THREE PAPERS ON LABOR ECONOMICS

By

Andy Chou

A DISSERTATION

Michigan State University

in partial fulfillment of the requirements

Submitted to

for the degree of

Economics – Doctor of Philosophy

2019

ABSTRACT

THREE PAPERS ON LABOR ECONOMICS

By

Andy Chou

Chapter 1: Fortunes and Misfortunes of the Dragon Sons: Direct and Cohort Effects of
Superstition on Education Attainment
In many parts of East Asia, the fertility rate spikes every 12 years, starting in the 1970s. Researchers
have linked this phenomenon to the belief that being born in years associated with the dragon zodiac
leads to better outcomes in life; yet the research linking birth years and education outcomes has
found mixed results. One potential explanation for the mixed results is opposing mechanisms:
being born in dragon zodiac years, the dragon direct effect, may be positive while being in a larger
cohort during dragon years, the dragon cohort effect, may be negative. I use the difference between
cutoff for determining school cohort and zodiac cohort to estimate the separate effects from each
mechanism. Using the Taiwan Social Change Survey, I find evidence of a positive direct effect and
a negative cohort effect for those born during dragon zodiac years. I also look at the impact of tiger,
a zodiac that is considered to be unfortunate, and found the inverse.

Chapter 2: Crowding Out the Shadow: Effect of School Construction on Private Supplemen-
tary Education in Taiwan
To reduce after-school private tutoring, the Taiwanese government built more high schools, reason-
ing that these new schools would increase the number of seats and reduce the pressure of getting
into high school. However, it’s relatively unknown whether this policy accomplished its intended
purpose. Using the 1991-2006 Survey of Family Income and Expenditure, I study the effect of pub-
lic high school availability on household spending on private tutoring. Exploiting variation in high
school construction across counties throughout the 1990s, I find that increase in the probability of
getting into a public high school is associated reductions in households’ spending and participation
on private tutoring.

Chapter 3: Schooling the Superstitious: Evidence from Taiwan
Researchers find both an increase in the prevalence of superstitious acts over time and a positive
association between an individual’s educational attainment and participation in superstitious activ-
ities. Yet, these studies do not address selection into education groups. In this paper I examine
the causal impact of education, exploiting the extension of compulsory education from six to nine
years in Taiwan in 1968. I conduct my analysis using the Taiwan Social Change Survey from 1984
to 2015. IV estimates show that educational attainment decreases belief in superstition and the
prevalence of superstitious activities. In contrast, I find that OLS estimates are biased upwards.

ACKNOWLEDGEMENTS

I am sincerely grateful to Todd Elder, the chair of my dissertation, for providing mentorship,
support, encouragement, and guidance to me during this process.
I additionally thank Scott
Imberman, Leslie Papke, and Eric Chang, my committee members, for their E.P.I.C. comments,
feedback, and support. I am grateful to Lori Jean Nichols, other supporting staff in the economics
department, and MSU Writing Center, whose hard work, enthusiasm, and encouragement have
been invaluable. Finally, I thank my parents and my friends for emotional, editorial, and research
support. Without their support this dissertation would take much longer to finish.

iv

TABLE OF CONTENTS

.
.

.
.

.
.

.
.

.
.

.
.

.
.

Introduction .

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vii
ix
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

LIST OF TABLES .
LIST OF FIGURES .
CHAPTER 1 FORTUNES AND MISFORTUNES OF THE DRAGON SONS: DIRECT
AND COHORT EFFECTS OF SUPERSTITION ON EDUCATION AT-
TAINMENT .
.
.
1.1
.
.
1.2 Literature Review .
.
.
1.3 Zodiac Years and Cohort Sizes . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.4 Data .
.
1.5 Zodiac Years and Education Attainment
. . . . . . . . . . . . . . . . . . . . . . .
1.6 Regression Model .
1.7 Regression Results

1
1
4
6
8
9
.
. 11
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
1.7.1 Robustness Checks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
. 16

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

1.8 Discussion and Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

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

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

. .

. .

. .

. .

.
.
.

.
.
.

.
.
.

.

.
.

.

.
.

.

.
.

.

.
.

CHAPTER 2 CROWDING OUT THE SHADOW: EFFECT OF SCHOOL CONSTRUC-

.

.

.
.
.

.
.
.

.
.
.

. .

Introduction .

2.4 Methods .

2.1
.
2.2 Literature Review .
2.3 Background .
.

TION ON PRIVATE SUPPLEMENTARY EDUCATION IN TAIWAN . . . 18
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
.
.
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.3.1 Taiwanese Education System . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.3.2
Private Tutoring in Taiwan . . . . . . . . . . . . . . . . . . . . . . . . . . 24
2.3.3 Education Reform in Taiwan . . . . . . . . . . . . . . . . . . . . . . . . . 25
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
.
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
.
2.4.1 Data .
2.4.2
Sample Selection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
2.4.3 Regression Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
Summary Graphs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
2.5.1
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
2.5.2 Regression Results
2.5.3 Robustness Checks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
2.6 Discussion and Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32

2.5 Results .

.
.
. .

.
.
. .

.
.
. .

. .

. .

. .

.

.

.

.

. .

Introduction .

3.1
.
3.2 Literature Review .
3.3 Background .
.

CHAPTER 3 SCHOOLING THE SUPERSTITIOUS: EVIDENCE FROM TAIWAN . . . 34
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
.
.
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
3.3.1 Education Reform in 1968 . . . . . . . . . . . . . . . . . . . . . . . . . . 38
3.3.2 Notable Superstitions in Taiwan . . . . . . . . . . . . . . . . . . . . . . . 39

.
.
.

.
.
.

.
.
.

.

.

v

.

.
.

.
.

.
.

. .

. .

. .

3.6

. .
.
.

.
3.4 Data .
3.5 Event Study Results .

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
.
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
. 41
. . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.5.1 Regression Model .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. 42
3.5.2 Results
.
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
IV Results .
.
.
. 43
. . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.6.1 Regression Model .
.
. . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. 43
3.6.2 Results
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
. 44
. 45
3.8 Discussion and Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45

3.7.1 Number of Cohorts in Sample . . . . . . . . . . . . . . . . . . . . . . .
3.7.2 Variable Addition Tests . . . . . . . . . . . . . . . . . . . . . . . . . . .

.
3.7 Robustness Checks

. .
.
.

. .
.
.

. .
.

.

.

.

.

.

.

.

.

.

.

.

APPENDICES .

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
APPENDIX A FIGURES AND TABLES FOR CHAPTER 1 . . . . . . . . . . . . . . 48
APPENDIX B FIGURES AND TABLES FOR CHAPTER 2 . . . . . . . . . . . . . . 64
APPENDIX C REGARDING THE OFFICIAL NUMBER OF PRIVATE TUTOR-

ING CENTERS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
APPENDIX D FIGURES AND TABLES FOR CHAPTER 3 . . . . . . . . . . . . . . 86
. 102

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

BIBLIOGRAPHY .

.

.

.

.

.

.

.

.

vi

LIST OF TABLES

Table A.1: Chinese Zodiacs and the Corresponding Years . . . . . . . . . . . . . . . . . . . 51

Table A.2: Zodiac and Log Annual Live Births, 1947-2016 - Overall . . . . . . . . . . . . . 52

Table A.3: Zodiac and Log Annual Livebirths, 1947-2016 - Individual Years

. . . . . . . . 52

Table A.4: Zodiac and Log Annual Livebirths, 1947-2016 - Other Zodiacs . . . . . . . . . . 53

Table A.5: Zodiac and Log School Cohort Live Births, 1970-2016 - Overall . . . . . . . . . 53

Table A.6: Zodiac and Log School Cohort Live Births, 1970-2016 - Individual Years . . . . 54

Table A.7: Zodiac and Monthly Thousand Live Births, 1970-2016 - Bounding School

Cohort Substitution .

.

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55

Table A.8: Descriptive Statistics - TSCS . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56

Table A.9: Role of Controls

.

.

.

.

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57

Table A.10:Different Education Attainment

. . . . . . . . . . . . . . . . . . . . . . . . . . 57

Table A.11:Heterogeneity - Gender . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58

Table A.12:Heterogeneity - Different Dragon/Tiger years . . . . . . . . . . . . . . . . . . . 58

Table A.13:Heterogeneity - Relative Age within School Year

. . . . . . . . . . . . . . . . . 59

Table A.14:Heterogeneity - Father Ethnicity . . . . . . . . . . . . . . . . . . . . . . . . . . 60

Table A.15:Zodiacs before 1970 .

.

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60

Table A.16:Means of Control Variables by Dragon Zodiac Group . . . . . . . . . . . . . . . 61

Table A.17:Means of Control Variables by Tiger Zodiac Group . . . . . . . . . . . . . . . . 62

Table A.18:Robustness Check: Accounting for Short-Term Switch in Fertility . . . . . . . . 63

Table A.19:Selection on Unobservables: DD Model - LPM . . . . . . . . . . . . . . . . . . 63

Table B.1: Percentage of School Children Going to Private Tutoring Centers by Grade

and Subject of Study (2001)

. . . . . . . . . . . . . . . . . . . . . . . . . . . . 71

vii

Table B.2: Important Events of Taiwanese Education Reform . . . . . . . . . . . . . . . . . 71

Table B.3: Percentage change in public high school density from 1994-2001 by County . .

. 72

Table B.4: Role of Controls - Probit

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73

Table B.5: Role of Controls - Tobit . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74

Table B.6: IV First Stage .

.

.

.

.

.

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75

Table B.7: IV Reduced Form - Probit

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

. 76

Table B.8: IV Reduced Form - Tobit . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

. 77

Table B.9: Comparison of Methods - OLS/IV . . . . . . . . . . . . . . . . . . . . . . . .

. 78

Table B.10: School Districts - Participation . . . . . . . . . . . . . . . . . . . . . . . . . .

. 79

Table B.11: School Districts - Spending - Tobit . . . . . . . . . . . . . . . . . . . . . . . .

. 80

Table B.12: Placebo Test .

.

.

.

.

.

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81

Table D.1: Effect of Reform on Education . . . . . . . . . . . . . . . . . . . . . . . . . . . 93

Table D.2: Reduced Form - Overall

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

. 94

Table D.3: Reduced Form - Use by Question Type

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

. 95

Table D.4: IV results - Overall

.

.

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

. 96

Table D.5: IV results - Use by Question Type . . . . . . . . . . . . . . . . . . . . . . . .

. 97

Table D.6: Heterogeneous Effects by Religious Affiliation - Overall

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

. 98

Table D.7: Heterogeneous Effects by Religious Affiliation - by Question Type . . . . . . .

. 99

Table D.8: Dropping Controls - Overall

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

. 100

Table D.9: Dropping Controls - Use by Question Type . . . . . . . . . . . . . . . . . . .

. 101

viii

LIST OF FIGURES

Figure A.1: Dragon Effect Timeline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48

Figure A.2: Live Birth by Birth Year, 1947-2016 . . . . . . . . . . . . . . . . . . . . . .

. 48

Figure A.3: Trends in Education Attainment . . . . . . . . . . . . . . . . . . . . . . . . .

. 49

Figure A.4: Comparison of Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50

Figure A.5: Seven-year-old Children in First Grade (per 1,000 seven-year-old children) . . . 51

Figure B.1: Current Taiwanese Education System . . . . . . . . . . . . . . . . . . . . . .

. 64

Figure B.2: Number of High School per 10,000 High School Age Children, 1982-2010 . . . 65

Figure B.3: Correlation between County Average Spending on test preparation and High

School Construction Intensity . . . . . . . . . . . . . . . . . . . . . . . . . . . 65

Figure B.4: Number of First Year Public High School Students over Number of Middle

School Graduates by track, 1983-2010 . . . . . . . . . . . . . . . . . . . . . . 66

Figure B.5: Student Teacher Ratio by high school track, 1991-2010 . . . . . . . . . . . . . 66

Figure B.6: Public High School Sizes, SY 1994-1995 . . . . . . . . . . . . . . . . . . . .

. 67

Figure B.7: Treated Outcomes by School Construction Intensity group . . . . . . . . . . . . 67

Figure B.8: Control Outcomes by School Construction Intensity group . . . . . . . . . . .

. 68

Figure B.9: Effect of School Availability by Household Income Level

. . . . . . . . . . . . 69

Figure B.10: Effect of School Availability by Household Head Education . . . . . . . . . . . 70

Figure C.1: Number of Private Tutoring Centers by Different Sources . . . . . . . . . . .

. 84

Figure C.2: Number of Private Tutoring Centers - Adjusted . . . . . . . . . . . . . . . . .

. 85

Figure D.1: Trends Over Time .

.

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

. 86

Figure D.2: Use and View of Superstition by Education Group . . . . . . . . . . . . . . . . 87

Figure D.3: Use of Superstition by Question Type and Education Group . . . . . . . . . .

. 88

ix

Figure D.4: Outcomes by Birth School Cohort

. . . . . . . . . . . . . . . . . . . . . . . . 89

Figure D.5: Use of Superstition by Question Type and Birth School Cohort

. . . . . . . . . 90

Figure D.6: Robustness Check - Bandwidth (Overall) . . . . . . . . . . . . . . . . . . . . . 91

Figure D.7: Robustness Check - Bandwidth (by Question Type)

. . . . . . . . . . . . . . . 92

x

CHAPTER 1

FORTUNES AND MISFORTUNES OF THE DRAGON SONS: DIRECT AND COHORT

EFFECTS OF SUPERSTITION ON EDUCATION ATTAINMENT

1.1 Introduction

There is a common belief among people from East Asian countries that being born in certain
years, under specific zodiac signs in the lunar calendar, determines a person’s fortunes in life. Previ-
ous researchers have noted increases in fertility rate during years associated with the dragon zodiac,
a symbol of good fortune, as evidence of zodiac superstition (Goodkind, 1991; Yip et al., 2002).
However, the research on effect of zodiac superstition on education outcomes has produced mixed
findings. Previous researchers noted that those born in years associated with the “fortunate” zodiac
experience both positive effects of beliefs coming from parental expectations, self-confidence, or
expectations from others and the negative effects of increased cohort size. On the other hand,
people born in a zodiac considered to be “unfortunate” are faced with a negative effect from beliefs
and a positive effect from decreased cohort size. The opposing nature of the two mechanisms
results in an ambiguous overall effect from the zodiac superstition (Mocan and Yu, 2017; Wong
and Yung, 2005; Do and Phung, 2010; Johnson and Nye, 2011; Senbet and Huang, 2012; Agarwal
et al., 2017).1

For this paper, I focus on the region of Taiwan.2 Children in Taiwan are required to be age 6
1Nunn and Sanchez de la Sierra (2017) provides another example where the effect at the
individual level differs from the overall effect. Using the example of bulletproof spells in Africa,
they argue that even though in theory individual belief in bulletproof spells is harmful for individual
safety, group beliefs in bulletproof spells may be beneficial due to positive externality of individuals
making effort to ensure safety of the group.

2While the lunar calender and the zodiacs have origins from China, fertility spikes, particularly
those during dragon zodiac years, didn’t appear in China until the 2000s. Goodkind (1991)
hypothesized the policies of the Chinese government prohibiting traditional practices are at play.
In Taiwan, fertility policies were fairly relaxed. The Taiwanese government made efforts to reduce
fertility during the 1960s and 1970s through promoting the use of contraceptives and implementing
education programs on family planning rather than through penal actions such as fines or jail

1

on September 1st to enter elementary school.3 Those born between the months of September and
December are required to enter school a year later with those born between the months of January
and August of the next calendar year. This means that the school cohorts containing those born in
dragon zodiac years also have those born in other zodiac years. Thus, I estimate the effect of being
born in the dragon zodiac year by comparing the education attainment among those born in these
years and those who are not within the same academic cohort. The effect of being in a larger cohort
due to increased fertility during dragon zodiac years is estimated by comparing those in the same
academic cohort as those born in the dragon zodiac years and those who are not.4

I document changes in cohort size during years associated with different zodiac signs (dragon,
tiger, sheep, etc) using birth data at the national level from the Taiwan Ministry of Interior. Similar
to previous research (Goodkind, 1993), I find that birth cohort size is 7 percent larger on average
during years associated with the dragon(fortunate) zodiac, while birth cohort size is 8 percent
smaller on average during years associated with the tiger(unfortunate) zodiac. These fertility
effects do not vary by gender but do vary across different years. I find no evidence of changes in
birth cohort size during years not associated dragon or tiger. I present evidence that the changes in
annual birth cohort sizes are not driven only by parents switching the timing of child birth within
school cohort. The changes in annual birth cohort result in changes in academic cohort sizes.

I use the Taiwan Social Change Survey from 1996 to 2016 to analyze the effects of different
mechanisms. For the rest of the paper, I refer the effect of being born in a certain zodiac as direct
effect and the effect of being in a cohort with a different cohort size as cohort effect. Overall I
find evidence of a similar-sized direct effect, about 2 percentage points, on probability of having
a college education in the academic track for those born in dragon and tiger years. The size
of the cohort effect is stronger for those exposed to people born in the dragon years, at around 4
sentences (Sun, 1989).

3In Hong Kong, a school year goes from September to August of next calendar year. However,
students as young as 5 years and 8 months old can entry primary school. In Singapore and Malaysia,
school year and calendar year are the same. These children need to be at least 6 years old to enter
primary school.

4See Figure A.1 for a graphical representation.

2

percentage points, than in tiger years. The effects are weaker on less selective measures of academic
achievement, suggesting academic competition may play a role.

The importance of separately identifying the two effects is highlighted by differing effects across
different years. Looking across time, the cohort effect for dragons and tigers roughly follows the
relative magnitude of the fertility spike at the national level. However, there is little evidence of the
direct effect in the 1970s when the fertility effects are first observed.

I examine the heterogeneity in the direct and cohort effects through subsample analysis. While
there is no fertility effect by gender or relative age within school year, I find direct effects largely
concentrated on males and those who are older within a school cohort. I do not find support for
the direct effect driven by minority immigrants. I find evidence that the cohort effects are related
to the relative changes in fertility.

This paper contributes to the literature on the effect of culture on economic outcomes.5 There are
many theories trying to explain the persistence of cultural beliefs in cases with a single mechanism
(Bénabou and Tirole, 2016; Fudenberg and Levine, 2006; Foster and Kokko, 2009; Guiso et al.,
2016). The results from this paper show there may be spillover effect as large as, or even larger than,
the effects operating through individual beliefs. The indefinite sign of the overall effect suggests a
possible reason for the persistence of superstitions: multiple counteracting mechanisms may make
evaluating the superstitions difficult or misleading. In addition, even if the individuals are correctly
evaluating the outcomes, the superstition may be sustained by groups that benefit from group effects
within the academic cohort, namely, the dragons in academic cohorts with dragons or the non-tigers
in academic cohorts with tigers.

This paper also contributes to the literature on the effect of cohort size. Many articles document
a negative relationship of cohort size on education and labor market outcomes (Bound and Turner,
2007; Connelly and Gottschalk, 1995; Welch, 1979). However, articles finding a positive relation-
ship between cohort size and education attainment suggest other mechanisms, such as increased
public spending or scale economies are also at play (Do and Phung, 2010; Reiling, 2016). This

5See Guiso et al. (2006) for a review.

3

paper uses a source of variation that is exogenous from decisions of previous generations and over-
comes the problem of distinguishing between cohort and age effects. The evidence from people
born in years associated with the dragon zodiac supports the idea of “cohort crowding.” However,
I only find minor evidence of positive effect on education attainment when cohort size is smaller
for people born during years associated with the zodiac tiger. The difference between dragon and
tiger cohort effects may imply the negative effects of cohort size scales up at a different pace than
that of the positive effects.

The rest of the paper is organized as follows: Section 2 reviews the literature. Section 3 discusses
impact of zodiac years on cohort sizes. Section 4 describes the data. Section 5 discusses summary
statistics on education attainment by different zodiac group. Section 6 presents the regression
model. Section 7 presents estimation results. Section 8 discusses the relevant issues and concludes.

1.2 Literature Review

In East Asia, the Chinese lunar calendar (or an adaptation of it) is commonly used in conjunction
with the Gregorian calendar.6 In the Chinese lunar calendar, each year is represented by a creature.7
The collection of creatures, called zodiacs, follows through a twelve-year cycle. There is a common
belief that people born in certain years share the characteristics of the zodiac they were born under.
An often studied zodiac is the zodiac dragon.8 Dragon, the only zodiac with no real-life counterpart,
is a symbol for mystical power coming from the heavens. Several papers have found that during
dragon zodiac years, fertility rates have spiked consistently (Goodkind, 1991). These papers linked
6Several holidays in Taiwan, such as the Chinese New Year, Dragon Boat Festival, and Mid-
Autumn Festival, are based on dates in the lunar calendar rather than dates in the Gregorian
calendar.

7The twelve creatures, listed sequentially, in the Chinese zodiac are: Rat, Ox, Tiger, Rabbit,
Dragon, Snake, Horse, Sheep, Monkey, Chicken, Dog, Pig. Table A.1 lists the corresponding years
in the Gregorian calendar for each zodiac. In Vietnam, rabbit is replaced by cat. The Western
equivalent of zodiacs are the horoscopes. However, the western horoscopes vary by month and not
by year.

