market.” KDI Policy Study

Cho, S.I. (2015), “Study on the Economic Eﬀect of Enhanced Fixed Book Pricing Scheme.” KDI

Policy Study 1–98.

Clapp, J.M. and Ross, S.L. (2004), “Schools and housing markets: An examination of school

segregation and performance in connecticut.” The Economic Journal, 114(499): 425–440.

Clapp, J.M., Nanda, A. and Ross, S.L. (2008), “Which school attributes matter? The inﬂuence
of school district performance and demographic composition on property values.” Journal of urban
Economics, 63(2): 451–466.

De los Santos, B. and Wildenbeest, M.R. (2017), “E-book pricing and vertical restraints.”

Quantitative Marketing and Economics 15(2): 85–122.

Ellison, G. and Ellison, S.F. (2018), “Match quality, search, and the Internet market for used

books” National Bureau of Economic Research, (No. w24197)

Epple, D. (1987), “Hedonic Prices and Implicit Markets: Estimating Demand and Supply Functions

for Diﬀerentiated Products.” Journal of Political Economy, 95(1): 59–80.

Figlio, D. N., and Maurice E. L. (2004), “What’s in a grade? School report cards and the

housing market.” American Economic Review, 94(3): 591–604

Fiva, J. H. and Kirkeben, L. J. (2011), “Information Shocks and the Dynamics of the Housing

Market.” The Scandinavian Journal of Economics, 113: 525-552.

Ghose, A., Smith, M.D. and Telang, R. (2006), “Internet exchanges for used books: An empir-
ical analysis of product cannibalization and welfare impact.” Information systems research, 17(1):
3–19.

Gibbons, S. and Machin, S. (2003), “Valuing English primary schools.” Journal of urban eco-

nomics, 53(2): 197–219.

Gibbons, S. and Machin, S. (2008), “Valuing school quality, better transport, and lower crime:

evidence from house prices.” Oxford Review of Economic Policy, 24(1): 99–119.

150

Gibbons, S., Machin, S. and Silva, O. (2013), “Valuing School Quality Using Boundary Discon-

tinuity Regressions.” Journal of Urban Economics, Volume 75: 15–28,

Grainger, C.A. (2012), “The distributional eﬀects of pollution regulations: Do renters fully pay for

cleaner air?” Journal of Public Economics, 96(9-10): 840–852.

Hamilton, T.L. and Phaneuf, D.J. (2015), “An integrated model of regional and local residential
sorting with application to air quality.” Journal of Environmental Economics and Management,
74: 71–93.

Imberman, S. A., and Lovenheim, M. F. (2016), “Does the market value value-added? Evidence
from housing prices after a public release of school and teacher value-added.” Journal of Urban
Economics, Volume 91: 104–121,

Kane, T. J., Riegg, S. K. and Staiger, D. O. (2006), “School Quality, Neighborhoods, and

Housing Prices.” American Law and Economics Review, 8 (2): 183–212.

Klaiber, H. A., and Phaneuf, D. J. (2010), “Valuing open space in a residential sorting model

of the Twin Cities.” Journal of Environmental Economics and Management, 60(2): 57–77.

Machin, S. (2011). “Houses and schools: Valuation of school quality through the housing market.”

Labour Economics, 18(6): 723–729.

Machin, S. and Salvanes, K.G. (2016). “Valuing school quality via a school choice reform.” The

Scandinavian Journal of Economics, 118(1): 3–24.

Nickell, S. (1981). “Biases in dynamic models with ﬁxed eﬀects.” Econometrica: Journal of the

Econometric Society: 1417–1426.

Nguyen-Hoang, P. and Yinger, J. (2011), “The capitalization of school quality into house values:

A review.” Journal of Housing Economics, 20(1): 30–48.

Poort, J. and van Eijk, N. (2017), “Digital ﬁxation: the law and economics of a ﬁxed e-book

price. ” International Journal of Cultural Policy 23(4): 464–481.

Reardon, S.F. (2011), “The widening academic achievement gap between the rich and the poor:

New evidence and possible explanations.” Whither Opportunity: 91–116.

Reardon, S.F. (2016), “School segregation and racial academic achievement gaps.” RSF

Reardon, S.F., Yun, J.T. and Eitle, T.M. (2000), “The changing structure of school segrega-
tion: Measurement and evidence of multiracial metropolitan-area school segregation, 1989-1995.”
Demography, 37(3): 351–364.

Reardon, S.F., Grewal, E.T., Kalogrides, D. and Greenberg, E. (2012), “Brown Fades: The
End of Court?Ordered School Desegregation and the Resegregation of American Public Schools.”
Journal of Policy Analysis and Management, 31(4): 876–904.

Reback, R. (2005), “House prices and the provision of local public services: capitalization under

school choice programs.” Journal of Urban Economics, 57(2): 275–301.

Ries, J. and Somerville, T. (2010), “School quality and residential property values: evidence from

Vancouver rezoning.” The Review of Economics and Statistics, 92(4): 928–944.

Rochet, J.C. and Tirole, J. (2003), “Platform competition in two-sided markets.” Journal of the

european economic association, 1(4): 990–1029.

Rochet, J.C. and Tirole, J. (2006), “Two-sided markets: a progress report.” The RAND journal

of economics, 37(3): 645–667.

151

Rosen, S. (1974), “Hedonic Prices and Implicit Markets: Product Diﬀerentiation in Pure Competi-

tion.” Journal of Political Economy, 82(1): 34–55.

Rysman, M. (2009), “The economics of two-sided markets.” Journal of Economic Perspectives,

82(1): 34–55.

Sieg, H., Smith, V.K., Banzhaf, H.S. and Walsh, R. (2004), “Estimating the general equi-
librium beneﬁts of large changes in spatially delineated public goods.” International Economic
Review, 45(4): 1047–1077.

Telser, L.G. (1960), “Why should manufacturers want fair trade?” The journal of law and eco-

nomics, 3: 86–105.

Tra, C. I. (2010), “A discrete choice equilibrium approach to valuing large environmental changes.”

Journal of Public Economics, 94(1-2): 183–196.

152