8Other known zodiac superstitions include firehorse women in Japan (Yamada, 2013), horse
zodiac in South Korea (Lee and Paik, 2006), and sheep zodiac in China (Mocan and Yu, 2017).
There is no evidence of superstition on the dragon zodiac for Japan and Korea.

4

this phenomenon to the belief that being born in years associated with the zodiac dragon can bring
a person good fortune and power in life.

Several previous studies have looked at the effect of being born in dragon zodiac years on
education and labor market outcomes, but the results have been mixed. Mocan and Yu (2017)
found positive evidence on education attainment and test scores using data in China. Liu (2015)
used Taiwanese data and found positive evidence on education attainment, but no effect on wages.
However, several studies found no evidence and sometimes even a negative impact of being born in
dragon years. Wong and Yung (2005) found no evidence on wages using the Hong Kong Census.
Sim (2015) used Singapore data and found a negative effect of being born in dragon zodiac years
on the probability of having a college degree. Agarwal et al. (2017) used a difference in differences
design to compare Chinese and non-Chinese and found a negative effect on income in Singapore.
One reason for the mixed findings is due to the multiple mechanisms triggered by the zodiac
superstition. For those born in dragon years, the positive effects from beliefs may be weakened by
the negative effects from larger cohort sizes. Previous studies tried to get around the effect of larger
cohort sizes by studying Asian immigrants in countries where Asians are minorities. Johnson and
Nye (2011) used US Census and Current Population Survey data and found being born in dragon
years is associated with increase in years of education attained. Senbet and Huang (2012) used
US Panel Study of Income Dynamics and found no dragon year effect on wages. However, there
are concerns about external validity due to differences between native and immigrant population.
Among the studies that look at countries where Chinese immigrants represent a significant portion,
only Do and Phung (2010) tried to account for cohort size effect by controlling for cohort size
in the regression. Yet cohort size may just be an indicator for local economic development. In
addition, increases in overall cohort size may not capture the sudden increase in cohort size during
“fortunate” zodiac years.

Most of the studies treat individuals within a zodiac year as homogeneous and did not explore
differences within zodiac group. Notable exceptions include Do and Phung (2010) and Agarwal
et al. (2017). Do and Phung (2010) used gender specific zodiac superstition in Vietnam to explore

5

the timing of the mechanism. Since gender cannot be observed before birth, they argued that the
zodiac superstitions were a result of pre-birth planning due to lack of differences between siblings
within households. However, most individuals in their sample, between ages 2 and 23, haven’t
completed their education.
It is unclear whether the null result in their study is due to sample
selection or zodiac superstition effects. Agarwal et al. (2017) looked at overall effects by gender
and time. They found the negative effect on only the younger cohort but did not find differing
effects by gender. Yet, the authors were unable to distinguish whether these differences or lack of
differences are determined by mechanisms in effect before entering the labor market or mechanisms
occurring during the labor market process.

In my paper, I look at Taiwan where there is a belief that dragon zodiac brings fortune while
the tiger zodiac brings misfortune.9 My paper differs from the literature in the following ways. I
estimate the effect of being born in a certain zodiac, named direct effect, and the effect of being in
a cohort with different cohort size, named cohort effect, separately. The separation is necessary as
it is believed that the direct and cohort effects are driven by different mechanisms. I further explore
the possible mechanisms by looking at different subsamples. In addition to variation by gender
and time, I also explored variations in relative age within school year and wave of immigration.
In terms of outcomes of interest, I look at education attainment with varying levels of selectivity.
This allows me to explore whether academic competition plays a role in realizing superstitions.

1.3 Zodiac Years and Cohort Sizes

Figure A.2 presents the total number of live births in each year between 1947 and 2016. There
was a sharp rise in birth number in the end of the 1940s, possibly due to the influx of Chinese
immigrants during the Chinese Civil War (Francis, 2011). The total number of births hovered
around 400,000 until the early 1980s. Since then, the number of births have been declining,
reaching a plateau in the last couple of years. There were spikes in cohort size during dragon
9Liu (2015) and Goodkind (1991) suggest the belief that tiger zodiac specifically brings mis-
fortune to females, as female born in tiger zodiac years are believed to make them stubborn and
unsuitable for wifely duties.

6

years since 1976. Before the 1970s, the cohort size stays flat (in 1952) or decreases (in 1964)
during dragon years. For the tiger years, birth numbers stay flat for 1962 and 1974 but decrease
consistently during tigers years since the 1980s. Goodkind (1991) hypothesize the lack of fertility
effect is due to changes in demographics, economic environment, or availability of modern birth
controls.

Table A.2 presents OLS results of log annual births on dummies for dragon and tiger years
at the national level. The regression results suggesting a 7.3 percent increase in number of live
births during dragon years and 7.9 percent decrease in number of live births during tiger years
relative to other years while accounting for a quadratic time trend overall. The interaction terms
between zodiac dummies and gender indicates that the changes during dragon or tiger years does
not vary by gender. Table A.3 present OLS results of log annual births on dummies for each
individual dragon or tiger years. The results suggest some variation in change in cohort size during
dragon years. Most notable is the statistically insignificant decrease during the dragon year of 1988.
This is possibly due to the response by the Taiwanese government to discourage birth in dragon
year(Goodkind, 1991). I do not find evidence of fertility spikes or drops in other zodiacs years, as
shown in Table A.4.

While there is evidence of increase in annual cohort size, the increase can be driven entirely by
changes of timing of fertility within academic cohort and result in no change in sizes of academic
cohorts. To test this, I acquired monthly birth data and calculated school cohort sizes using months
of birth. Table A.5 presents the overall result, while table A.6 presents results by each individual
year. The regression results suggest a statistically significant decrease during tiger years but a
statistically insignificant increase during dragon years. However, looking at individual dragon
cohorts suggest there are still increases in cohort sizes. The overall insignificant result may be
driven by the lack of increase for the 1988 dragon year.

An alternate way to look at the issue of annual birth increases being driven by fertility substitution
is to consider a particular scenario. Suppose the change in fertility is driven entirely by parents
changing their fertility timing within academic cohort, then there must be a same-magnitude but

7

opposite-signed shift in fertility in the birth months of individuals who would be in the same
academic cohort as people born in dragon or tiger years.10 I therefore regress the number of live
births on zodiac dummies and dummies for those who are not dragons or tigers but in the same
academic cohort as those born in dragon or tiger years. These results are presented in table A.7.
Suppose annual increases in fertility during dragon years are entirely driven by switching within
academic cohort, then the sum of the coefficient on dragon dummy and the coefficient on exposed
to dragon dummy should not be statistically different than zero. A F-test yields a F statistic of 5.20,
rejecting the null hypothesis at the 95 percent confidence level. Fort the tiger zodiac, the changes in
fertility for those non-tigers in the tiger cohort is the same sign as the change for those born during
tiger years. This suggests the annual fertility increase is not entirely driven by fertility switches
within academic cohort.

1.4 Data

I look at the impact of zodiac superstition on education attainment using the Taiwan Social
Change Survey (TSCS). TSCS is a biannual survey conducted by researchers in the Academia
Sinica. TSCS is nationally representative sample of adults, with sample size of around 2000 per
survey.11
I merged surveys collected between 1990 and 2016 that recorded the birth month of
the respondent. The resulting dataset includes 41 surveys spanning over 24 years. To look at
completed education, I only include those age 25 or above in the sample. Because the fertility
effects are only observed after 1970s, I focus my analysis to those born after 1970. TSCS consists
of basic characteristics such as gender and education attainment and additional themed questions
that rotates every five years. Table A.8 presents summary statistics of the variables I used in my
regression analysis.

Education level is coded into five categories: elementary school degree or below, middle school
10Fertility switches across academic cohorts would cause a real change in cohort sizes so they

are not a concern for this bounding argument.

11Surveys before 2000 sampled those age 20 to 65. Surveys in 2000 and 2001 removed the
restrictions on maximum age. Surveys starting in 2002 lowered the minimum age in the sample to
18 to conform with social surveys in other countries.

8

degree, high school degree, associates degree, and bachelor’s degree or above. Other than the survey
in 2003, all the other surveys distinguish between vocational and academic track in high school and
post-secondary education. Unless otherwise asked in the survey, education level includes those
who completed education and those dropped out before completing.

Zodiac dummies are constructed based on self-reported birth year and month. I approximate
the lunar calendar based on birth month. I coded each lunar year to start in February and end in
January of the next calendar year.

The control variables are mostly selected to be characteristics determined before birth. The
exception is parental occupation, which is during age 15 or 18 of the respondent. Parental ethnicity is
coded into three categories: Taiwanese (Fukien and Hakka Taiwanese), Mainlander, and Aborigine.
Birth place is defined as being born in the two municipal cities and the five provincial cities:
Taipei, Kaohsiung, Keelung, Hsinchu, Taichung, Chiayi, and Tainan.12 Religion is coded into five
categories: No religion, Folk religion, Buddhist, Daoist, and other religions(Christianity and Islam).
Occupations are coded into five levels according to Hwang (2003): managers and professionals;
technician and professional assistants; technical workers; machine operators and assemblers, sales
and service personnel; non-technical, labor, and agricultural workers.

1.5 Zodiac Years and Education Attainment

Overall Trend Figure A.3 presents percentage of individuals with different levels of education,
including high school, college or college in academic track, by birth lunar year. Overall the
individuals in later generations are more likely to have a high school, college, or college degree in
academic track. Percentage of individuals with a high school degree rose from about 60 percent
for those born in 1960 to over 90 percent for those born after 1970s. Percentage of individuals with
college degrees saw the most dramatic rise, with only 30 percent for those born in 1960 and 80
percent for those born in the 1980s. Percentage of individuals with college degrees in the academic
track also saw some significant rise, with 10 percent for those born in 1960 and about 40 percent for
12This designation applies between 1982 and 2009. Hsinchu city and Chiayi city became a

provincial city in 1982. Redesignation in 2010 changed the status of several cities and counties.

9

those born after late 1980s. There is evidence of spikes in dragon years and drops in tiger years but
only during 1986 tiger year and 1988 dragon year for those in the college academic track. There is
a drop in percentage of people having college degree for those born in 1988. There were no spikes
or drops in percentage with high school education, possibly due to the prevalence of people having
high school education.

By Treated Group Figure A.4 presents differences in percentage of persons having different
education attainment for different zodiac groups: dragons, non-dragons in dragon cohort, tigers,
non-tigers in tiger cohort, and others. Dragons and tigers refer to those born in the dragon and
tiger zodiac years. Non-dragons in dragon cohort and non-tigers in tiger cohort refers to those
born between September and January of the year before a dragon or tiger zodiac year and those
born between January and September of the year after a dragon or tiger zodiac year. Because of
the school year cutoff in Taiwan, these groups enter school along with those born in dragon or
tiger zodiac years. The “others” group refers to those not in the aforementioned groups. The three
figures on the left focus on outcomes during dragon years while the three figures on the right focus
on outcomes during tiger years. According to TCSC, 30.69 percent of people born in dragon years
have a college education in the academic track, slightly lower than 31.36 percent for people born
in other years (excluding tiger years). The higher education attainment of people born in dragon
years, compared to those who are supposed to be in the same school years, suggests the direct effect
is positive. In addition, the lower prevalence of people with college education in the academic
track, between non-dragons in dragon cohort and those who are not in the dragon cohort, suggests
the cohort effect is negative. Furthermore, people born in tiger years exhibit the opposite pattern as
people born in dragon years. However, those who are supposed to be in the same academic cohort
as people born in tiger years have a lower probability of having college education compared to those
not in the same school years as tigers. This suggests the cohort effect is negative or statistically
insignificant for people born in tiger years. Differences in means for having college education in
vocational track and high school education do not show particular differences among zodiac groups.

10

1.6 Regression Model

To account for the heterogeneity across different school years that are time invariant13, I estimate

direct and cohort effect using the following models:

Yi = β0 + β1zZodiaciz + SYi + β3Xi + 
i

Yi = β0 + β2zZodiacSYiz + ZYi + β3Xi + 
i

(1.1)

(1.2)

where Yi is education attainment for individual i. Zodiaciz is a vector of dummies for individual
i being born under zodiac z. ZodiacSYiz is a vector of dummies for being in the same school year
as people born in the zodiac years z. In my case, z = dragon or tiger. SYi, ZYi are birth academic
year/zodiac year fixed effects. Xi is a vector of control variables. These include parental education,
parental ethnicity, parental religion, place of birth, birth year, survey year, birth month fixed effects,
and birth order.

I estimated both equations using Probit and clustered my standard errors at the birth lunar year
level. The direct effects are estimated as β1z whereas the cohort effects are estimated as β2z. When
estimating the cohort effects, I dropped the individuals born in tiger or dragon zodiac years to avoid
confounding with the direct effects. To account for the missing data in the TSCS, I added a category
for missing variables and included a dummy that equals one if the control variable is missing.

1.7 Regression Results

Overall Results Table A.9 presents Probit marginal effects of dragon and tiger dummies on a
dummy for having college education in the academic track when all the controls are added. Column
1 only includes dragon and tiger dummies. Column 2 adds school year fixed effects and estimates
equation 1. Column 3 estimates equation 2, on dragon and tiger school cohort dummies and
13The papers in the dragon superstition effect literature address the issue by limiting the sample to
years close to the dragon zodiac years. Another reason to use fixed effects is to make my estimates
comparable to the estimates in the cohort size effect literature. While the articles in the cohort size
literature use data aggregated at the county or state level, they all include year fixed effects in their
estimation.

11

zodiac year fixed effects, while those born in dragon and tiger zodiac years are dropped. For the
dragon direct effect on having college education in the academic track, the results go from 0.4 to
2.0 percentage points when adding school year fixed effects. For the tiger direct effect on having
college education in academic track, the results stayed similar in size, going from -2.1 to -2.4
percentage points but with larger standard errors. The dragon cohort effect is much larger, at -4.6
percentage points, than the tiger cohort effect, at -1.8 percentage points. The difference between
the movement in coefficients of dragon and tiger direct effects after accounting for cohort effects is
consistent with the relative size of cohort effects.

On Different Education Attainment Table A.10 presents probit marginal effects of dragon and
tiger dummies on different levels of education attainment. Columns 1 and 2 are on college education
at the academic track. Columns 3 and 4 are on college education at any track. Columns 5 and 6 are
on high school education at any track. The direct effects weakens as the education level becomes
less competitive. The cohort effects are similar between college education at academic track or
college education in general but is much weaker for having high school education.

By Gender Previous research found male bias exhibited in within household resource distribution
in Taiwan (Parish and Willis, 1993). While I find no male bias in terms of fertility change during
dragon and tigers years, the male bias can still exhibit in education attainment. Table A.11 presents
probit marginal effects allowing for heterogeneous male and female effects on various outcomes.
While there is no difference in fertility spikes by gender, the results suggest males and females are
affected differently during dragon and tiger years. Dragon and tiger direct effects apply mostly to
males, with dragon direct effects estimated at 3.8 percentage points and tiger direct effect estimated
at -3.9 percentage points on having college education at academic track. There is no dragon or tiger
direct effect on females despite the groups having similar education attainment.14 A similar trend
14One explanation for the gender heterogeneity may be the difference in mandatory military
service. In Taiwan, males are required to serve in the military once they turn 18, while females do
not have service requirements. The service requirements can be delayed if the person is enrolled
in a domestic college or graduate degree program but not a foreign one. If a male student wishes

12

is observed for having high school education. There is also some evidence of resources shifting to
women during tigers years with a tiger direct effect of 3.3 percentage points for university degree in
vocational track and tiger direct effect of 1.5 percentage points for high school degree on women.
The effects on having college education at vocational track are mostly small. The dragon cohort
effects are similar in magnitude for males and females. However, the tiger cohort effect is mostly
on the females and in the opposite direction than previous literature would suggest.

Across Time Table A.12 presents Probit marginal effects from equations 1 and 2 when dummies
for each individual dragon or tiger years are included. The results suggest that the effects on fertility
is related to cohort effect but not the direct effect. Exploiting the differences in size of fertility
effects, I find little evidence of direct effect for the 1976 dragons and 1974 tigers, even though there
was an effect on fertility for both years. I also find strong evidence of direct effect for the 1988
dragons and 1986 tigers, even though there was no effect on fertility for the 1988 dragons. Overall
the relative size of the cohort effect goes the same direction as the size of the fertility effect. I
find negative and significant cohort effect for the 1976 dragons but no cohort effect for the 1988
dragons. I also find a negative cohort effect for the 1974 tigers and a positive cohort effect for the
1986 tigers.

I investigate variations created by the timing of the
By Relative Age within Academic Cohort
school year cutoff. Individuals born between February and August are placed in an earlier school
year while individuals born between September and January have to wait and enroll in the next
to study abroad, he has to do it before he turns 18 or after he finishes his service. According to the
Ministry of Education, the total number of student visas for all ages and programs range between
27,101 in 1998 to 37,171 in 2006. These numbers account for less than one percent of the students
in one student cohort. It is unlikely that the gender difference is driven by female students studying
abroad earlier. Another possibility is that the differences are driven by the timing of labor market
entry. The males born in 1976 face a two year military service, turn 18 in 1994, and should be
done with military service in 1996 if there’s no delay. The males born in 1988 face a one year and
four month service (less with certain academic credits), turn 18 in 2006, and should be done with
military service in 2007 if there’s no delay. 1994, 1996, 2006, and 2007 are all during economic
expansions. Thus, the results are unlikely to be driven by timing of labor market entry.

13

school year. Those born between September and January are thus “older” within their respective
academic cohort. Table A.13 presents Probit marginal effects allowing estimates to vary by relative
age within school year. I run the regressions only on the dragon zodiac or the tiger zodiac due to the
sample for estimating young tiger cohort effect and old dragon cohort effect overlaps when using
zodiac year fixed effects. I find that the positive dragon direct effect applies only to the dragons
relatively older within school cohort. Among those affected by the tiger superstition, the direct
effect is similar in magnitude among them. However, the estimates on the effects for the young
tigers are noisier. I find a strong negative cohort effect for the older dragons, while the cohort effect
for the young tigers is statistically indistinguishable from zero. I cannot separately estimate the
young dragon and old tiger cohorts effects and thus cannot infer whether cohort composition plays
a role in determining cohort effects.

Minority Status as a possible mechanism I explore whether the dragon and tiger effects are
driven by minority immigrant status15 by using the difference between different waves of Chinese
immigrants in Taiwan. If the dragon and tiger effects are driven by minority immigrant status, then
we should expect people with parents who are more recent Chinese immigrants to have stronger
effects. Table A.14 presents Probit marginal effects for equation 2 and 3 separately by the ethnic
group of the father. The Taiwanese group refers to parents who were born in Taiwan at the time
of the Chinese Civil War in early 1950s. The Chinese group refers to people who immigrated to
Taiwan during the Chinese Civil War. I find a weaker dragon direct effect for those with Chinese
parents.

Dragon and Tiger effects before 1970s While there is no observed fertility effect before the
1970s, it is possible the zodiac superstition exhibits itself in direct effects on individuals born
before the 1970s. Table A.15 presents Probit marginal effects on those born before 1970. Columns
1 and 2 present results for having a college education. Columns 3 and 4 present results for having
15See Goodkind (1995) for discussion of the minority immigrant status hypothesis applied in

explaining the variation of dragon fertility effects in Malaysia.

14

a high school education. Columns 5 and 6 present results for having a middle school education.16
I find no evidence of a positive dragon effect on any of the education outcomes, but there is some
evidence of tiger direct effect. One explanation for existence of tiger effect but no dragon effect, is
that it is more costly to increase education than to decrease it.17

1.7.1 Robustness Checks

In this section, I consider four different types of selection issues, selection on observables, selection
on unobservables, short-term switching, and delayed school entry, that could bias the estimates for
direct and cohort effects. The results suggest selection is not a big concern.

Selection on Observables Table A.16 and A.17 present summary statistics for observable char-
acteristics by different treatment groups. Table A.16 presents results for individuals affected by the
belief in dragon zodiac, while Table A.17 presents results for those affected by the belief in tiger
zodiac. Columns 2 to 4 are means for the different groups while columns 5 and 6 are differences in
means and significance star from t-tests. Overall the differences are mostly statistically insignificant
across groups except a few significant differences across the groups. For those affected by the belief
in dragon zodiac, they have a larger family size comparing within school years with dragons, more
likely to be older and less likely to have a Chinese Nationalist father comparing across school years
without dragons. For those affected by the belief in tiger zodiac, they are less likely to be male
within school year and more likely to be older across school years.

Short Term Switching Modern birth technology such as Caesarean section allows for parents to
choose the hour of the their children in a limited window. I account for the short-term switching
by estimating the DD model without individuals born in the beginning and end of the lunar year
16There were no separate track for college before 1997 since those in vocational track are not

expected to get a higher degree beyond high school for their jobs.

17Another known zodiac superstition with historical evidence is the firehorse superstition in

Japan. It is believed that women born in firehorse years brings misfortune.

15

(January and February in the Gregorian calendar). Table A.18 presents these results. The results
are similar in magnitude compared to the DD estimates.

Delayed School Entry One possible way parents can avoid the dragon cohort effect related to
increases in cohort size is to delay the entry of their children. If a substantial number of parents do
this compared to other years, then we should see a spike in seven-year-old children in first grade one
school year after they were supposed to enter. From Figure A.5, I do not find evidence of increased
delayed entry related to people born in dragon years.

Selection on Unobservables While there is not a lot of evidence that the various zodiac groups
are different in terms of observable characteristics, it is possible these groups are different based
on unobservable characteristics. For example, parents might know about the cohort size increases
in dragon years and avoid having children around that time. The parents who did are likely to be
positively selected. The cohort effect might be a result of the rest of the cohort being negative
selected, rather than a result of competition.
I follow the set up in Altonji et al. (2005) to test
whether these selection issues affect my results. Using the Stata commands from Oster (2017), the
results suggest selection on unobservables is not an issue. Selection on unobservables needs to go
a different direction than selection on observables to explain away the dragon direct effect, tiger
direct effect, and dragon cohort effect, while they need to be about 2.1 times as strong as selection
on observables to explain away the tiger cohort effect.18 The corresponding LPM regressions
estimates and delta estimates are presented in Table A.19.

1.8 Discussion and Conclusion

Informal institutions, such as culture, play a role in determining economic outcomes. These
institutions affect economic outcomes directly through changes in beliefs or through selection.
However, individuals may react to the superstition and create spillover or cohort effects. Evaluating
the impact of culture becomes harder when different mechanisms go in opposite directions. In

18The delta estimates have very large standard errors and need to be interpreted with caution.

16

this paper, I study the effects of zodiac superstition on education outcomes in Taiwan. I develop
a new method using institutional details to separately estimate the effects coming from different
mechanisms. I find evidence of both direct and cohort effects, even though estimates using methods
from other papers suggest no statistically significant effect overall.

I find parallels between the dragon(fortunate) zodiac and the tiger(unfortunate) zodiac but not
on all effects. The direct effects for dragons and tigers are very similar in magnitude overall and
in many of the subsamples for those born after 1970s. However, the dragon cohort effect is larger
in magnitude compared to the tiger cohort effect even though the fertility effects are similar in
magnitude. Historically, there is some evidence of a tiger direct effect but not a dragon direct effect.
I find direct effects to be strongest among males and those who should be older within their academic
cohort. I do not find evidence of the direct effect related to fertility or driven by ethnic minority
status. These differences may reflect variations in beliefs or economic environment, through which
the beliefs operate. Further research is needed to distinguish between the two possibilities.

The findings from this paper present challenges to both theoretical and empirical work on
It is possible that superstitions persist because the individuals are making wrong
superstition.
conclusions on the effect of superstition and fail to update their actions. Moreover, the existence
of spillover effects by superstition highlight the importance of having larger scale field studies in
addition to small scale randomized controlled trials in evaluating psychological effects at scale.
The evidence from subsample analysis questions the exogeneity assumption other papers applied
to use zodiac superstition as an instrumental variable.

The existence of zodiac effects also have implications for evaluating other policies. One
particular example is related to the interpretation of the effect of entrance exam change during the
year 2000 in Taiwan. The datasets used to evaluate the reform contain a group with people born
in dragon years and a group who were not. A direct comparison between the two groups is thus
aggregating both the effect of the reform and the dragon effects. Further studies should try to
account for the dragon effect or use different comparison groups.

17

CHAPTER 2

CROWDING OUT THE SHADOW: EFFECT OF SCHOOL CONSTRUCTION ON

PRIVATE SUPPLEMENTARY EDUCATION IN TAIWAN

2.1 Introduction

In recent years, researchers have noticed a rising trend in the prevalence of participating in
after-school tutoring programs (Dang and Rogers, 2008). These after-school tutoring programs,
often called shadow education, are most pronounced in East Asian countries. In countries such
as Taiwan, where there are high stakes entrance exams for secondary and tertiary education, many
students participate in private after-school tutoring (Yi and Wu, 2004). The long hours students
spend in private tutoring classes and the associated high fees raise concerns about students’ mental
health (Lin and Tsai, 2007) and educational inequality (Liu, 2006).

In the context of Taiwan, educational reformers have long argued that establishing more schools
would increase the likelihood of students getting into advanced levels of education and, thus,
decrease the use of private tutoring.1 This philosophy is exemplified by policy recommendations
in Taiwan from the early 1990s. For example, a report by the Executive Yuan Education Reform
Committee recommended that the government establish more high schools and colleges in addition
to implementing a multi-phased high school entrance program2 to reduce the stress of students
(Executive Yuan, 1996a). These recommendations were later implemented. Between 1991 and
2006, the number of high schools increased from 389 to 474, representing an increase of about
1.5 schools per 10,000 children aged 15 to 17. Hence, the percentage of middle school graduates
entering high school increased from 63.85 percent in 1991 to 79.24 percent in 2006. However,
1One of the reasons for the large-scale construction of middle schools that is part of the expansion
of compulsory schooling in 1968 is due to “the lack of capacity of middle schools, causing cramming
in elementary school students”(Ministry of Education, 1985).

2Most students in the past get into high schools/colleges through passing a joint entrance exam.
The multi-phased high school entrance program allowed students to get into high schools in ways
which standardized exam scores are weighted less.

18

the number of private tutoring centers drastically increased during the same period (Ministry of
Education, 2012).3 This led some scholars to conclude that other factors such as culture (Lin and
Huang, 2009) or school quality (Chou, 2003) were at play.

Even if the Taiwanese government could consistently build high quality schools or somehow
abate testing culture, it is not certain that an increase in school availability leads to a decrease in
private tutoring usage. The education literature is divided on whether private tutoring is a substitute
or a complement to formal education (Kwo and Bray, 2014). Furthermore, recent theoretical models
suggest that increasing school availability has an overall ambiguous effect on private tutoring for
entrance exam preparation (Chang, 2011; Chu, 2015).

Additionally, several alternative factors may explain the increase in the number of private
tutoring centers. Previous studies indicate private tutoring spending on a given child is positively
correlated with parental education and household income and is negatively correlated with total
number of children in a household (Sun and Hwang, 1996; Lin, 2001; Lin and Chen, 2006;
Liu, 2006; Lin and Huang, 2009). As the general population in Taiwan becomes more formally
educated, earns higher family income, and has fewer children, it is difficult to dismiss the effects of
demography on the increase in the number of private tutoring centers. In addition, shifts in the job
market may contribute to changes in private tutoring spending. The unemployment rate in Taiwan
has increased overall and relative to those with more education. As it becomes harder to find a
job with lower levels of education, going to private tutoring centers and getting higher levels of
education becomes more attractive.

This article examines the impact of high school construction in the late 1990s in Taiwan. I
use longitudinal household spending data from the Survey of Family Income and Expenditure
collected by the Taiwanese government from 1991 to 2006. Due to the comprehensive nature of
supplementary education in Taiwan, I look at household spending on private tutoring on academic
subjects and on technical skills.
I include county and year fixed effects, as well as control for
household demographics and changes in the macroeconomic environment. My main independent
3There are some errors with the official number of private tutoring centers. See Appendix B for

more details.

19

variable is a proxy for probability of entering high school, which is measured by the number of
first year high school students divided by the number of middle school graduates from the previous
school year. As a robustness check to make sure confounding effects from other concurrent
education policy changes are not biasing the baseline results, I also use high school density, which
is the number of high schools per 10,000 individuals aged 15-17, as an instrument for high school
probability.

I find that an increase in school availability reduced households’ spending on preparation for
entrance exams by the intensive and the extensive margin. Trends in average spending suggest
counties with higher construction intensity have slower growth. The trends in participation are
similar for counties with different construction intensity. Regression results suggest a 1 percent
increase in school availability leads to a decrease of NT 309 dollars for private tutoring class
spending and an increase of NT 313 dollars in skill class spending. Furthermore, I find that a 1
percentage point increase in school availability leads to a 0.2 percentage point decrease in private
tutoring class participation and 0.2 percentage point increase in skill class participation. While there
is concern the effect of school construction applies only to low income households or households
with less educated household head, I find that the effects apply across households with different
income and household head education.

This article contributes to the literature on school construction (Chou et al., 2010; Duflo, 2001;
Spohr, 2003). Many articles found a positive effect on education attainment and labor market
outcomes. The findings in this study suggests that changes in quality of education also indirectly
affect these estimates. The direction of the bias depends on whether private tutoring is beneficial
or harmful towards the student’s academic achievement.

This article also contributes to the literature that studies the relationship between public and
private inputs in education.4 Previous papers found responses from changes in school environment
in terms of parental involvement (Houtenville and Conway, 2008; Pop-Eleches and Urquiola, 2013).
My paper complements the literature by considering another aspect of response: use of private

4See Zhan (2014) and the articles cited within.

20

tutoring. Altogether, this evidence contributes to the evaluation of the impact of education inputs
on academics by considering secondary effects (Das et al., 2013).

The rest of the article is organized as follows. Section 2 reviews the literature. Section 3
discusses the background to supplementary education in Taiwan, including an overview of the
education system in the region and institutional details on its supplementary education. Section 4
describes the data and the empirical methodology. Section 5 presents estimation results. Section 6
discusses the relevant issues and concludes.

2.2 Literature Review

Scholars have pointed to various policy recommendations on regulating after-school tutoring
programs (Dang and Rogers, 2008), but there is little empirical evidence on the impact of these
regulations. One notable exception is research in South Korea. Researchers have looked at the
effect of operating hours restriction (Kim and Chang, 2010; Choi and Cho, 2016; Choi and Choi,
2016; Kim, 2016), High School Equalization Policy (Kim, 2004; Byun, 2010), and imposition of
after-school programs in formal schools (Bae et al., 2010).

Existing empirical studies on private tutoring in Taiwan have focused on two questions: Who
uses private tutoring, and is it effective? There is much more agreement regarding the answer to
the first question than the second. Researchers have found private tutoring usage to be positively
associated with parents’ education, father’s occupation, family income, region of residence, and
negatively associated with number of siblings. There is also some evidence of male children
receiving more private tutoring than female children. Additionally, usage of private tutoring does
not just increase linearly by household income (Sun and Hwang, 1996; Lin, 2001; Lin and Chen,
2006; Liu, 2006; Lin and Huang, 2009). On the other hand, the effect of private tutoring on
academic achievement is highly debated, with some articles arguing that private tutoring has a
positive effect (Sun and Hwang, 1996; Liu, 2006; Lee and Paik, 2006) and some others arguing
there is no effect or a negative effect (Lin and Chen, 2006; Kuan and Lee, 2010). There is also
some evidence that the decision of whether or not to use private tutoring is not exogenous(Yin

21

et al., 2012).

The academic discussion on changes in private tutoring usage over time in Taiwan has remained
theoretical. Chou (2003) pointed to the rising number of private tutoring centers and the high rate
of students using private tutoring in Taiwan and concluded the education reform was a failure.
Liu (2006) mentioned that one of the reasons for the rising number of private tutoring centers was
due to changes in demographic structure. Lin and coauthor hypothesized that as private tutoring
usage increases, there may be changes in culture or quality of private tutoring centers that further
influence the changes in number of private tutoring centers (Lin and Huang, 2009; Lin, 2012).

Economists have studied the interaction between public and private provision of goods and
services. Some notable examples are in health insurance (Brown and Finkelstein, 2008; Gruber and
Simon, 2008; Hamersma and Kim, 2013) and charitable giving (Andreoni and Payne, 2003, 2011;
Kingma, 1989). In education, researchers have looked at the interaction between charter school
enrollment and private school enrollment (Chakrabarti and Roy, 2016) and the interaction between
public and private provision of preschool education (Bassok et al., 2014; Bastos and Straume,
2016).

While private tutoring and privatized education share some similarities, the relationship between
demand of formal education supplied by different providers does not directly translate to the
relationship between demand of formal and informal education. The time occupied by formal
education does not overlap with the time occupied by informal education and many students use
informal education to complement their formal education. On the other hand, restrictions created
by limits on student stamina, curriculum of the formal schools, and cheating by teachers in formal
schools5can result in students substituting between formal and informal education. Yuan and Zhang
(2015) tested this relationship using Chinese data and found a negative relationship between local
government spending on public schools and household tutoring spending. However, it is unclear
whether the relationship identified is causal. My study uses a shift in school availability to address
the endogeneity concern.

5For empirical work on private tutoring corruption in Georgia, see Kobakhidze (2014). For

22

One mechanism driving the relationship between formal and informal education is through
outcomes realized in the future. Recent scholars have modeled the decision of private tutoring
usage and admission into higher education using overlapping generations models and found that
increase in admission slots may increase overall spending on private tutoring (Chang, 2011; Chu,
2015). This result is due to different effects of increasing enrollment on people in different parts of
the income-ability distribution. Imagine the case where all schools are of the same quality. Given
the limited number of enrollments, only people with a high enough income-ability combination will
go to the next level of education. When enrollment increases, the cutoff for the lowest income-ability
combination to get into the next level of education also changes. The enrollment changes create
an incentive for people who were just below the pre-expansion income-ability cutoff to participate
or invest more in private tutoring. On the other hand, those who are above the pre-expansion
income-ability cutoff have incentive to invest less in private tutoring or not participate in private
tutoring after the expansion. The effect on two different types of people move the overall spending
on private tutoring in different directions, thus making it unclear whether either one dominates the
other without formal modeling.

2.3 Background

2.3.1 Taiwanese Education System

In Taiwan, it is required by law for kids to go to school starting at the age of 6. The first nine
years of education (six years of elementary school and three years of middle school) are mandatory.
The curriculum of elementary and middle school varies little among schools. Beyond the middle
school level, schools are divided into two tracks: academic and vocational (see Figure B.1).
Students in academic programs focus on academic subjects such as Chinese, English, mathematics,
history, geography, chemistry, and physics; students in the vocational track focus on vocational
skills such as cooking, car repair, or hair styling. Academic programs are offered by high schools
and universities, while vocational programs are hosted by vocational high schools, junior colleges,
theoretical discussion, see Biswal (1999).

23

In general, students need to take an entrance exam to enter high schools
or technical colleges.
or colleges. The program and track students study during high school is determined with the
entrance exam scores. It is very rare for high school students to switch programs, tracks, or schools.
The process of getting into high schools or colleges, especially high-ranked institutions, is very
competitive. As a result, many students in Taiwan seek help outside of regular school hours by
hiring private tutors or enrolling in tutoring classes.

2.3.2 Private Tutoring in Taiwan

Private tutoring is regulated as part of supplementary education in Taiwan. According to Article
3 of the Supplementary Education Act, “Supplementary education can be divided into three types:
supplementary compulsory education, supplementary advanced education and short-term tutorial
education...Citizens seeking to acquire general knowledge and skills may receive short-term tutorial
education.” Short-term tutorial education is hosted by private tutoring centers. All registered
private tutoring centers require a qualified homeroom teacher and a fixed classroom location in
order to charge fees and advertise to the public. Other than medical sciences, short-term tutorial
education centers can offer any subject.6 Similar to the division in the formal education system,
the subjects are divided into two categories: Arts & Sciences and Vocational. However, students
cannot obtain degrees from short term tutorial education, compared to the formal and other two
types of supplementary education.

Different types of people attend different types of private tutoring. In 1998, 76.69 percent of
the individuals who attended private tutoring were students, while 14.11 percent were working.
Students were more likely to enroll in private tutoring offering academic, art or business subjects,
while those working were more likely to enroll in private tutoring in industrial, sports, or housework
subjects (Ministry of Education, 1999). Among the students, those in elementary schools tended to
6One notable exception is driving schools. Early reports include driving schools as part of
the vocational short-term tutorial classes. Starting in 1989, driving schools are regulated by the
Department of Transportation. They are no longer in the statistics on number of short-term tutorial
classes by the Department of Education (Ministry of Education, 1985).

24

go to private tutoring for skills such as foreign language, music and arts. Beyond the middle school
level, more students went to private tutoring related to school work. There was also some evidence
that students in different tracks in high school had different behavior in using private tutoring. In
general, students in academic high schools had a much higher rate of using private tutoring for
almost every subject (Ministry of Education, 2001).7

2.3.3 Education Reform in Taiwan

The education reform in the 1990s brought about a comprehensive set of changes across all levels
of education (Table B.2).8 There were a number of changes directly affecting the availability of
high schools. The first type of change was in total number of schools. According to the Ministry
of Education (Ministry of Education, 2010), there were 402 high schools in the 1994-1995 school
year; by the 2009-2010 school year, that number increased to 486. In terms of population between
the ages 15 and 17, in the 1994-1995 school year, the students represented 78.57 percent of the
population; by the 2009-2010 school year, the number increased to 92.45 percent (Figure B.2).

The intensity of high school construction also varies across different counties. Table B.3
summarizes the percentage change in public high school density across different counties between
1994 and 2001. The changes in density of high schools vary from almost no change in Tainan
City to almost doubling in Taipei County.
I further examined the relationship between school
construction intensity and average private tutoring spending in a county to see if there is any
evidence the government is building public schools at counties where private tutoring is more
prevalent. Interestingly, I didn’t find any relation between county high school construction intensity
and pre-reform average private tutoring spending (Figure B.3).9

Further, the Taiwanese government adjusted the distribution of schools across different tracks.
7The results of this report is reproduced in table B.1.
8Earlier policy changes in hopes to affect private tutoring include various changes of the

Supplementary Education Act and suspension of private tutoring licenses (Ting, 2005).

9I ran similar analysis on average county private tutoring spending in 1991 and 1992. I also find

no relation between county school construction intensity and average private tutoring spending.

25

In the 1994-1995 school year, the ratio of high school students in academic versus vocational track
was about 3:7, with 31.9 percent in academic high schools and 68.1 percent in vocational high
schools. The Ministry of Education slowly adjusted this ratio to 5:5. By the 2009-2010 school year,
53.2 percent of high school students were in academic high schools, whereas 46.8 percent were in
vocational high schools (Figure B.4).

Moreover, there was evidence of decreasing class sizes in high schools.10 Official statistics
showed a decrease in average number of students in a class for both the academic track and the
vocational track. In the 1991-1992 school year, the average number of students taught by a high
school teacher was around 22 for the academic track and around 21 for those in the vocational track.
By the 2010-2011 school year, the number decreased to around 19 for those in either the academic
track or the vocational track (see Figure B.5).

The education reform also brought about changes that would affect high schools’ quality. In
addition to academic and vocational tracks, the Ministry of Education also created a hybrid track
between the academic and the vocational tracks, officially known as comprehensive high schools.
Students in comprehensive high schools take common courses in their first year of high school and
delay their decision to choose between the academic and the vocational track in their second year
of high school. Vocational high schools can have both vocational and comprehensive programs in
the same school. However, the Ministry of Education often includes numbers of comprehensive
high schools as part of the academic track in official statistics.

Another change brought by the education reform was the way students get into high schools.
Before the education reform, academic high schools, vocational high schools, and junior colleges
held their entrance exams separately. Students had to decide where to apply to before knowing
their exam scores. After the education reform, the entrance exams were standardized on a national
level. The students can then use the score on the new exam to apply to each school respectively.
10Unlike for middle schools and elementary schools, there was no official target on class size
for high schools. However, the high school class size declines continuously while the birth rate
flattened out in the 1990s. This suggests forces other than natural births to affect high school class
size.

26

After reform, other aspects of student behavior, such as awards or extracurricular activities, were
given more consideration in the admission process compared to before.

2.4 Methods

2.4.1 Data

The Survey of Family Income and Expenditure (SFIE) is an annual survey administered by the
Taiwanese government. SFIE is nationally representative and covers detailed information on
various sources of household income and spending. Beginning in 1991, SFIE started to include
subcategories of education expenditures including (1) tuition and other related expenditures, such
as textbooks, stationery, supplementary readings; (2) spending on private tutors for the purpose of
entrance exam preparation; (3) spending on individual private tutors;111mm and (4) spending on
skill classes, such as piano, painting, computer, and dance lessons. In 2007, the survey design of
SFIE changed and new subcategories of education spending were used. I therefore limit my data
to 1991-2006. This includes time periods before and after the majority of school construction and
is useful to study the changes before and after the school construction.

The unit of analysis of SFIE is household. Spending data in SFIE is recorded as household
yearly totals. Therefore, I cannot directly observe how much each household spends on each
particular child if there is more than one child in the household.
I remedy this by controlling
for the number of students in each level of education and the percentage of male students in each
household. This limitation forces me to look at spending on private tutoring for students at all
levels of education. However, looking at spending on private tutoring for students at all levels of
education allows me to account for early planning behavior where parents send their kids to private
tutoring centers much earlier than the entrance exam.

I’ve also collected data from various branches of the Taiwanese government. The data on
number of schools, number of students, and number of graduates are obtained from the Ministry of
11In Taiwan, the courses offered in private tutoring centers are often small group classes. Very

few households (less than 1 percent in my sample) hire individual tutors.

27

Education. Data on population by age is obtained from the Ministry of the Interior. Unemployment
rate data is acquired from the Directorate General of Budget, Accounting and Statistics. In general,
the year in data from the Ministry of Education refers to school year, while data from SFIE and
other sources are in calendar year.

2.4.2 Sample Selection

SFIE is a random sample of all households in Taiwan. Considering the main purpose of private
tutoring is test preparation, I limit the sample of my analysis to households with students. I further
limit the relationship between students in the household and head of household to be of the own
children of the Head of Household to account for the students registered in another household for
the purpose of enrollment.12 I also ran regressions including students with any type of relationship
with the Head of Household. The regression coefficients are similar.

2.4.3 Regression Model

The empirical model is specified as follows:

Yi jt = α0 + α1Hjt + α2Xi jt + dj + dt + 
i jt

(2.1)

Where Yi jt is outcome of interest (log spending/participation on private tutors,skill classes, or
tuition) of household i in county j in year t.

Hjt denotes availability of public high schools in county j in year t, measured by number of first
year high school students divided by number of middle school graduates. I refer to this measure
as public high school probability as it is a proxy to the ex-post probability of entering public high
school. I choose this measure instead of high school density due to the differences in high school
sizes. As presented in figure B.6, while the majority of schools have between 1,500 and 2,000
12In Taiwan, physical households and registered households may not necessarily be the same. A
student can attend a school in an area different from the one they physically live in by registering
under another household, even if the student is not related to any of the members in the household
they registered in. Since SFIE asked about both registered but not cohabiting and not registered
but cohabiting members, I can get around the issue by limiting the relationship of the respondents.

28

students, schools can have over 3,000 students or under 1,000 students. Regressing on high school
density assumes a one-thousand student high school provides the same opportunity in attending
high school as a three-thousand student high school, whereas regressing on high school probability
treats all seats equal.

Xi jt denotes control variables. These include household characteristics such as age, education
level, and occupation of head of household, family type (single parent family, nuclear family,
grandparent family, extended family, others), number of male students in the household, number of
students in the household by each education level (elementary, junior high, senior high, college),
log household annual income, region of residence of household (urban, town, and rural), and county
unemployment rate.13dj, dt are county and year fixed effects.

I estimated the model using Tobit on private tutoring spending and Probit on private tutoring

participation. Standard errors are clustered at the county-year level.

To isolate the effects of school construction from other programs during the education reform,
I use school density, measured by number of schools per ten thousand people age 15-17, as an
instrument for high school probability.

2.5 Results

2.5.1 Summary Graphs

Figure B.7 shows the yearly trend in different categories of education spending by school con-
struction intensity between 1991 and 2006. The higher intensity group refers to counties with
an above median percentage change in public school density between 1994 and 2004. In 1994,
counties with more schools constructed spent 20% more in private tutoring classes and 33% more
in skill classes compared to those with less schools constructed. By 2004, the gap shrunk to 7% for
private tutoring class and -7% for skill classes. For private tutoring class participation, skill class
participation, and tuition spending, the two groups have similar trends or the gap between the two
13Due to the change in sample design of the Manpower Survey in 1993, the county unemployment

for 1991 and 1992 may not be directly comparable with the rest of the years.

29

groups have not changed significantly over time. On the other hand, with regards to households,
average tuition spending, income, head education, and student number, presented in figure B.8,
are mostly parallel during this time period. These results suggest there is a negative relationship
between school construction intensity and spending on private tutoring classes and skill classes,
while there is no significant relationship on participation in private tutoring classes and skill classes
and tuition spending.

2.5.2 Regression Results

Tables B.4 presents Probit marginal effects on private tutoring class participation and skill classes
participation. The results from the top to bottom panel are total private tutoring spending, private
tutoring spending on academic subjects, and private tutoring spending on non-academic skills. The
first column estimates the regression model controlling for only year fixed effects. The second
column adds county fixed effects. The third column adds controls for the head of household. The
fourth column adds controls for type of family and number of children in each gender and education
group. The fifth column adds controls for macroeconomic environment. The estimates are positive
without other control variables and negative when county fixed effects and other control variables
are added for private tutoring class participation. The estimates remain positive after adding
county fixed effects and other controls for skill class participation. The results imply an increase
in 1 percentage point of probability of getting into public high schools leads to a 0.2 percentage
point decrease in private tutoring class participation and 0.2 percentage point increase in skill class
participation. The size of the effect may not be economically significant.

Tables B.5, organized the same as table B.4, presents Tobit estimates on private tutoring class
spending and skill class spending. I estimate models with the same functional form, except the
independent variable is total amount of spending. The estimated coefficient goes from positive to
negative after adding county fixed effects and stayed negative with many types of controls added.
The coefficients imply an increase in 1 percentage point of probability of getting into public high
schools leads to a decrease of NT 309 dollars for private tutoring class spending and an increase

30

of NT 313 dollars in skill class spending. This represents a 1.3 percent decrease from average
spending in private tutoring and a 6 percent increase from average skill class spending.

The school availability measures captures all changes that affect the probability of entering high
school, including concurrent changes such as class sizes and tracks. I isolate the effect of school
construction by using the density of schools, measured by number of high schools per 10,000
individuals age 15-17 in a county, as an instrument for high school availability. The IV first stage
is presented in table B.6 while the reduced forms are presented in tables B.7 and B.8. Table B.9
presents OLS and IV results with all the control variables. The IV estimates are similar to the OLS
estimates but with larger standard errors.

Previous researchers reported increased competition for top high schools after the education
reform is implemented (Chou, 2003). If this is caused by the school construction, then there should
be a positive relationship between school construction and private tutoring spending or participation
among households with high income or highly educated head. Figure B.9 presents OLS and Probit
estimates when I allow the effect of school expansion to differ by household income level, while
figure B.10 presents estimates when the effect differs by education level of household head. I find
evidence of a negative effect of school construction at the highest income and household education
level. Both pieces of evidence do not support the aforementioned hypothesis.

2.5.3 Robustness Checks

One concern for county fixed effect is that several counties are located close to each other. The
regression results can be driven by students attending schools at a different county than the one they
live in. I estimate the effect of school construction using variation across school districts in 2016,
which merges some of the smaller districts, as the source of variation. The results are presented in
tables B.10 and B.11. The results are qualitatively similar to my main estimates.

It is possible that the regression results are driven by other unobservable changes variant over
year and county. These changes would also affect older adults who are taking private lessons. For
example, a 30-year-old taking computer lessons for work may be affected by changes in the labor

31

market but not changes in high school education policy. Table B.12 presents results on skill classes
spending on households with those, aged 25 to 50, who had completed their education and working.
I find no significant effect of school expansion on skill class spending and participation for this
group. This suggests the effect is not driven by macroeconomic forces unrelated to students.

2.6 Discussion and Conclusion

Private tutoring has long been prevalent in many countries, especially in East Asia, and the
issue has been long recognized by the public and the government in Taiwan. Historically, the usual
policy to reduce spending on private tutoring in Taiwan is to increase public school enrollment.
However, there is no empirical evaluation of whether this policy has worked. Using SFIE, I found
that within each county and year, the effect of school expansion on household spending on private
tutoring is negative but statistically insignificant in some models.

The purpose of this article is not to say other factors such as culture or quality of newly
constructed schools have no effect on household expenditure on private tutoring. In fact, it is likely
that there is some cultural shift during the education expansion. The purpose of this article is to
reexamine the claim that school expansion will decrease household spending on private tutoring
across the board. If we make that assumption, then we may overstate the influence of culture and
understate the potential effect of policy interventions on private tutoring.

Due to the limitation of data, I can only observe spending and participation in private tutoring
at the household level. Ultimately, policymakers are interested in how building more schools
is associated with individual spending and individual participation in private tutoring. I cannot
rule out the possibility of Simpson’s Paradox where household spending is on average weakly
negatively correlated with school availability but individual spending is positively correlated with
school availability. However, this result is still qualitatively different from conclusions drawn
looking only at number of registered private tutoring centers over time. Further research is required
to understand the effects of school expansion on individual spending and participation.

Another important aspect of the effect of school expansion is that the effect seems to be

32

different for households with different characteristics. Previous theoretical work suggests school
expansion will have different effects on households with different mixes of student ability and
household income. Due to the lack of information on student characteristics, I can only explore the
heterogeneous effect of school expansion on households with different incomes. I find evidence
that as household income increases, the effect of school expansion on household private tutoring
spending goes from negative to positive but statistically insignificant. This raises concerns that
school expansion could affect the inequality of education spending. Further research is needed to
disentangle the mechanisms causing the differing effect and the causal outcome of school expansion
on education spending inequality.

At the same time the expansion in high school happened, there was also a similar expansion for
college education. However, due to a lack of variation of college availability within any given year
and that college expansion, which happened between 1987 and 2007, started before the first year
of the SFIE data and ended after the last of the SFIE data, I cannot distinguish effects of college
expansion from time trends that may be caused by changes in culture or quality of private tutoring
centers. Further research is needed to identify the true effects of college expansion.

33

CHAPTER 3

SCHOOLING THE SUPERSTITIOUS: EVIDENCE FROM TAIWAN

3.1 Introduction

Many of us have our own routines or superstitions. Michael Jordan, one of the greatest basketball
players who ever played, is known for wearing his college shorts during games for good luck (Lee,
2013). While the majority of superstitions on a small scale are harmless, superstitions exercised
at a large scale may have harmful effects, sometimes at the expense of others. For example, Chou
(2019) found that in response to the superstition of the fortunate dragon zodiac, those born in
dragon years are more likely to attend college. Yet exposure to the larger academic cohort, due to
the fertility increase during dragon years, is associated with lower educational attainment.

Traditionally, psychologists consider use of superstitions to be the result of a lack of cognitive
ability. As cognitive ability improves with more education, it is hypothesized that uses or beliefs
in superstition will go down (Musch and Ehrenberg, 2002; Pennycook et al., 2012; Aarnio and
Lindeman, 2005). However, the empirical data doesn’t necessarily substantiate this claim.
In
Taiwan, the education level of the population has increased. According to the Ministry of Interior,
between 1985 and 1995, the number of college educated students increased from 1.6 million to 2
million.1 Chiu (1993, 1999) found that between 1985 and 1995 there was an increased prevalence
of people reporting the use of traditional rituals, such as Feng Shui and fortune telling. In addition,
Chiu found non-negative associations between educational attainment and reported use of Feng
Shui and fortune telling.2

There are multiple sources of bias that possibly explain the mismatch between the theory and
the empirical data. One source of bias is reverse causation. Previous research has found following
a superstition leads to better performance in experimental settings (Damisch et al., 2010; Siniver

1These numbers represent 9.39 percent and 10.24 percent of the population, respectively.
2Chiu’s results are replicated in figures D.1 to D.3.

34

and Yaniv, 2015). For example, the use of a “lucky pencil” may lead to more confidence, better
test scores, and higher educational attainment. The non-negative relationship between educational
attainment and usage of superstitious behaviors may simply capture the impact of following a
superstition on educational attainment, and not the other way around. Another source of bias is
omitted variable bias. Both educational attainment (Brodaty et al., 2014; Checchi et al., 2014) and
religious beliefs (Nielsen et al., 2017; Leon and Pfeifer, 2017) have been found to be correlated with
people’s risk attitudes. People who are more risk-averse may be more likely to follow superstitions
and also may be more likely to invest more years in education as opposed to working.3 The non-
negative relationship between educational attainment and usage of superstitious behaviors can be
driven by risk attitudes and not the impact of education on superstitious beliefs.

Taiwan is a prime location to study the relationship between education and uses and beliefs
of superstition because of the 1968 education reform, which extended compulsory education from
six to nine years. This created a natural experiment that resulted in sharp changes in educational
attainment. For my analysis, I use the Taiwan Social Change Survey (TSCS) between 1984 and
2015.
I focus my analysis on Feng Shui and fortune telling as previous research (Chiu, 1993,
1999) found non-negative associations between educational attainment and use of Feng Shui and
fortune telling. I examine both reported usage and reported superstitious beliefs. I also consider
the difference in types of usage, between active or current use versus previous use.

I first conduct an event study analysis to examine the impact of the reform on education,
superstitious beliefs, and superstitious behaviors.
I then use the event study specification to
construct an IV estimator where the first stage is the impact of the reform on years of education. I
compare IV estimates to OLS estimates using the method in Chiu (1993) to see the direction of the
bias.

3Brodaty et al. (2014) and Checchi et al. (2014) both found education to be negatively correlated
with risk-aversion using European data, meaning that those who are more risk-averse are less likely
to obtain college degrees. Theoretically, it is not certain whether spending more years in education
is more or less risky than spending more years working. Given the different education system
and labor market in Taiwan, it is entirely possible that risk-aversion and educational attainment are
positively correlated.

35

The event study specification suggests that being exposed to the reform is associated with an
increase of 0.4 years of education. My IV estimates suggest a negative relationship between years
of education on superstitions. The point estimates suggest one more year of education decreases
the probability of agreeing with Feng Shui and fortune telling by 0.4 and 1.2 percentage points and
decreases the probability of following Feng Shui and foretune telling by 0.7 and 0.3 percentage
points. However, the OLS results suggest a positive relationship between years of education and
use/belief in fortune telling, with one more year of education associated with a 1 percentage point
increase in agreeing with and using fortune telling. My IV estimates are more negative than the
OLS estimates, suggesting other factors, such as selection, may be biasing up the OLS estimates.
This paper contributes to the literature on external benefits of education.4 Previous literature
has investigated the effect of education on health, political participation, and crime. Lochner (2011)
suggested that one mechanism which can explain the effect of education on health outcomes is a
better understanding of health information. This paper tests a generalized version of the knowledge
mechanism and reports new evidence in support of it.5

This paper also contributes to the literature on the economics of religion.6 The majority of
the literature focuses on studying monotheistic western religions. Taiwan is a region where the
traditional religion is polytheistic and exhibits a diversity of beliefs. Extending the literature to
these regions complements and extends previous studies.

The rest of the paper is organized as follows: Section 2 reviews the literature. Section 3
discusses the background of the education reform and the superstitions. Section 4 describes the
data. Section 5 discusses the event study results. Section 6 discusses the IV and OLS results.
Section 7 discusses robustness checks. Section 8 provides additional discussion and conclusion.

4See McMahon (2004) and Lochner (2011) for review.
5There is a parallel literature on the impact of education on scientific knowledge (Drummond
In a sense superstitions are real life counterpart of type 1 errors while

and Fischhoff, 2017).
incorrect beliefs on scientific knowledge are examples of type 2 errors.

6See Iannaccone (1998) and Iyer (2016) for review.

36

3.2 Literature Review

Previous studies have explored the relationship between education and superstitious beliefs.
Many of them expected to find a negative relationship as they believed education is an indicator of
critical thinking (Musch and Ehrenberg, 2002; Pennycook et al., 2012; Aarnio and Lindeman, 2005).
However, the evidence is mixed. Some studies found negative relationships between education and
superstitious beliefs (Za’Rour, 1972; Orenstein, 2002; Otis and Alcock, 1982) while others found
insignificant or positive relationships (Roe, 1999; Wolfradt et al., 1999; Chiu, 1993, 1999).

One issue with such studies is that these conclusions ignore other factors that are both correlated
with education and superstitious beliefs. For example, risk attitudes are found to be correlated with
both educational attainment (Brodaty et al., 2014; Checchi et al., 2014) and religious beliefs (Nielsen
et al., 2017; Leon and Pfeifer, 2017). If more risk-averse individuals have higher education and are
more superstitious, then the positive relationship can be explained by failing to include measures
of risk aversion in the empirical model.

Additionally, education is thought to be a treatment to superstitious beliefs. Vyse (2014)
suggested teaching critical thinking and promoting science education as potential methods to
address superstitious thinking. Previous lab studies found lower superstitious beliefs after attending
critical thinking classes (Banziger, 1983; McLean and Miller, 2010; Manza et al., 2010; Wilson,
2018). However, studies that looked at different types of education found mixed results (Jahoda,
1968; Pasachoff et al., 1970; Salter and Routledge, 1971). Given that entrance into these different
types of education is not randomly selected, it is not clear whether these results are driven by
selection or whether education showed heterogeneous treatment effects.

Recent work by Mocan and Yu (2017) sought to address the selection issue apparent in field
studies by exploiting sharp changes in education as a result of changes in compulsory education
laws. They looked at the overall impact of a variety of reforms in Europe from the 1950s to 1980s.
Using the European Values Survey, they found a negative impact of education on the belief in the
power of lucky charms and the use of horoscopes in daily life.
This paper is in spirit similar to Mocan and Yu (2017).

I exploit changes in compulsory

37

education laws in Taiwan as a shock to educational attainment to identify the causal effect of
education on superstitions. My paper builds on Mocan and Yu (2017) in a number of ways. First,
my paper focuses on one region, whereas the effect in Mocan and Yu (2017) may be driven by
averaging of the countries’ different treatment effects. Second, my paper looks at Taiwan, where
there is a more diverse set of religions compared to Europe. This welcoming attitude toward
different religions may apply to superstitious beliefs, resulting in different effects. Additionally, in
terms of outcomes examined, my paper looks at both belief and usage. Given that usage can also
capture behaviors from the past, this study is able to capture a broader array of habits. Comparing
a current outcome such as belief and a mixed outcome such as usage can provide information on
whether these outcomes update over time or not.

3.3 Background

3.3.1 Education Reform in 1968

In 1968, the Taiwanese government extended tuition-free public education from 6 to 9 years.7 By
the 1960s, elementary education was almost universal. Yet only about 60 percent of elementary
school graduates chose to go to middle school (Chou et al., 2010). Part of this was due to the
admission process. Before this change, students who wished to continue their education after
elementary school had to take a competitive entrance exam to enroll in middle school. The
expansion abolished the middle school entrance exam and changed the middle school admission
process to be based on residence rather than test scores.

Another part of the 1968 education reform involved increasing the capacity of public education.
The Taiwanese government opened 150 new middle schools in 1968, a 50 percent increase from the
previous academic year. The Taiwanese government built more schools in counties that had fewer
schools before the reform (Chou et al., 2010). In addition, government expenditure rose by 15-20
percent for six years starting from 1968. Part of this increase was due to an rise in new teachers, as
the student-teacher ratio roughly kept pace with pre-reform levels (Spohr, 2003).

7Mandatory attendance in primary and secondary education wasn’t written into law until 1982.

38

The 1968 reform resulted in an increase in total years of education.8 The increase in education
is linked to increases in wages (Spohr, 2003; Clark and Hsieh, 2000) and increases in female
labor force participation (Tsai et al., 2009). Several studies also exploited the variation in school
construction across counties and found that increases in parental education impacted the education
(Tsai et al., 2011) and health (Chou et al., 2010) outcomes of their children.

3.3.2 Notable Superstitions in Taiwan

Feng Shui, literally translated as “wind and water,” is the description of one’s surroundings. In
addition, followers of Feng Shui believe that if the surroundings, which can be as small as one’s
room or as large as an entire country, are arranged following certain principles, the resulting
environment can bring about good outcomes or deter bad outcomes. For example, one Feng Shui
principle suggests avoiding aligning the front and back doors to prevent the good energy from
escaping.

Fortune telling refers to the practice of “learning” one’s life outcomes using information such
as timing of birth, facial structure, features on one’s palm, or tarot cards. A common practice is to
learn of future fortunes to take advantage of or to learn of the misfortunes to avoid.

Choosing dates refers to the practice of consulting a Chinese lunar calender to decide the date
of important occasions, such as wedding, funerals, opening a business, or moving. In the Chinese
lunar calender, each day is marked with a list of suitable or unsuitable activities. Believers make
important decisions based on these recommendations, sometimes along with recommendations
from a fortune teller.

Feng Shui, fortune telling, and choosing dates all rely on the belief that there is an underly-
ing force resulting in real-life outcomes. For fortune telling and choosing dates, the relation is
determined by the relationship in timing. For Feng Shui, the relation is determined by spatial
8Spohr (2003) reported an increase of 0.4 years of education for males and 0.25 years of
education for females while Tsai et al. (2009) reported an increase of 1.5 years for males and 2.1
years for females. My estimates are closer to those reported in Spohr (2003).

39

relationships. Individuals can consult experts on these subjects in order to get a prediction or a
prescription.

While Feng Shui, fortune telling, and choosing dates have origins dating back to ancient imperial
China, researchers have noted changes in the delivery or content of these superstitions based on
market forces. Li (1990) noted the increasing popularity of Feng Shui that reflected short-term
results as a response to market demand. Chiu (2006) noted the inclusion of western astrology in
traditional fortune telling services in response to the increasing popularity of western astrology.
The fortune telling industry has also adapted, creating numerous fortune telling websites to attract
customers, as the internet has become crucial for everyday purposes (Shuai et al., 2017).

3.4 Data

The Taiwan Social Change Survey (TSCS) is a nationally representative survey at the individual
level collected by researchers at Academia Sinica. The surveys were first collected in 1984
then continued annually starting from 1990. Each survey consists of two rounds, which were
administered at different times of the year. All surveys include basic demographic information
such as gender, age, marital status, education level9, religious affiliation, employment status, and
ethnicity10. Each survey round has a theme that repeats roughly every five years. I include surveys
that ask about religious beliefs and behaviors and one round that asks about communication
patterns. The wording of a question may be different in different years. Unless otherwise specified,
the wording of the question is the same as previously mentioned.

In the 1984, 1990, 1995, and 2000 survey, the religious behaviour questions were asked as
“Have you done the following?” with a list of ten occult, religious, and witchcraft behaviors. I
focus on fortune telling and Feng Shui because an early study by Chiu (1993) found non-negative
correlations between education level and responding, “yes,” to this question. In 2003, the question
9Education level was asked in years during the first survey in 1984, later surveys asked education

level in categories.

asked about ethnicity of each parent.

10Own ethnicity was asked before 1991, starting from the 1992 survey separate questions are

40

was worded as “Which of the following ways have you conducted fortune telling?” I recode this
question to a dummy variable, with 1 equalling “yes” to any of the subcategories of fortune telling.
In the rest of the surveys there are separate sections devoted to different types of religious behavior.
I use the questions “Have you used a fortune teller by your own will?” “In the past 5 years, have
you or your family examined Feng Shui?” “In recent years, have you consulted a fortune teller of
the following type?” For the last question I recode the variable as a dummy that equals one if the
respondent responded “yes” to any of the subcategories.

In a separate section, respondents are asked about their beliefs with the question “Do you believe
in the following” in the section head and several statements listed below. I only used the statements
“You need to choose dates for having weddings, having funerals, opening a business, and moving.”
and “You should consider Feng Shui before purchasing a house.” The options for the questions
range from 1 “Strongly Agree” 2 “Agree” 3 “Disagree” 4 “Strongly Disagree” 5 “No opinion.” I
recode the values to a dummy, with 1 equalling agree and strongly agree. I also include responses
to the question “Do you think fortune telling is accurate?” with responses ranging from 1 to 5 as
before in the 1994, 1999, and 2004 surveys.

3.5 Event Study Results

3.5.1 Regression Model

Following Spohr (2003), I estimate the impact of the reform using an event study specification.
The estimating equation is specified as follows

Yict = β0 + β1Tic + β2Xict + µc + ηt + vict

(3.1)

where Yict denotes the outcome of interest, years of education or measures of superstition of
individual i born in academic cohort c, and surveys in year t. Ti is a set of dummy that equals 1
if the individual is 12 or younger in September 1968. In the event study specification, Ti is a set
of dummies denoting individuals born up to 3 years before or after the first cohort exposed to the
education reform. Xi is a set of control variables, including a dummy for being born in the city,

41

a dummy for being married, religion fixed effects, a dummy for being employed, and household
monthly income in 2016 NT Dollars. µc denotes cohort time trends. ηt denotes survey fixed effects.
The standard errors are clustered at the birth school cohort level. The sample for the main
results includes those born 6 years before and after the first cohort exposed to the reform. To
account for the missing data in the TSCS, I add a category for missing variables and included a
dummy that equals one if the control variable is missing.

3.5.2 Results

Figures D.4 and D.5 present averages of educational attainment and various superstitious outcomes
by each academic cohort. I plotted separate quadratic trends before and after the first cohort exposed
to the reform to visualize the impact of the reform. There is a distinct jump in years of education
starting from the first cohort exposed to the reform. However, the impact of the reform on various
superstitious outcomes are mixed.

Table D.1 presents estimates on the impact of education reform. The first column estimates
the effect of the reform without controls, the second column controls for quadratic trend over birth
cohorts, gender, and a dummy for birth, and the third column uses an event study specification to
explore the dynamic effect of the reform. The results suggest the reform increases years of education
by 1.4 years. The effect lowers to 0.6 years once adding controls. Event study specification suggests
there are no pretrends of the reform effect on educational attainment.

It is possible that superstitious views and use change randomly across cohorts and the resulting
IV estimates are driven by spurious correlation. In order to address concerns of spurious correlation,
I estimate the reduced form regression of superstitious outcomes on treatment dummies using an
event study specification. Tables D.2 and D.3 present these results. The estimates show the effect
to be largely concentrated on the first and second cohort exposed to the reform. This suggests the
results are driven by the reform and not spurious correlation.

42

3.6 IV Results

3.6.1 Regression Model

To compare OLS and IV results, I estimate the following model

Sict = α0 + α1Eic + α2Xict + µc + ηt + 
ict

(3.2)

where Sict denotes a variable on superstitious belief or superstitious behavior. Eic denotes years
of education in the OLS estimation or predicted years of education using equation 1 as first stage
in IV estimation.

3.6.2 Results

Table D.4 presents OLS and IV regression coefficients of Years of Education on various superstitious
outcomes. The top panel presents results on superstitious views while the bottom panel presents
results on superstition use. All regressions control for quadratic time trend, pre-treatment controls,
and post-treatment controls. The first stage F statistic are in the bottom of the table. Columns 1 and
2 report the results when the outcome variable is about fortune telling, columns 3 and 4 report the
results when the outcome variable is about Feng Shui, columns 5 and 6 report the results when the
outcome variable is about choosing dates for special occasions. For all three outcomes, the OLS
is more positive than the IV estimates, suggesting an upward bias of the OLS estimates. The OLS
and IV results go in the same direction for Feng Shui and choosing dates, suggesting an increase
in education is associated with a decrease in superstition belief. For fortune telling, the estimate
goes from positive to statistically insignificant. The IV estimate for fortune telling suggests there
are similar levels of belief in fortune telling across years of education. For the bottom panel, the
regression coefficient in column 1 implies a one-year increase in education is associated with a one
percentage point increase in the probability of using fortune telling. Similar to the results on views
of superstitions, the OLS regression coefficients for uses of superstition are more positive than IV

43

estimates. Both IV estimates for fortune telling and Feng Shui are negative, suggesting increases
in education reduces engagement of superstitious acts.

To explore possible mechanisms driving the relationship between education and superstition
usage, I run the same regression on different types of questions asked. Table D.5 presents the
results. The top panel presents results on fortune telling while the bottom panel presents the results
on Feng Shui. Columns 1 and 2 present the results when the question doesn’t specify when the
timing of the usage was. Columns 3 and 4 present the results when the question includes the
wording of “in the past year” or “in the past 5 years” Columns 5 and 6 includes questions with
the additional wording of voluntary/active. These results control for time, pre-treatment variables,
and post-treatment variables. Overall there is a positive relationship between education and use
of fortune telling in any time or in recent times using OLS, yet all the estimates become more
negative once using IV. This suggests a positive bias for OLS. The relationship between education
and fortune telling seems to be driven by individuals’ voluntarily use while Feng Shui use is driven
by overall use but not recent use.

Tables D.6 and D.7 present IV estimation results across different religious groups. Column 1
presents results on those with no religion. Column 2 presents results on those affiliated with folk
religion. Column 3 presents results on Buddhist. Column 4 presents results on Daoist. Column 5
presents results on other religions, including Christianity and Islam. I find the effect of education
to be strongest on those with no religious affiliation.

3.7 Robustness Checks

3.7.1 Number of Cohorts in Sample

To test the robustness of the estimates, I re-estimated equations 1 and 2 by changing the number
of cohorts included in the sample. I presented these estimates in Figures D.6 and D.7. Overall,
the estimates are qualitatively similar. The IV estimates are consistently more negative than OLS
estimates. These results suggest OLS estimates are biased upward.

44

3.7.2 Variable Addition Tests

To test whether selection plays a role in driving the results, I re-estimate the results while dropping
a control variable one at a time to see how the estimates change. The results are presented in
Tables D.8 and D.9. The estimates are qualitatively similar when dropping most single variables.
The biggest difference seems to be from dropping gender.

3.8 Discussion and Conclusion

In this paper, I examined the impact of education on superstitious beliefs and behaviors in
I exploited the education reform in 1968 to construct an IV estimator to examine the
Taiwan.
causal effect of education. I found that IV estimates are either negative or statistically insignificant
despite some OLS estimates being positive. These results suggest education reduces superstitious
beliefs and actions. In addition, selection or other unobserved factors may result in correlations in
a different direction.

My estimates support the hypothesis that education decreases superstition through an increase
in knowledge or critical thinking. These estimates, while significant, are small compared to the
prevalence of superstitious beliefs and actions. While it is possible social benefits or costs are taken
into consideration when forming superstitious beliefs or engaging in superstitious acts, I cannot
separate out the effects of these mechanisms due to data limitations. I leave this empirical challenge
to a future study.

One caveat to note concerns policy implications. The results point to the overall impact of the
education reform. However, due to data limitations, I cannot separately identify the exact impact
of specific pieces of the education reform. The exact impact of education expansion, increasing
education funding, and expanding compulsory education individually requires further study.

Another caveat to note regards the external validity of the IV method. IV identifies local average
treatment effects, which means these estimates apply to compliers of middle school education in
the late 1960s/early 1970s in Taiwan. Caution needs to be taken when extending these estimates
to non-compliers, other time periods, or other levels of education. The question of whether these

45

results apply to higher levels of education requires further study.

46

APPENDICES

47

APPENDIX A

FIGURES AND TABLES FOR CHAPTER 1

Figure A.1: Dragon Effect Timeline

Note: Solid lines are the calendar year cutoffs. Dashed lines are the school year cutoffs. This is for
children born in 2000 entering elementary school.

Figure A.2: Live Birth by Birth Year, 1947-2016

Note: Green lines are years associated with the zodiac dragon. Orange lines are years
associated with the zodiac tiger. Source: Taiwan Ministry of Interior.

48

Figure A.3: Trends in Education Attainment

Note: Green lines are years associated with the zodiac dragon. Orange lines are years
associated with the zodiac tiger.

49

Figure A.4: Comparison of Methods

(a) College (Academic) - Dragon

(b) College (Academic) - Tiger

Note: Tiger Zodiac is not included in the analysis.

Note: Dragon Zodiac is not included in the analysis.

(c) College (Any) - Dragon

(d) College (Any) - Tiger

Note: Tiger Zodiac is not included in the analysis.

Note: Dragon Zodiac is not included in the analysis.

(e) High School - Dragon

(f) High School - Tiger

Note: Tiger Zodiac is not included in the analysis.

Note: Dragon Zodiac is not included in the analysis.

50

Figure A.5: Seven-year-old Children in First Grade (per 1,000 seven-year-old children)

Note: Years after 2000 are not included due to increased enforcement of school entry laws following the amendment
of the school entry laws in 1999.

Table A.1: Chinese Zodiacs and the Corresponding Years

Zodiac

Rat
Ox
Tiger
Rabbit
Dragon
Snake
Horse
Sheep
Monkey
Chicken

Dog
Pig

Years in Gregorian Calendar

1936, 1948, 1960, 1972, 1984, 1996, 2008
1937, 1949, 1961, 1973, 1985, 1997, 2009
1938, 1950, 1962, 1974, 1986, 1998, 2010
1939, 1951, 1963, 1975, 1987, 1999, 2011
1940, 1952, 1964, 1976, 1988, 2000, 2012
1941, 1953, 1965, 1977, 1989, 2001, 2013
1942, 1954, 1966, 1978, 1990, 2002, 2014
1943, 1955, 1967, 1979, 1991, 2003, 2015
1944, 1956, 1968, 1980, 1992, 2004, 2016
1945, 1957, 1969, 1981, 1993, 2005, 2017
1946, 1958, 1970, 1982, 1994, 2006, 2018
1947, 1959, 1971, 1983, 1995, 2007, 2019

51

Table A.2: Zodiac and Log Annual Live Births, 1947-2016 - Overall

Male

Dragon

Tiger

Dragon X Male

Tiger X Male

(1)

(2)

Overall Gender Specific
0.071∗∗∗
0.071∗∗∗
(0.019)
(0.017)
0.073∗∗∗
0.073∗∗∗
(0.028)
(0.020)
−0.079∗
−0.080∗∗∗
(0.029)
(0.042)
−0.001
(0.039)
−0.001
(0.057)
140
Yes

Observations
Quadratic Trend
Note: Standard errors in parentheses.

140
Yes

* p < 0.10, ** p < 0.05, *** p < 0.01.

Table A.3: Zodiac and Log Annual Livebirths, 1947-2016 - Individual Years

1974/1976

1986/1988

1998/2000

2010/2012

(1)

(2)

Tiger Years Dragon Years
0.077∗∗∗
−0.079∗∗∗
(0.010)
(0.011)
−0.157∗∗∗
−0.014
(0.011)
(0.012)
0.115∗∗∗
−0.072∗∗∗
(0.016)
(0.013)
−0.232∗∗∗
0.186∗∗∗
(0.032)
(0.026)
140
140
Yes
Yes

Observations
Quadratic Trend
Note: Standard errors in parentheses.

* p < 0.10, ** p < 0.05, *** p < 0.01.

52

Table A.4: Zodiac and Log Annual Livebirths, 1947-2016 - Other Zodiacs

(5)
Dog

(6)
Pig

(7)
Rat

Male

Zodiac Dummy

Zodiac Dummy X Male

(2)
Sheep Monkey Chicken

(3)

(4)

(1)
Horse
0.071∗∗∗ 0.071∗∗∗ 0.071∗∗∗ 0.071∗∗∗ 0.071∗∗∗ 0.072∗∗∗ 0.071∗∗∗
(0.018)
(0.019)
(0.018)
0.007 −0.046 −0.044
0.043
(0.041)
(0.032)
(0.047)
0.002 −0.003 −0.001
0.003
(0.042)
(0.055)
(0.068)
140
140
140
Yes
Yes
Yes

(0.018)
−0.004
(0.042)
−0.001
(0.058)
140
Yes

(0.018)
0.043
(0.050)
0.004
(0.068)
140
Yes

(0.018)
0.030
(0.049)
0.008
(0.067)
140
Yes

(0.017)

(0.057)

(0.082)
140
Yes

Observations
Quadratic Trend
Note: Standard errors in parentheses. * p < 0.10, ** p < 0.05, *** p < 0.01.

Table A.5: Zodiac and Log School Cohort Live Births, 1970-2016 - Overall

Young Dragon School Year

Old Dragon School Year

Young Tiger School Year

Old Tiger School Year

(1)

Overall
0.016
(0.035)
0.052
(0.033)
−0.082∗
(0.047)
−0.081∗∗
(0.030)

Observations
Quadratic Trend
Note: Standard errors in parentheses.

46
Yes

* p < 0.10, ** p < 0.05, *** p < 0.01.

53

Table A.6: Zodiac and Log School Cohort Live Births, 1970-2016 - Individual Years

1973/1975

1974/1976

1985/1987

1986/1988

1997/1999

1998/2000

2009/2011

2010/2012

(1)

(2)

Tiger Years Dragon Years
−0.112∗∗∗
(0.035)
−0.094∗∗∗
(0.032)
−0.099∗∗∗
(0.017)
−0.106∗∗∗
(0.018)
0.043∗
(0.021)
−0.011
(0.022)
−0.189∗∗∗
(0.036)
−0.143∗∗∗
(0.039)

0.006
(0.024)
0.080∗∗∗
(0.022)
−0.052∗∗
(0.023)
−0.020
(0.023)
0.072∗∗∗
(0.020)
0.079∗∗∗
(0.021)
0.108∗
(0.058)
0.144∗∗
(0.064)

Observations
Quadratic Trend
Note: Standard errors in parentheses.

46
Yes

* p < 0.10, ** p < 0.05, *** p < 0.01.

46
Yes

54

Dragon Zodiac

Tiger Zodiac

Exposed to Dragon

Exposed to Tiger

Overall Substitution Bound
2.438∗∗∗
(0.333)
−2.795∗∗∗
(0.305)

2.211∗∗∗
(0.342)
−3.022∗∗∗
(0.314)
−1.085∗∗∗
(0.292)
−1.126∗∗∗
(0.339)
565
Yes

Table A.7: Zodiac and Monthly Thousand Live Births, 1970-2016 - Bounding School Cohort
Substitution

(1)

(2)

Observations
Quadratic Trend
Note: Standard errors in parentheses.

565
Yes

* p < 0.10, ** p < 0.05, *** p < 0.01.

55

Table A.8: Descriptive Statistics - TSCS

count

mean

sd

min max

0
0
0

0
0
0
0

1996
1970
1971

.095386
.1806313
.0939283
.1895044

.2937566
.3847246
.2917381
.3919212

.3076127
.6141463
.9210293

.4615202
.4868117
.2697016

15461
15778
15778

15778
15778
15778
15778

Dependent Variables
College and above(Academic)
College and above
High School and above
Independent Variable
Dragon Direct Effect
Dragon Cohort Effect
Tiger Direct Effect
Tiger Cohort Effect
Control Variables
Survey Year
Birth School Cohort
Birth Lunar Year
Birth Month
1
Male
0
Born in City
0
Father Bachelor’s Education
0
Father Associate’s Education
0
Mother Bachelor’s Education
0
Mother Associate’s Education
0
Father Chinese Nationalist
0
Mother Chinese Nationalist
0
Father Folk Religion
0
Mother Folk Religion
0
Father Occupation Rank
0
Mother Occupation Rank
0
Sibling Rank(1=Oldest)
1
Family Size
1
Oldest Sibling
0
Note: The 2003 survey did not distinguish the education track of the individuals.

4.73301
4.962798
4.953068
3.442201
.4997053
.4856143
.2543349
.275402
.1762288
.1945923
.2914674
.1989696
.4962795
.4906294
1.263339
1.237803
1.331591
1.290055
.4835189

2009.751
1977.369
1977.929
6.715807
.5176195
.3807916
.069513
.0826752
.0320833
.0394167
.0937337
.0412912
.4378549
.4023355
2.448011
1.713018
2.201145
3.301837
.3724648

15778
15778
15778
15778
15778
7403
12156
12156
12000
12000
15384
14870
1585
942
4174
1690
4191
4191
4191

1
1
1

1
1
1
1

2016
1991
1991
12
1
1
1
1
1
1
1
1
1
1
5
5
11
11
1

56

Table A.9: Role of Controls

Dependent Variable: College Education(Academic)

(3)

Dragon Direct Effect

Tiger Direct Effect

(1)
0.004
(0.007)
−0.021∗∗∗
(0.006)

(2)
0.020
(0.015)
−0.024
(0.015)

Tiger Cohort Effect

Dragon Cohort Effect

−0.046∗
(0.025)
−0.018
(0.018)
12547
Observations
.311
Ymean
Year Fixed Effects
Zodiac
* p < 0.1, ** p < 0.05, *** p < 0.01. Standard errors clustered at birth

15461
.308
No

15461
.308

Academic

year level.

Note: All regressions include survey year, birth month fixed effects, and

birth year in controls.

Table A.10: Different Education Attainment

College
Academic

College

Any

High School

Any

0.015
(0.010)
−0.012
(0.009)

0.007
(0.005)
−0.001
(0.006)

0.020
(0.015)
−0.024
(0.015)

Dragon Direct Effect

Tiger Direct Effect

Dragon Cohort Effect

Tiger Cohort Effect

−0.010
(0.018)
−0.013
(0.012)
12791
.922
Academic Zodiac Academic Zodiac Academic Zodiac

Observations
Ymean
Year Fixed Effects
* p < 0.1, ** p < 0.05, *** p < 0.01. Standard errors clustered at birth year level.
Note: All regressions include survey year, birth month fixed effects, and birth year in controls.

−0.041∗
(0.022)
−0.024∗
(0.013)
12791
.615

−0.046∗
(0.025)
−0.018
(0.018)
12547
.311

15461
.308

15778
.614

15749
.921

57

Dragon Direct Effect

Tiger Direct Effect

Observations
Ymean
Year Fixed Effects

Dragon Cohort Effect

Tiger Cohort Effect

Table A.11: Heterogeneity - Gender

College
Academic

Male

Female
−0.000
0.038
(0.007)
(0.027)
−0.039∗∗∗ −0.006
(0.024)
(0.010)
7470
7991
.306
.31

College

Any

Female

Male

0.023
(0.019)
−0.046∗∗∗
(0.003)
8167
.601

0.008
(0.012)
0.025
(0.020)
7611
.628

High School

Any

Male
Female
0.018∗∗∗ −0.004
(0.007)
(0.005)
−0.016∗∗
0.015∗∗
(0.006)
(0.007)
7598
7974
.914
.927

Academic Academic Academic Academic Academic Academic
−0.055
(0.045)
−0.001
(0.029)
6519
.308
Zodiac

−0.047∗∗∗ −0.020
(0.012)
(0.023)
−0.071∗∗∗ −0.002
(0.015)
(0.014)
6622
6131
.916
.627
Zodiac
Zodiac

−0.040∗∗
(0.017)
−0.040
(0.032)
6028
.314
Zodiac

0.003
(0.013)
−0.019∗
(0.011)
6131
.927
Zodiac

−0.038
(0.044)
0.019
(0.028)
6660
.605
Zodiac

Observations
Ymean
Year Fixed Effects
* p < 0.1, ** p < 0.05, *** p < 0.01. Standard errors clustered at birth year level.
Note: All regressions include survey year, birth month fixed effects, and birth year in controls.

Table A.12: Heterogeneity - Different Dragon/Tiger years

1976 Dragon Direct Effect

1988 Dragon Direct Effect

1974 Tiger Direct Effect

1986 Tiger Direct Effect

1976 Dragon Cohort Effect

1988 Dragon Cohort Effect

1974 Tiger Cohort Effect

1986 Tiger Cohort Effect

College
Academic

0.011
(0.017)
0.053∗∗∗
(0.013)
−0.002
(0.006)
−0.076∗∗∗
(0.013)

College

Any

High School

Any

0.022∗
(0.011)
−0.020∗∗
(0.009)
−0.002
(0.006)
−0.049∗∗∗
(0.013)

0.009
(0.006)
−0.013
(0.013)
−0.001
(0.007)
−0.002
(0.011)

−0.016
(0.016)
0.025
(0.034)
−0.022
(0.014)
0.046∗∗
(0.019)
12791
.922
Academic Zodiac Academic Zodiac Academic Zodiac

−0.062∗∗∗
(0.019)
0.003
(0.009)
−0.043∗∗
(0.019)
0.042∗∗∗
(0.006)
12547
.311

−0.048∗∗
(0.019)
−0.015
(0.055)
−0.038∗∗
(0.017)
0.028
(0.035)
12791
.615

15461
.308

Observations
Ymean
Year Fixed Effects
* p < 0.1, ** p < 0.05, *** p < 0.01. Standard errors clustered at birth year level.
Note: All regressions include survey year, birth month fixed effects, and birth year in controls.

15778
.614

15749
.921

58

High School

Any

−0.007
(0.005)
0.020∗∗∗
(0.007)

15461
.308

0.002
(0.022)
−0.060∗∗
(0.025)
14273
.613

−0.008
(0.016)
−0.066∗∗∗
(0.024)
13998
.308

0.019∗
(0.011)
−0.028∗
(0.014)
14273
.921
Academic Zodiac Academic Zodiac Academic Zodiac
−0.023
(0.031)
−0.024∗∗∗
(0.009)

−0.017∗∗∗
(0.006)
0.016∗∗∗
(0.006)

0.012
(0.010)
−0.036∗∗∗
(0.013)

15778
.614

15749
.921

Young Dragon Direct Effect

Old Dragon Direct Effect

Young Dragon Cohort Effect

Old Dragon Cohort Effect

Observations
Ymean
Year Fixed Effects

Young Tiger Direct Effect

Old Tiger Direct Effect

Young Tiger Cohort Effect

Old Tiger Cohort Effect

Table A.13: Heterogeneity - Relative Age within School Year

College
Academic

−0.015
(0.022)
0.054∗∗∗
(0.008)

College

Any
−0.012∗∗
(0.006)
0.043∗∗∗
(0.011)

0.002
(0.009)
−0.020∗
(0.011)
14296
.922
Academic Zodiac Academic Zodiac Academic Zodiac

Observations
Ymean
Year Fixed Effects
* p < 0.1, ** p < 0.05, *** p < 0.01. Standard errors clustered at birth year level.
Note: All regressions include survey year, birth month fixed effects, and birth year in controls.

−0.008
(0.022)
0.014
(0.019)
14010
.31

−0.023∗∗
(0.009)
0.012
(0.023)
14296
.616

15778
.614

15461
.308

15749
.921

59

Dragon Direct Effect

Tiger Direct Effect

Observations
Ymean
Year Fixed Effects

Dragon Cohort Effect

Tiger Cohort Effect

Table A.14: Heterogeneity - Father Ethnicity

College
Academic

Taiwanese
0.025∗∗
(0.011)
−0.028
(0.021)
13355
.311

Chinese
−0.009
(0.055)
0.001
(0.031)
1457
.346

College

Any

High School

Any

Taiwanese
0.018∗∗
(0.008)
−0.028∗∗∗
(0.010)
13625
.627

Chinese
Taiwanese
0.007∗∗
0.025
(0.003)
(0.052)
0.086∗∗∗ −0.005
(0.018)
(0.005)
13601
1494
.62
.934

Chinese
0.042∗∗∗
(0.006)
−0.011
(0.025)
1190
.886

Academic Academic Academic Academic Academic Academic
−0.043∗∗
(0.020)
−0.010
(0.016)
10846
.313
Zodiac

−0.163∗∗∗ −0.015
(0.015)
(0.063)
−0.183∗∗∗ −0.010
(0.031)
(0.010)
11052
1203
.934
.625
Zodiac
Zodiac

−0.117
−0.033
(0.021)
(0.085)
−0.144∗∗∗ −0.005
(0.045)
(0.011)
11052
1179
.628
.353
Zodiac
Zodiac

0.015
(0.057)
0.026
(0.038)
997
.892
Zodiac

Observations
Ymean
Year Fixed Effects
* p < 0.1, ** p < 0.05, *** p < 0.01. Standard errors clustered at birth year level.
Note: All regressions include survey year, birth month fixed effects, and birth year in controls.

Table A.15: Zodiacs before 1970

College

Any

High School

Any

Middle School

Any

−0.006∗
(0.003)
−0.019∗∗∗
(0.003)

0.003
(0.004)
−0.007∗∗∗
(0.003)

0.002
(0.004)
−0.012∗∗∗
(0.001)

Dragon Direct Effect

Tiger Direct Effect

Dragon Cohort Effect

Tiger Cohort Effect

0.008
(0.007)
−0.002
(0.006)
31430
.633
Academic Zodiac Academic Zodiac Academic Zodiac

Observations
Ymean
Year Fixed Effects
* p < 0.1, ** p < 0.05, *** p < 0.01. Standard errors clustered at birth year level.
Note: All regressions include survey year, birth month fixed effects, and birth year in controls.

0.009
(0.009)
0.007
(0.010)
30740
.131

0.001
(0.010)
0.001
(0.007)
31430
.29

37914
.128

38747
.281

38747
.619

60

Table A.16: Means of Control Variables by Dragon Zodiac Group

(2)-(3)

(1)-(2)
−0.010
0.010
(0.015)
(0.019)
0.040∗∗∗
−0.014
(0.014)
(0.019)
−0.015
0.000
(0.020)
(0.027)
0.007
0.004
(0.008)
(0.011)
−0.008
−0.018
(0.009)
(0.011)
−0.012
0.008
(0.006)
(0.008)
−0.009
−0.005
(0.007)
(0.007)
−0.014
0.012
(0.009)
(0.011)
−0.002
−0.001
(0.006)
(0.008)
−0.009
0.059
(0.045)
(0.062)
0.041
0.020
(0.064)
(0.090)
−0.011
−0.006
(0.073)
(0.092)
−0.047
−0.042
(0.116)
(0.144)
−0.112
0.110
(0.071)
(0.091)
0.221∗∗∗ −0.074
(0.083)
(0.069)
0.054∗∗
−0.029
(0.034)
(0.026)

Male

Born between September and January

Born in City

Father Bachelor’s Education

Father Associate’s Education

Mother Bachelor’s Education

Mother Associate’s Education

Father Chinese

Mother Chinese

Father Folk Religion

Mother Folk Religion

Father Occupation Rank

Mother Occupation Rank

Sibling Rank(1=Oldest)

Family Size

Oldest Sibling

Observations

15778

15778

7403

12156

12156

12000

12000

15384

14870

1585

942

4174

1690

4191

4191

4191

(1)

Dragons
0.518
(0.500)
0.453
(0.498)
0.369
(0.483)
0.078
(0.268)
0.061
(0.240)
0.028
(0.166)
0.028
(0.166)
0.093
(0.291)
0.040
(0.196)
0.480
(0.502)
0.458
(0.502)
2.438
(1.209)
1.623
(1.204)
2.210
(1.339)
3.422
(1.222)
0.387
(0.488)

Non-Dragons

in Dragon Cohort

(2)

0.529
(0.499)
0.467
(0.499)
0.384
(0.487)
0.074
(0.262)
0.079
(0.270)
0.040
(0.196)
0.033
(0.179)
0.082
(0.274)
0.041
(0.198)
0.489
(0.502)
0.438
(0.500)
2.444
(1.298)
1.664
(1.199)
2.100
(1.275)
3.201
(1.174)
0.415
(0.493)

(3)

Others
0.519
(0.500)
0.427
(0.495)
0.383
(0.486)
0.067
(0.251)
0.087
(0.281)
0.032
(0.175)
0.042
(0.200)
0.095
(0.294)
0.043
(0.202)
0.430
(0.495)
0.397
(0.490)
2.456
(1.259)
1.711
(1.243)
2.212
(1.323)
3.275
(1.299)
0.362
(0.481)

Standard errors in parentheses. * p < 0.10, ** p < 0.05, *** p < 0.01 for t-tests in the last two columns.
The 2001 surveys ask parents’ occupations when the respondent is 18 while all other surveys asks when respondent is age 15.

61

Table A.17: Means of Control Variables by Tiger Zodiac Group

Male

Born between September and January

Born in City

Father Bachelor’s Education

Father Associate’s Education

Mother Bachelor’s Education

Mother Associate’s Education

Father Chinese

Mother Chinese

Father Folk Religion

Mother Folk Religion

Father Occupation Rank

Mother Occupation Rank

Sibling Rank(1=Oldest)

Family Size

Oldest Sibling

Observations

15778

15778

7403

12156

12156

12000

12000

15384

14870

1585

942

4174

1690

4191

4191

4191

(1)

Tigers
0.491
(0.500)
0.464
(0.499)
0.396
(0.489)
0.072
(0.259)
0.089
(0.285)
0.031
(0.175)
0.043
(0.202)
0.092
(0.290)
0.041
(0.197)
0.421
(0.495)
0.365
(0.484)
2.405
(1.266)
1.857
(1.280)
2.155
(1.328)
3.359
(1.338)
0.403
(0.491)

(2)

Non-Tigers

in Tiger Cohort

0.527
(0.499)
0.467
(0.499)
0.359
(0.480)
0.068
(0.253)
0.075
(0.264)
0.031
(0.175)
0.038
(0.192)
0.095
(0.293)
0.034
(0.182)
0.426
(0.496)
0.423
(0.497)
2.457
(1.314)
1.712
(1.228)
2.275
(1.439)
3.400
(1.358)
0.355
(0.479)

(3)

Others
0.519
(0.500)
0.427
(0.495)
0.383
(0.486)
0.067
(0.251)
0.087
(0.281)
0.032
(0.175)
0.042
(0.200)
0.095
(0.294)
0.043
(0.202)
0.430
(0.495)
0.397
(0.490)
2.456
(1.259)
1.711
(1.243)
2.212
(1.323)
3.275
(1.299)
0.362
(0.481)

Standard errors in parentheses. * p < 0.10, ** p < 0.05, *** p < 0.01 for t-tests in the last two columns.
The 2001 surveys ask parents’ occupations when the respondent is 18 while all other surveys asks when respondent is age 15.

(1)-(2)
−0.036∗∗
(0.018)
−0.003
(0.018)
0.038
(0.025)
0.004
(0.010)
0.014
(0.011)
0.000
(0.007)
0.004
(0.008)
−0.002
(0.011)
0.006
(0.007)
−0.004
(0.057)
−0.058
(0.075)
−0.052
(0.089)
0.145
(0.143)
−0.120
(0.100)
−0.041
(0.098)
0.048
(0.035)

(2)-(3)

0.008
(0.014)
0.039∗∗∗
(0.014)
−0.025
(0.019)
0.001
(0.008)
−0.012
(0.009)
−0.000
(0.006)
−0.003
(0.006)
−0.001
(0.008)
−0.009
(0.006)
−0.005
(0.045)
0.026
(0.059)
0.001
(0.067)
0.001
(0.107)
0.063
(0.074)
0.125∗
(0.072)
−0.007
(0.027)

62

Table A.18: Robustness Check: Accounting for Short-Term Switch in Fertility

College
Academic

College
Vocational

High School

Any

0.013
(0.013)
−0.019∗
(0.011)

0.006
(0.005)
0.000
(0.005)

0.017
(0.012)
−0.022
(0.016)

Dragon Direct Effect

Tiger Direct Effect

Dragon Cohort Effect

Tiger Cohort Effect

−0.008
(0.017)
−0.012
(0.011)
10790
.923
Academic Zodiac Academic Zodiac Academic Zodiac

−0.044∗∗
(0.017)
−0.024∗∗
(0.009)
10790
.617

−0.043∗∗
(0.020)
−0.017
(0.013)
10578
.31

Observations
Ymean
Year Fixed Effects
* p < 0.1, ** p < 0.05, *** p < 0.01. Standard errors clustered at birth year level.
Note: All regressions include survey year, birth month fixed effects, and birth year in controls.

13319
.922

13348
.615

13077
.306

Table A.19: Selection on Unobservables: DD Model - LPM

Dependent Variable: College Education (Academic)

(3)

(4)

Dragon Direct Effect

Tiger Direct Effect

Dragon Cohort Effect

Tiger Cohort Effect

(2)
0.020
(0.015)
−0.023
(0.016)

(1)
0.012
(0.016)
−0.017
(0.013)

δ
−0.260
−0.432
−0.045∗ −0.046∗ −3.834
(0.024)
(0.025)
−0.019 −0.018
2.161
(0.019)
(0.017)
12547
12547
.31
.31
.0289
.116
Yes
No
Academic Academic Zodiac Zodiac

15461
.31
.0299
No

15461
.31
.118
Yes

Observations
Ymean
Adjusted R-squared
Controls
Year Fixed Effects
* p < 0.1, ** p < 0.05, *** p < 0.01. Standard errors clustered at birth year level.
Note: All regressions include survey year, birth month fixed effects, and birth year in
controls. Columns 1 and 2 present regression coefficients while column 3 present esti-
mated delta from Oster (2017). Standard errors for estimates of delta parameter are not
presented due to their magnitude.

63

APPENDIX B

FIGURES AND TABLES FOR CHAPTER 2

Figure B.1: Current Taiwanese Education System

Source: Education in Taiwan (2015-2016) by The Ministry of Education, Taiwan, R.O.C.

64

Figure B.2: Number of High School per 10,000 High School Age Children, 1982-2010

Source: The Ministry of Education, Taiwan, R.O.C.

Figure B.3: Correlation between County Average Spending on test preparation and High School
Construction Intensity

Note: Size of the dot is the county population size in 1994

65

Figure B.4: Number of First Year Public High School Students over Number of Middle School
Graduates by track, 1983-2010

Source: The Ministry of Education, Taiwan, R.O.C.

Figure B.5: Student Teacher Ratio by high school track, 1991-2010

Source: The Ministry of Education, Taiwan, R.O.C.

66

Figure B.6: Public High School Sizes, SY 1994-1995

Source: The Ministry of Education, Taiwan, R.O.C.

Figure B.7: Treated Outcomes by School Construction Intensity group

(a) Private Tutoring Spending

(b) Private Tutoring Participation

(c) Skill Classes Spending

(d) Skill Classes Participation

Note: Average Spending is among those who report nonzero spending. Participation is among households with
students. More or less intensity refers to above or below median county public high school construction density.
County with median intensity, Taipei City, is in the less intensity group. By construction, more than 99 percent of
the households with students report positive tuition spending for every year. I therefore omit the figure on household
participation on tuition spending.

67

Figure B.8: Control Outcomes by School Construction Intensity group

(a) Tuition Spending

(b) Household Income

(c) Household Head with College Education

(d) Average Household Student Number

Note: Average Spending is among those who report nonzero spending. Participation is among households with
students. More or less intensity refers to above or below median county public high school construction density.
County with median intensity, Taipei City, is in the less intensity group. By construction, more than 99 percent of
the households with students report positive tuition spending for every year. I therefore omit the figure on household
participation on tuition spending.

68

Figure B.9: Effect of School Availability by Household Income Level

(a) Spending on Private Tutoring (Academic)

(b) Participation on Private Tutoring (Academic)

(c) Spending on Private Tutoring (Skills)

(d) Participation on Private Tutoring (Skills)

Note: All regressions control for household head age, education level, and occupation fixed effects, family type fixed
effects, number of male students, number of students at each education level, urbanicity fixed effects, and county
unemployment rate. Income quintiles are calculated within each year.

69

Figure B.10: Effect of School Availability by Household Head Education

(a) Spending on Private Tutoring (Academic)

(b) Participation on Private Tutoring (Academic)

(c) Spending on Private Tutoring (Skills)

(d) Participation on Private Tutoring (Skills)

Note: All regressions control for household head age, education level, and occupation fixed effects, family type fixed
effects, number of male students, number of students at each education level, urbanicity fixed effects, and county
unemployment rate.

70

Table B.1: Percentage of School Children Going to Private Tutoring Centers by Grade and Subject
of Study (2001)

Attending
Private
Tutoring
43.54
53.87

38.47
48.42

29.59

13.76

Subject of Study (can select multiple)

Schoolwork Foreign
Language
19.56
34.28

28.38
27.12

32.61
44.02

25.07

11.58

9.06
6.64

3.42

1.09

Computer

Sports Arts

Other

1.76
2.73

0.96
0.92

1.51

0.51

1.61
2.62

0.85
0.52

1.23

0.46

4.21
6.07

2.90
3.25

2.74

1.05

3.98
7.37

1.09
0.81

0.41

0.44

Overall
Elementary
School
Middle School
Academic High
School
Comprehensive
High School
Vocational High
School

Source: Report on Status of Student Learning and Living in Taiwan (High School and below), Department

of Statistics, The Ministry of Education, Taiwan, R.O.C. (2001)

1994

1996
1997
2000

2001

2002
2005

Table B.2: Important Events of Taiwanese Education Reform

Protest for Education Reform
Committee for Education Reform is established
Establish Comprehensive High School
Establish Complete High School
Government stop publishing textbooks for elementary and middle
schools and retreat to the role of reviewing textbooks published by
private companies
Introduce Joint Entrance Exam for High Schools and Junior Colleges.
Expand methods for entering high school
Introduce New Curriculum Guidelines for Elementary and Middle
Schools
Expand methods for entering universities
Government resumed publishing textbooks
Source: The Ministry of Education, Taiwan, R.O.C. website

71

Table B.3: Percentage change in public high school density from 1994-2001 by County

Taipei City
Kaohsiung City
Taipei County
Yilan County
Taoyuan County
Hsinchu County
Miaoli County
Taichung County
Changhua County
Nantou County
Yunlin County
Chiayi County
Tainan County
Kaohsiung County
Pingtung County
Taitung County
Hualien County
Penghu County
Keelung City
Hsinchu City
Taichung City
Chiayi City
Tainan City

Percentage
Change
in
Density
61.53(cid:156)
87.20(cid:156)
199.00(cid:156)
38.79(cid:156)
50.62(cid:156)
45.74(cid:156)
76.32(cid:156)
98.92(cid:156)
29.85(cid:156)
50.39(cid:156)
69.03(cid:156)
26.40(cid:156)
29.15(cid:156)
146.15(cid:156)
66.59(cid:156)
79.74(cid:156)
73.19(cid:156)
32.82(cid:156)
111.25(cid:156)
82.55(cid:156)
29.07(cid:156)
24.12(cid:156)
2.19(cid:156)

Population
in 1994
2,653,578
1,416,248
3,260,731
464,359
1,483,955
401,188
558,191
1,379,949
1,281,296
546,091
753,791
564,381
1,069,339
1,179,635
909,110
254,718
358,247
92,645
364,520
338,140
832,654
260,368
702,658

Source: The Ministry of Education, Taiwan, R.O.C. and The Ministry of the Interior, Taiwan,
R.O.C.

72

Table B.4: Role of Controls - Probit

(1)

(2)

(3)

(4)

(5)

Panel A: Private Tutoring Participation (Any Type)

Public High School Probability 0.002∗∗∗ −0.002∗∗ −0.001
(0.001)
96669
0.581

(0.001)
96669
0.581

(0.000)
96669
0.581

Observations
Ymean

−0.002∗∗∗ −0.001∗
(0.001)
(0.001)
96669
96669
0.581
0.581

Panel B: Private Tutoring Participation (Academic)

Public High School Probability 0.001∗∗∗ −0.002∗∗ −0.002
(0.001)
96669
0.503

(0.001)
96669
0.503

(0.000)
96669
0.503

Observations
Ymean

−0.003∗∗∗ −0.002∗∗
(0.001)
(0.001)
96669
96669
0.503
0.503

Panel C: Private Tutoring Participation (Skills)

0.001∗
(0.001)
96669
0.169

0.002∗∗
(0.001)
96669
0.169

0.002∗∗
(0.001)
96669
0.169

Public High School Probability 0.002∗∗∗ 0.001∗
(0.001)
96669
0.169

(0.000)
96669
0.169

Observations
Ymean
Controls
County Fixed Effects
Head Characteristics
Family Type
Macro Conditions
* p < 0.1, ** p < 0.05, *** p < 0.01. Standard errors clustered at birth year level.
Note: Head Controls include age, education level, and occupation fixed effects of household head.
Family type include family type fixed effects, number of male students, and number of students at
each education level. Macro controls include urbanicity fixed effects and county unemployment rate.

Yes
Yes
Yes
No

Yes
Yes
Yes
Yes

No
No
No
No

Yes
No
No
No

Yes
Yes
No
No

73

Table B.5: Role of Controls - Tobit

(1)

(2)

(3)

(4)

(5)

Panel A: Private Tutoring Spending (Any Type)

Public High School Probability 0.432∗∗∗ −0.191
(0.118)
96669
28.913

(0.045)
96669
28.913

Observations
Ymean

−0.076
(0.128)
96669
28.913

−0.307∗∗∗ −0.192∗
(0.110)
(0.117)
96669
96669
28.913
28.913

−0.410∗∗∗ −0.309∗∗
(0.124)
(0.130)
96669
96669
23.033
23.033

0.223
(0.180)
96669
5.880

0.313∗
(0.176)
96669
5.880

Panel B: Private Tutoring Spending (Academic)

Observations
Ymean

(0.034)
96669
23.033

(0.134)
96669
23.033

Public High School Probability 0.320∗∗∗ −0.292∗∗ −0.209
(0.141)
96669
23.033
Panel C: Private Tutoring Spending (Skills)
0.401∗∗
(0.176)
96669
5.880

Public High School Probability 0.461∗∗∗ 0.242
(0.176)
96669
5.880

(0.061)
96669
5.880

Observations
Ymean
Controls
County Fixed Effects
Head Characteristics
Family Type
Macro Conditions
* p < 0.1, ** p < 0.05, *** p < 0.01. Standard errors clustered at birth year level.
Note: Head Controls include age, education level, and occupation fixed effects of household head.
Family type include family type fixed effects, number of male students, and number of students at
each education level. Macro controls include urbanicity fixed effects and county unemployment rate.

Yes
Yes
Yes
Yes

Yes
Yes
Yes
No

Yes
Yes
No
No

Yes
No
No
No

No
No
No
No

74

Table B.6: IV First Stage

Dependent Variable: Public High School Density
(2)

(1)

(3)

(4)

(5)

(0.005)
96669
2.155
439.435

(0.005)
96669
2.155
371.251

(0.003)
96669
2.155
27.367

Public High School Probability 0.048∗∗∗ 0.046∗∗∗ 0.046∗∗∗ 0.046∗∗∗ 0.045∗∗∗
(0.005)
96669
2.155
297.215

Observations
Ymean
F Statistic
County Fixed Effects
Head Characteristics
Family Type
Macro Conditions
* p < 0.1, ** p < 0.05, *** p < 0.01. Standard errors clustered at birth year level.
Note: Head Controls include age, education level, and occupation fixed effects of household head.
Family type include family type fixed effects, number of male students, and number of students at
each education level. Macro controls include urbanicity fixed effects and county unemployment rate.

(0.005)
96669
2.155
310.733

Yes
Yes
Yes
Yes

Yes
Yes
Yes
No

No
No
No
No

Yes
No
No
No

Yes
Yes
No
No

75

Table B.7: IV Reduced Form - Probit

(1)

(2)

(3)

(4)

(5)

Panel A: Private Tutoring Participation (Any Type)

Public High School Density 0.006
(0.004)
96669
0.581

Observations
Ymean

−0.023∗∗∗ −0.017∗ −0.020∗∗ −0.017∗
(0.009)
(0.009)
96669
96669
0.581
0.581

(0.009)
96669
0.581

(0.009)
96669
0.581

Panel B: Private Tutoring Participation (Academic)

Public High School Density 0.005
(0.004)
96669
0.503

Observations
Ymean

−0.026∗∗∗ −0.020∗∗ −0.023∗∗ −0.021∗∗
(0.009)
(0.009)
96669
96669
0.503
0.503

(0.009)
96669
0.503

(0.010)
96669
0.503

Panel C: Private Tutoring Participation (Skills)

Public High School Density 0.008∗∗∗ −0.004
(0.007)
96669
0.169

(0.003)
96669
0.169

0.002
(0.007)
96669
0.169

−0.001
(0.007)
96669
0.169

−0.000
(0.007)
96669
0.169

Observations
Ymean
Controls
Yes
County Fixed Effects
Yes
Head Characteristics
Yes
Family Type
Macro Conditions
No
* p < 0.1, ** p < 0.05, *** p < 0.01. Standard errors clustered at birth year level.
Note: Head Controls include age, education level, and occupation fixed effects of household head.
Family type include family type fixed effects, number of male students, and number of students at
each education level. Macro controls include urbanicity fixed effects and county unemployment
rate.

Yes
Yes
Yes
Yes

Yes
Yes
No
No

Yes
No
No
No

No
No
No
No

76

Table B.8: IV Reduced Form - Tobit

(1)

(2)

(3)

(4)

(5)

Panel A: Private Tutoring Spending (Any Type)

Public High School Density 1.601∗∗ −3.480∗∗∗ −2.119
(1.376)
96669
28.913

(0.723)
96669
28.913

(1.262)
96669
28.913

Observations
Ymean

−2.874∗∗ −2.121∗
(1.255)
(1.275)
96669
96669
28.913
28.913

Panel B: Private Tutoring Spending (Academic)

Public High School Density 1.112∗ −3.729∗∗∗ −2.728∗∗ −3.357∗∗∗ −2.667∗∗
(1.243)
96669
23.033

(1.266)
96669
23.033

(0.625)
96669
23.033

(1.373)
96669
23.033

(1.289)
96669
23.033

Observations
Ymean

Panel C: Private Tutoring Spending (Skills)

Public High School Density 2.304∗∗∗ −1.462
(1.692)
96669
5.880

(0.817)
96669
5.880

0.334
(1.756)
96669
5.880

−0.635
(1.698)
96669
5.880

−0.275
(1.709)
96669
5.880

Observations
Ymean
Controls
Yes
County Fixed Effects
Yes
Head Characteristics
Yes
Family Type
Macro Conditions
No
* p < 0.1, ** p < 0.05, *** p < 0.01. Standard errors clustered at birth year level.
Note: Head Controls include age, education level, and occupation fixed effects of household head.
Family type include family type fixed effects, number of male students, and number of students at
each education level. Macro controls include urbanicity fixed effects and county unemployment
rate.

Yes
Yes
Yes
Yes

Yes
Yes
No
No

Yes
No
No
No

No
No
No
No

77

Table B.9: Comparison of Methods - OLS/IV

Spending

OLS

IV

Participation
IV

Probit

Dependent Variable: Private Tutoring Classes
Public High School Probability −0.309∗∗ −0.473∗∗ −0.002∗∗ −0.011∗∗
(0.005)
96669

(0.001)
96669

(0.234)
96669

(0.124)
96669

Observations
Dependent Variable: Skill Classes
Public High School Probability 0.313∗ −0.101
(0.311)
96669

(0.176)
96669

0.002∗∗ −0.001
(0.006)
(0.001)
96669
96669

Observations
* p < 0.1, ** p < 0.05, *** p < 0.01. Standard errors clustered at birth year level.
Note: All regressions control for household head age, education level, and occupation
fixed effects, family type fixed effects, number of male students, number of students at
each education level, urbanicity fixed effects, and county unemployment rate.

78

Table B.10: School Districts - Participation

(1)

(2)

(3)

(4)

(5)

Panel B: Private Tutoring Participation (Academic)

Public High School Probability −0.001
(0.000)
96669
0.581

Observations
Ymean

−0.003∗∗∗ −0.003∗∗ −0.004∗∗∗ −0.003∗∗∗
(0.001)
(0.001)
96669
96669
0.581
0.581

(0.001)
96669
0.581

(0.001)
96669
0.581

Panel B: Private Tutoring Participation (Academic)

Public High School Probability −0.000
(0.000)
96669
0.503

Observations
Ymean

−0.004∗∗∗ −0.003∗∗∗ −0.004∗∗∗ −0.003∗∗∗
(0.001)
(0.001)
96669
96669
0.503
0.503

(0.001)
96669
0.503

(0.001)
96669
0.503

Panel C: Private Tutoring Participation (Skills)

Public High School Probability −0.000
(0.000)
96669
0.169

0.001
(0.001)
96669
0.169

0.001
(0.001)
96669
0.169

0.000
(0.001)
96669
0.169

0.001
(0.001)
96669
0.169

Observations
Ymean
Controls
District Fixed Effects
Head Characteristics
Family Type
Macro Conditions
* p < 0.1, ** p < 0.05, *** p < 0.01. Standard errors clustered at birth year level.
Note: Head Controls include age, education level, and occupation fixed effects of household head.
Family type include family type fixed effects, number of male students, and number of students at
each education level. Macro controls include urbanicity fixed effects and county unemployment rate.

Yes
Yes
Yes
Yes

Yes
Yes
Yes
No

Yes
Yes
No
No

Yes
No
No
No

No
No
No
No

79

Table B.11: School Districts - Spending - Tobit

(1)

(2)

(3)

(4)

(5)

Panel A: Private Tutoring Spending (Any Type)

Public High School Probability −0.078
(0.068)
96669
28.913

Observations
Ymean

−0.471∗∗∗ −0.359∗∗ −0.634∗∗∗ −0.464∗∗∗
(0.147)
(0.152)
96669
96669
28.913
28.913

(0.155)
96669
28.913

(0.144)
96669
28.913

Panel B: Private Tutoring Spending (Academic)

Public High School Probabilityt −0.036
(0.066)
96669
23.033

Observations
Ymean

−0.552∗∗∗ −0.465∗∗∗ −0.711∗∗∗ −0.547∗∗∗
(0.141)
(0.148)
96669
96669
23.033
23.033

(0.151)
96669
23.033

(0.142)
96669
23.033
Panel C: Private Tutoring Spending (Skills)

0.101
(0.255)
96669
5.880

0.233
(0.253)
96669
5.880

0.029
(0.262)
96669
5.880

0.087
(0.265)
96669
5.880

Public High School Probability −0.080
(0.075)
96669
5.880

Observations
Ymean
Controls
District Fixed Effects
Head Characteristics
Family Type
Macro Conditions
* p < 0.1, ** p < 0.05, *** p < 0.01. Standard errors clustered at birth year level.
Note: Head Controls include age, education level, and occupation fixed effects of household head.
Family type include family type fixed effects, number of male students, and number of students at
each education level. Macro controls include urbanicity fixed effects and county unemployment rate.

Yes
Yes
Yes
Yes

Yes
Yes
Yes
No

Yes
No
No
No

Yes
Yes
No
No

No
No
No
No

80

Table B.12: Placebo Test

Counties

Districts

Public High School Probability

Spending Participation Spending Participation
0.188
(0.254)
10580
1.074

−0.473
(0.448)
10580
Observations
Ymean
1.074
* p < 0.1, ** p < 0.05, *** p < 0.01. Standard errors clustered at birth year level.
Note: All regressions control for household head age, education level, and occupation fixed effects,
family type fixed effects, family size fixed effects, urbanicity fixed effects, and county unemployment
rate. Sample is restricted to households with persons age 26-50.

−0.001
(0.001)
10580
0.050

0.000
(0.001)
10580
0.050

81

REGARDING THE OFFICIAL NUMBER OF PRIVATE TUTORING CENTERS

APPENDIX C

The phenomenon of the increasing number of private tutoring centers has been documented exten-
sively in the education literature in Taiwan. Many of these articles used official statistics from the
Ministry of Education for their analysis.1 However, different publications report different numbers
of private tutoring centers for the same year. If I juxtapose the number of private tutoring centers
reported by articles, magazines, and government reports over the years, I find that for a given year,
the number of private tutoring centers reported is decreasing as the year the data was accessed
increases.2 Thus, the issue is one of reporting.

For example, for the year 2001, there were about 7,000 private tutoring centers reported. The
number decreased to 6,151 reported in 2009, 5,884 reported in 2010, and for the reports published
in 2016, the official number of private tutoring centers in 2001 is only about 4,500.
I see the
same trend if we look at different months within the same year or focus on subcategories of private
tutoring centers that specialize in test preparation.

My hypothesis for this phenomenon is that the official numbers are actually the number of
private tutoring centers that survived until the time of access. After contacting the Ministry of
Education, they confirmed that the historical trend of private tutoring centers does not include
private tutoring centers that were discontinued/abolished. This significantly underestimates the
past number of private tutoring centers and always shows a rising trend in private tutoring centers
even if the numbers are actually decreasing.

It is important to note I am not arguing there is no increase in number of private tutoring centers
1The Ministry of Education created an online depository to track the private tutoring centers in
2001. The website is https://bsb.kh.edu.tw/. The Ministry of Education has updated the historical
trend in 2017.

2According to the National Digital Library of Theses and Dissertations in Taiwan, there are 286
master theses or PhD dissertations citing the official number of private tutoring centers from the
Ministry of Education as of July 30th of 2018. I only select a number of publications that reported
more than one year of numbers they observed from the official website.

82

at all. If I only use the number of private tutoring centers in the last year of each study, then I can
get closer to what the true time trend should be. I redrew the trend using the last year of data from
each study and compared that to the official data. I found an increase before 2011, but after 2011,
there seems to be no change or a slight decrease in the number of private tutoring centers.

Even if the data problem gets fixed, we still need to take caution interpreting the time trend
in official number of private tutoring centers. Ultimately, the statistic researchers care about is
time use of school children. However, the official number of private tutoring centers is actually
measuring the number of private tutoring center licenses. The law only requires private tutoring
centers to register their address, number of classrooms, size of the classroom and the total size of the
land the private tutoring center is using. There is a lot of variation in number of classes, students,
and lesson hours among different private tutoring centers. So far there is no quantitative evidence
on the changes of the services the private tutoring centers provide. Given qualitative evidence of
newer private tutoring centers tend to be smaller and more geographically dispersed (Liu, 2006),
we can expect some of the changes in number of private tutoring centers to be driven by changes
in the nature of private tutoring centers. In addition, it is possible that due to the establishment
of the online registry of private tutoring centers, previously unregistered private tutoring centers
are incentivized to register and the number of registered private tutoring centers can increase as a
result.

In my paper, I used data collected from households. Households have less incentive to lie about
their private tutoring spending. The household data also includes various demographic data that
may explain the changes in private tutoring. My paper should be interpreted as exploring impacts
from the demand side while assuming linear trends in supply of private tutoring. The exact impact
of changes in the supply side of private tutoring requires further study.

83

Figure C.1: Number of Private Tutoring Centers by Different Sources

(a) All Categories

(b) All Categories (within 2016)

(c) Academic & Foreign Language

(d) Academic

Note: Sources are listed in the references. For 2016 data, June numbers are from Chen (2016). The rest of the data
from 2016 are collected by the author.

84

Figure C.2: Number of Private Tutoring Centers - Adjusted

Note: 2016 data is collected by the author. Adjusted numbers are taken using only the last year
reported in each publication.

85

APPENDIX D

FIGURES AND TABLES FOR CHAPTER 3

Figure D.1: Trends Over Time

(a) Educational attainment

(b) Superstition Use

(c) Superstition View

(d) Fortune Telling Use by Question Type

(e) Feng Shui Use by Question Type

Source: Author’s calculations using TSCS.

86

Figure D.2: Use and View of Superstition by Education Group

(a) View of Fortune Telling

(b) Use of Fortune Telling

(c) View of Feng Shui

(d) Use of Feng Shui

(e) View of Choosing Dates

Source: Author’s calculations using TSCS.

87

Figure D.3: Use of Superstition by Question Type and Education Group

(a) Fortune Telling Use - Any

(b) Feng Shui Use - Any

(c) Fortune Telling Use - Recent

(d) Feng Shui Use - Recent

(e) Fortune Telling Use - Voluntary

Notes: Source: Author’s calculations using TSCS.

88

Figure D.4: Outcomes by Birth School Cohort

(a) Years of Education

(b) View of Fortune Telling

(c) Use of Fortune Telling

(d) View of Feng Shui

(e) Use of Feng Shui

(f) View of Choosing Dates

Notes: The birth cohort of 1955 includes those born between September 1955 and August 1956. Fitted values are
fitted using a local quadratic function. Source: Author’s calculations using TSCS.

89

Figure D.5: Use of Superstition by Question Type and Birth School Cohort

(a) Fortune Telling Use - Any

(b) Feng Shui Use - Any

(c) Fortune Telling Use - Recent

(d) Feng Shui Use - Recent

(e) Fortune Telling Use - Voluntary

Notes: The birth cohort of 1955 includes those born between September 1955 and August 1956. Fitted values are
fitted using a local quadratic function. Source: Author’s calculations using TSCS.

90

Figure D.6: Robustness Check - Bandwidth (Overall)

(a) View of Fortune Telling

(b) Use of Fortune Telling

(c) View of Feng Shui

(d) Use of Feng Shui

(e) View of Choosing Dates

Notes: Bandwidth refers to the number of academic cohort from the cutoff cohort of 1954 included in the analysis.
For example, bandwidth of 3 means the sample includes 6 birth cohorts: 3 cohorts born after 1954 and 3 cohorts born
before 1954.

91

Figure D.7: Robustness Check - Bandwidth (by Question Type)

(a) Fortune Telling Use - Any

(b) Feng Shui Use - Any

(c) Fortune Telling Use - Recent

(d) Feng Shui Use - Recent

(e) Fortune Telling Use - Voluntary

Notes: Bandwidth refers to the number of academic cohort from the cutoff cohort of 1954 included in the analysis.
For example, bandwidth of 3 means the sample includes 6 birth cohorts: 3 cohorts born after 1954 and 3 cohorts born
before 1954.

92

Table D.1: Effect of Reform on Education

(1)

(2)

1.420∗∗∗ 0.563∗∗∗
(0.106)
(0.132)

(3)

Exposed to Reform

3 Cohort Before Reform

2 Cohort Before Reform

1 Cohort Before Reform

1st Cohort Exposed to Reform

2nd Cohort Exposed to Reform

3rd Cohort Exposed to Reform

Birth School Cohort

Birth School Cohort Squared

Male

Born in City

−0.056
(0.129)
0.078
(0.148)
−0.111
(0.173)
0.435∗∗
(0.179)
0.378∗∗
(0.149)
0.206
(0.212)
4.333 −21.895
(13.528)
(33.289)
−0.001
0.006
(0.003)
(0.009)
1.759∗∗∗ 1.758∗∗∗
(0.086)
(0.086)
1.412∗∗∗ 1.411∗∗∗
(0.143)
(0.143)
9513
9513
10.136
10.136
13.895
28.349

Observations
Ymean
F Statistic
* p < 0.1, ** p < 0.05, *** p < 0.01. Standard errors clustered at birth year

9513
10.136
115.054

Note: Sample limited to those born 6 years around the treated cohort. F
statistic calculated by partially out birth school cohort, birth school cohort
squared, gender, and city dummy.

level.

93

Table D.2: Reduced Form - Overall

Superstition View

Fortune Telling

Feng Shui

Choosing Dates

Dummy Event Study Dummy Event Study Dummy Event Study
−0.038
(0.037)

−0.051∗
(0.025)

−0.015
(0.042)

0.014
(0.013)
0.001
(0.015)
0.016
(0.021)
−0.016
(0.015)
−0.047
(0.038)
−0.027∗
(0.015)
4147
0.791

4147
0.791

Exposed to Reform

3 Cohort Before Reform

2 Cohort Before Reform

1 Cohort Before Reform

1st Cohort Exposed to Reform

2nd Cohort Exposed to Reform

3rd Cohort Exposed to Reform

Observations
Ymean

1860
0.346

Exposed to Reform

3 Cohort Before Reform

2 Cohort Before Reform

1 Cohort Before Reform

1st Cohort Exposed to Reform

2nd Cohort Exposed to Reform

3rd Cohort Exposed to Reform

Fortune Telling

Feng Shui

Dummy Event Study Dummy Event Study
−0.014
(0.018)

−0.008
(0.008)

0.023∗
(0.012)
−0.005
(0.016)
0.007
(0.034)
0.037∗∗
(0.018)
−0.050
(0.041)
−0.031
(0.021)
4127
0.607

0.017
(0.014)
−0.004
(0.020)
−0.000
(0.017)
−0.012
(0.017)
0.002
(0.019)
0.026∗
(0.016)
6482
0.155

0.002
(0.030)
0.002
(0.035)
0.043
(0.049)
0.038
(0.037)
−0.068
(0.049)
−0.018
(0.033)
1860
0.346

4127
0.607

Superstition Use

−0.004
(0.032)
−0.006
(0.053)
−0.001
(0.027)
−0.016
(0.026)
−0.005
(0.022)
−0.016
(0.018)
7731
0.350

94

Observations
Ymean
* p < 0.1, ** p < 0.05, *** p < 0.01. Standard errors clustered at birth year level.
Note: All regressions controls for gender, place of birth, marital status, religion, employment status, household income,
quadratic age, and survey year.

7731
0.350

6482
0.155

Table D.3: Reduced Form - Use by Question Type

Fortune Telling Use
Any

Recent

Voluntary

Dummy Event Study Dummy Event Study Dummy Event Study
−0.009
(0.027)

−0.031
(0.028)

−0.039
(0.034)

0.000
(0.030)
0.030
(0.049)
0.031
(0.036)
0.036
(0.035)
−0.044
(0.044)
−0.025
(0.026)
2761
0.377

2761
0.377

Exposed to Reform

3 Cohort Before Reform

2 Cohort Before Reform

1 Cohort Before Reform

1st Cohort Exposed to Reform

2nd Cohort Exposed to Reform

3rd Cohort Exposed to Reform

Observations
Ymean

4348
0.353

Exposed to Reform

3 Cohort Before Reform

2 Cohort Before Reform

1 Cohort Before Reform

1st Cohort Exposed to Reform

2nd Cohort Exposed toReform

3rd Cohort Exposed to Reform

Any

Recent

Dummy Event Study Dummy Event Study
−0.024∗∗
(0.009)

0.018
(0.015)

−0.034
(0.026)
−0.014
(0.055)
0.003
(0.040)
−0.044
(0.027)
−0.006
(0.033)
−0.054∗∗
(0.025)
2996
0.186

0.009
(0.018)
−0.042
(0.029)
−0.040∗
(0.024)
−0.020
(0.023)
−0.006
(0.028)
0.007
(0.022)
2601
0.215

0.000
(0.017)
−0.014
(0.037)
−0.031
(0.022)
−0.045∗
(0.026)
0.001
(0.027)
−0.020
(0.016)
4348
0.353
Feng Shui Use

2996
0.186

0.020∗
(0.012)
0.022
(0.014)
0.026∗∗
(0.013)
−0.003
(0.012)
0.009
(0.012)
0.040∗∗∗
(0.011)
3881
0.115

95

Observations
Ymean
* p < 0.1, ** p < 0.05, *** p < 0.01. Standard errors clustered at birth year level.
Note: All regressions controls for gender, place of birth, marital status, religion, employment status, household income,
quadratic age, and survey year.

3881
0.115

2601
0.215

Table D.4: IV results - Overall

Superstition View

Fortune Telling
IV
OLS
Years of Education 0.010∗∗∗
0.004
(0.003)
(0.007)
1853
1853
0.347
0.347
457.736

Observations
Ymean
First Stage F

OLS

Feng Shui
IV
−0.013∗∗∗ −0.012
(0.002)
(0.008)
4120
4120
0.607
0.607

1172.889

Superstition Use

Choosing Dates
IV
OLS

−0.017∗∗∗ −0.020∗∗∗
(0.001)
(0.007)
4140
4140
0.791
0.791

1400.681

Fortune Telling
OLS
IV
Years of Education 0.011∗∗∗ −0.007
(0.008)
7722
0.350

(0.001)
7722
0.350

OLS
−0.000
(0.001)
6475
0.155

Feng Shui
IV
−0.003
(0.004)
6475
0.155
288.623

Observations
Ymean
First Stage F
* p < 0.1, ** p < 0.05, *** p < 0.01. Standard errors clustered at birth year level.
Note: All regressions controls for gender, place of birth, marital status, religion, employment status,

7908.610

household income, quadratic age, and survey year.

96

Table D.5: IV results - Use by Question Type

Fortune Telling Use

Any

OLS
Years of Education 0.014∗∗∗
(0.002)
4346
0.354

Observations
Ymean
First Stage F

IV
0.004
(0.008)
4346
0.354

10031.399

Recent

OLS
IV
0.004∗∗ −0.004
(0.001)
(0.007)
2989
2989
0.185
0.185
64.262

Voluntary

OLS
IV
0.008∗∗∗ −0.025∗∗
(0.002)
(0.010)
2754
2754
0.376
0.376
73.020

Feng Shui Use

Any

Recent

IV
0.007
(0.006)
2596
0.214
453.691

OLS
Years of Education −0.001
(0.002)
3879
0.115

IV

−0.009∗∗∗
(0.004)
3879
0.115

OLS
0.000
(0.003)
2596
0.214

Observations
Ymean
First Stage F
* p < 0.1, ** p < 0.05, *** p < 0.01. Standard errors clustered at birth year level.
Note: All regressions controls for gender, place of birth, marital status, religion, employment status,

1801.054

household income, quadratic age, and survey year.

97

Table D.6: Heterogeneous Effects by Religious Affiliation - Overall

Affiliation

No Religion Folk Religion Buddhist

Daoist Western Religion

Panel A: Agree with Fortune Telling

Years of Education

Observations
First Stage F

−0.023
(0.017)
206

110.263

0.021
(0.016)
635
35.573

0.019
(0.019)
242
.

Years of Education

Observations
First Stage F

Years of Education

Observations
First Stage F

Years of Education

Observations
First Stage F

Years of Education

−0.021∗∗
(0.010)
605

157.309

−0.028∗∗
(0.011)
609

257.725

−0.033∗∗
(0.013)
918
35.143

−0.001
(0.008)
835
90.085

Observations
First Stage F
* p < 0.1, ** p < 0.05, *** p < 0.01. Standard errors clustered at birth year level.

−0.009
(0.011)
564

119.645

−0.012
(0.009)
1313
440.585

−0.013∗∗
(0.006)
1322
837.060

−0.004
(0.008)
2363

5244.540

−0.004
(0.005)
2139

1393.710

Panel B: Agree with Feng Shui

−0.010
(0.014)
1319
315.514

0.013
(0.011)
527
.

Panel C: Agree with Choosing Dates

−0.017
(0.015)
1320

−0.012
(0.008)
532

13905.934

171.443
Panel D: Fortune Telling Use
−0.003
(0.017)
828

−0.011
(0.014)
2483

1705.718

107.881

Panel E: Feng Shui Use
0.008
(0.011)
755
73.800

−0.002
(0.011)
2157
925.725

−0.020
(0.025)
206
20.597

−0.033
(0.024)
348

7982.958

−0.029
(0.024)
349

4448.813

0.001
(0.014)
655
94.311

−0.004
(0.008)
581

131.235

98

Table D.7: Heterogeneous Effects by Religious Affiliation - by Question Type

No Religion Folk Religion Buddhist

Daoist Western Religion

Panel F: Any Fortune Telling Use

Years of Education

Observations
First Stage F

−0.018
(0.019)
557
25.345

0.008
(0.016)
1468

1694.443

0.005
(0.019)
379

1222.106

0.009
(0.007)
1173

5354.398

Panel G: Recent Fortune Telling Use

0.002
(0.010)
932

1277.171

−0.002
(0.015)
406
.

0.001
(0.011)
1003
255.547

Panel H: Voluntary Fortune Telling Use

Years of Education

Observations
First Stage F

Years of Education

Observations
First Stage F

−0.021∗
(0.013)
330
16.071

−0.029∗
(0.017)
286

1676.891

Years of Education

Observations
First Stage F

0.006
(0.006)
557
25.338

0.017
(0.023)
294
9.184

0.003
(0.026)
318
63.594

−0.021
(0.028)
306
66.787

−0.009
(0.017)
294
9.175

−0.003
(0.010)
287
51.786

108.799

−0.021
(0.019)
861

−0.006
(0.023)
353
77.343
Panel I: Any Feng Shui Use
0.005
(0.013)
379

−0.007
(0.012)
1468

2679.406

1220.469
Panel J: Recent Feng Shui Use

−0.013
(0.012)
948
93.557

−0.010
(0.008)
1173

3145.246

0.004
(0.007)
966

298.727

Years of Education

−0.001
(0.012)
278
24.174

0.011
(0.010)
689
12.176

0.008
(0.013)
376

Observations
First Stage F
* p < 0.1, ** p < 0.05, *** p < 0.01. Standard errors clustered at birth year level.

128.696

99

Table D.8: Dropping Controls - Overall

Dropped Variable

None

Born in City Male

Married

Religion Employed

Income

Agree Fortune Telling

0.029∗∗∗
(0.010)
1853
84.017

0.002
(0.008)
1853
225.056

0.002
(0.007)
1853

.

−0.001
(0.008)
1853

.

0.007
(0.006)
1853
46.217

Agree Feng Shui

Agree Choosing Dates

0.002
(0.010)
4120
450.525

−0.010
(0.009)
4140
912.472

−0.014∗
(0.007)
4120

1.17e+05

−0.013∗
(0.008)
4120

1619.493

−0.010
(0.008)
4120

5207.283

−0.013∗
(0.007)
4120

1442.792

−0.022∗∗∗ −0.021∗∗∗ −0.021∗∗∗ −0.020∗∗∗
(0.006)
(0.006)
4140
4140

(0.007)
4140

(0.006)
4140

12667.817

2325.373

.

1787.811

Fortune Telling Use

0.024∗∗ −0.004
(0.010)
(0.007)
7722
7722
451.746

6119.848

−0.006
(0.008)
7722

2146.130

−0.005
(0.007)
7722

1822.214

−0.002
(0.007)
7722
679.665

−0.003
(0.004)
6475
366.103

0.002
(0.004)
6475
285.451

−0.001
(0.003)
6475

1234.193

Years of Education

Observations
First Stage F

0.004
(0.007)
1853

1992.529

Years of Education −0.012
(0.008)
4120
905.304

Observations
First Stage F

−0.000
(0.007)
1853
759.748

−0.012
(0.010)
4120

1628.965

Years of Education −0.020∗∗∗
(0.007)
4140

Observations
First Stage F

1362.223

Years of Education −0.007
(0.008)
7722

Observations
First Stage F

5000.205

−0.016∗∗
(0.006)
4140

2047.151

−0.018∗
(0.009)
7722
230.011

Years of Education −0.003
(0.004)
6475
288.816

Feng Shui Use
−0.018∗∗∗ −0.004
(0.004)
(0.005)
6475
6475
Observations
First Stage F
343.864
100.263
* p < 0.1, ** p < 0.05, *** p < 0.01. Standard errors clustered at birth year level.

0.005
(0.006)
6475
125.791

100

Table D.9: Dropping Controls - Use by Question Type

Dropped Variable

None

Born in City Male

Married

Religion Employed

Income

0.005
(0.008)
4346

1258.973

−0.005
(0.006)
2989
74.957

0.005
(0.007)
4346
347.276

−0.003
(0.006)
2989
206.697

0.007
(0.007)
4346
419.903

−0.001
(0.005)
2989
174.404

Any Fortune Telling Use
0.006
(0.007)
4346

0.018∗
(0.010)
4346
325.638

1239.743

−0.005
(0.011)
4346

3427.246

Recent Fortune Telling Use
0.018∗∗ −0.004
−0.007
(0.009)
(0.006)
(0.007)
2989
2989
2989
74.148
59.244
154.139
Voluntary Fortune Telling Use
−0.034∗∗∗
(0.010)
2754
50.427

(0.010)
2754
251.584

Any Feng Shui Use

Years of Education

Observations
First Stage F

0.004
(0.008)
4346

4029.688

Years of Education −0.004
(0.007)
2989
61.057

Observations
First Stage F

Years of Education −0.025∗∗
(0.010)
2754
72.894

Observations
First Stage F

Years of Education −0.009∗∗∗
(0.004)
3879
952.509

Observations
First Stage F

0.029∗∗ −0.020∗∗ −0.023∗∗ −0.026∗∗ −0.014∗
(0.008)
(0.013)
2754
2754
312.432
81.158

(0.011)
2754
128.313

(0.011)
2754
76.699

−0.007
(0.007)
3879
296.182

−0.018∗∗∗ −0.011∗∗∗ −0.008∗∗
(0.003)
(0.004)
3879
3879
550.352

(0.003)
3879

.

0.000
(0.004)
3879

3611.439

−0.004
(0.004)
3879

1471.580

.
Recent Feng Shui Use

Years of Education

0.004
(0.006)
2596
Observations
First Stage F
255.388
* p < 0.1, ** p < 0.05, *** p < 0.01. Standard errors clustered at birth year level.

0.007
(0.006)
2596
623.366

−0.004
(0.013)
2596
316.349

0.014∗∗
(0.006)
2596
277.788

0.004
(0.006)
2596
242.233

0.006
(0.006)
2596
147.274

0.004
(0.005)
2596
361.687

101

BIBLIOGRAPHY

102

BIBLIOGRAPHY

Aarnio, K. and Lindeman, M. (2005). Paranormal beliefs, education, and thinking styles. Person-

ality and Individual Differences, 39(7):1227 – 1236.

Agarwal, S., Qian, W., Sing, T. F., and Tan, P. L. (2017). Dragon babies. George-
town McDonough School of Business Research Paper No. 3032575. Available at SSRN:
https://ssrn.com/abstract=3032575.

Akabayashi, H. (2008). Lives of the firehorse cohort: What the statistics show. Japanese Economy,

35(3):34–54.

Almond, D., Chee, C. P., Sviatschi, M. M., and Zhong, N. (2015). Auspicious birth dates among

chinese in california. Economics & Human Biology, 18(Supplement C):153 – 159.

Altonji, J. G., Elder, T. E., and Taber, C. R. (2005). Selection on observed and unobserved variables:
Assessing the effectiveness of catholic schools. Journal of Political Economy, 113(1):151–184.
Andreoni, J. and Payne, A. A. (2003). Do government grants to private charities crowd out giving

or fund-raising? American Economic Review, 93(3):792–812.

Andreoni, J. and Payne, A. A. (2011). Is crowding out due entirely to fundraising? evidence from
a panel of charities. Journal of Public Economics, 95(5–6):334 – 343. Charitable Giving and
Fundraising Special Issue.

Angrist, J. D. and Pischke, J.-S. (2008). Mostly Harmless Econometrics: An Empiricist’s Compan-

ion. Princeton University Press.

Antipov, E. A. and Pokryshevskaya, E. B. (2015). Are buyers of apartments superstitious? evidence

from the russian real estate market. Judgment and Decision Making, 10(6):590–592.

Bae, S., Oh, H., Kim, H., Lee, C., and Oh, B. (2010). The impact of after-school programs on
educational equality and private tutoring expenses. Asia Pacific Education Review, 11(3):349–
361.

Banziger, G. (1983). Normalizing the paranormal: Short-term and long-term change in belief in the
paranormal among older learners during a short course. Teaching of Psychology, 10(4):212–214.
Bassok, D., Fitzpatrick, M., and Loeb, S. (2014). Does state preschool crowd-out private provision?
the impact of universal preschool on the childcare sector in oklahoma and georgia. Journal of
Urban Economics, 83:18 – 33.

Bastos, P. and Straume, O. R. (2016). Preschool education in brazil: Does public supply crowd out

private enrollment? World Development, 78:496 – 510.

Biswal, B. P. (1999). Private tutoring and public corruption: A cost-effective education system for

developing countries. The Developing Economies, 37(2):222–240.

103

Bound, J. and Turner, S. (2007). Cohort crowding: How resources affect collegiate attainment.

Journal of Public Economics, 91(5):877 – 899.

Brodaty, T., Gary-Bobo, R. J., and Prieto, A. (2014). Do risk aversion and wages explain educational

choices? Journal of Public Economics, 117:125 – 148.

Brown, J. R. and Finkelstein, A. (2008). The interaction of public and private insurance: Medicaid
and the long-term care insurance market. The American Economic Review, 98(3):1083–1102.
Byun, S.-Y. (2010). Does policy matter in shadow education spending? revisiting the effects of the

high school equalization policy in south korea. Asia Pacific Education Review, 11(1):83–96.

Bénabou, R. and Tirole, J. (2016). Mindful economics: The production, consumption, and value

of beliefs. Journal of Economic Perspectives, 30(3):141–64.

Chakrabarti, R. and Roy, J. (2016). Do charter schools crowd out private school enrollment?

evidence from michigan. Journal of Urban Economics, 91:88 – 103.

Chang, J.-J. (2011). Higher education and cramming for examinations: The roles of educational
quotas and low-tuition policies on social mobility. Journal of Social Sciences and Philosophy,
23(3):pp.311–339. (In Chinese).

Checchi, D., Fiorio, C. V., and Leonardi, M. (2014). Parents’ risk aversion and children’s educational

attainment. Labour Economics, 30:164–175.

Chen, J.-Y. (2016). Education reform digging a bigger hole: Cram school number grew 8.8 times

in the past 20 years. Common Wealth Magazine. (in Chinese).

Chin, J. M. and Wen, T. H. (2014). Better or worse: Examining the outcomes of the 20-year

educational reforms in taiwan. Journal of Education Research, 245:005–021.

Chiu, H. (1993). Change and variations of occult, witchcraft, religious behaviors. Proceedings
of the National Science Council, Republic of China, Part C: Humanities and Social Sciences,
3(2):125–143.

Chiu, H. (1999). Occultism revival and social change in taiwan. Taiwanese Journal of Sociology,

22:1–45. in Chinese.

Chiu, H. (2006). Study on the Popularization of Astrology in Taiwan, volume 1 of Religion,
Occultism, and Social Change, chapter 11, pages 293–324. Taipei: Kuei Kuan Books Ltd. in
Chinese.

Choi, H. and Choi, A. (2016). Regulating private tutoring consumption in korea: Lessons from

another failure. International Journal of Educational Development, 49:144 – 156.

Choi, J. and Cho, R. M. (2016). Evaluating the effects of governmental regulations on south korean

private cram schools. Asia Pacific Journal of Education, 36(4):599–621.

Chou, A. (2019). Fortunes and misfortunes of the dragon sons: Direct and cohort effects of

superstition on education attainment. Unpublished Manuscript.

104

Chou, C. P. (2003). The great experiment of Taiwanese education (1987-2003). Taipei: Psychology.

(In Chinese).

Chou, S.-Y., Liu, J.-T., Grossman, M., and Joyce, T. (2010). Parental Education and Child Health:
Evidence from a Natural Experiment in Taiwan. American Economic Journal: Applied Eco-
nomics, 2(1):33–61.

Chu, H.-L. (2015). Private Tutoring, Wealth Constraint and Higher Education. Pacific Economic

Review, 20(4):608–634.

Chung, R., Darrat, A. F., and Li, B. (2014). Superstitions and stock trading: some new evidence.

Journal of the Asia Pacific Economy, 19(4):527–538.

Clark, D. and Hsieh, C.-T. (2000). Schooling and labor market impact of the 1968 nine-year
education program in taiwan. Technical report, Princeton University, Woodrow Wilson School
of Public and International Affairs.

Connelly, R. and Gottschalk, P. (1995). The effect of cohort composition on human capital

accumulation across generations. Journal of Labor Economics, 13(1):155–176.

Damisch, L., Stoberock, B., and Mussweiler, T. (2010). Keep your fingers crossed! how superstition

improves performance. Psychological Science, 21(7):1014–1020.

Dang, H.-A. and Rogers, F. H. (2008). The growing phenomenon of private tutoring: Does it deepen
human capital, widen inequalities, or waste resources? The World Bank Research Observer,
23(2):161–200.

Das, J., Dercon, S., Habyarimana, J., Krishnan, P., Muralidharan, K., and Sundararaman, V. (2013).
School inputs, household substitution, and test scores. American Economic Journal: Applied
Economics, 5(2):29–57.

Do, Q.-T. and Phung, T. D. (2010). The importance of being wanted. American Economic Journal:

Applied Economics, 2(4):236–53.

Drummond, C. and Fischhoff, B. (2017). Individuals with greater science literacy and education
have more polarized beliefs on controversial science topics. Proceedings of the National Academy
of Sciences, 114(36):9587–9592.

Duflo, E. (2001). Schooling and labor market consequences of school construction in indonesia:

Evidence from an unusual policy experiment. American Economic Review, 91(4):795–813.

Executive Yuan (1996a). Executive Yuan Education Reform Committee final report. (In Chinese).

Executive Yuan (1996b). Executive Yuan Education Reform Committee final report. (In Chinese).
Fortin, N. M., Hill, A. J., and Huang, J. (2014). Superstition in the housing market. Economic

Inquiry, 52(3):974–993.

Foster, K. R. and Kokko, H. (2009). The evolution of superstitious and superstition-like behaviour.

Proceedings of the Royal Society of London B: Biological Sciences, 276(1654):31–37.

105

Francis, A. M. (2011). Sex ratios and the red dragon: using the chinese communist revolution
to explore the effect of the sex ratio on women and children in taiwan. Journal of Population
Economics, 24(3):813–837.

Fudenberg, D. and Levine, D. K. (2006). Superstition and rational learning. The American

Economic Review, 96(3):630–651.

Goodkind, D. M. (1991). Creating new traditions in modern chinese populations: Aiming for birth

in the year of the dragon. Population and Development Review, 17(4):663–686.

Goodkind, D. M. (1993). New zodiacal influences on chinese family formation: Taiwan, 1976.

Demography, 30(2):127–142.

Goodkind, D. M. (1995). The significance of demographic triviality: Minority status and zodiacal

fertility timing among chinese malaysians. Population Studies, 49(1):45–55.

Grech, V. (2015). The influence of the chinese zodiac on the male-to-female ratio at birth in hong

kong. Journal of the Chinese Medical Association, 78(5):287 – 291.

Gruber, J. and Simon, K. (2008). Crowd-out 10 years later: Have recent public insurance expansions

crowded out private health insurance? Journal of Health Economics, 27(2):201 – 217.

Guiso, L., Sapienza, P., and Zingales, L. (2006). Does culture affect economic outcomes? Journal

of Economic Perspectives, 20(2):23–48.

Guiso, L., Sapienza, P., and Zingales, L. (2016). Long-term persistence. Journal of the European

Economic Association, 14(6):1401–1436.

Hamersma, S. and Kim, M. (2013). Participation and crowd out: Assessing the effects of parental

medicaid expansions. Journal of Health Economics, 32(1):160 – 171.

Houtenville, A. J. and Conway, K. S. (2008). Parental effort, school resources, and student

achievement. The Journal of Human Resources, 43(2):437–453.

Huang, Y.-J. and Chen, C.-W. (2008). Academic cram schooling, academic performance, and
opportunity of entering public universities. Bulletin of Educational Research, 54(1):117–149.
(in Chinese).

Hwang, Y.-J. (2003). The construction and assessment of the new occupational prestige and
socioeconomic scores for taiwan: The indigenization of the social science and sociology of
education research. Bulletin of Educational Research, 49(4):1–31.

Iannaccone, L. R. (1998). Introduction to the economics of religion. Journal of Economic Literature,

36(3):1465–1495.

Iyer, S. (2016). The new economics of religion. Journal of Economic Literature, 54(2):395–441.
Jahoda, G. (1968). Scientific training and the persistence of traditional beliefs among west african

university students. Nature, 220(5174):1356.

106

Johnson, N. D. and Nye, J. V. (2011). Does fortune favor dragons? Journal of Economic Behavior

& Organization, 78(1):85–97.

Kabatek, J. and Ribar, D. C. (2016). Not your lucky day: Romantically and numer-
IZA working paper. Available at SSRN:

ically special wedding date divorce risks.
https://ssrn.com/abstract=2849753.

Kim, H. (2004). Analyzing the effects of the high school equalization policy and the college
entrance system on private tutoring expenditure in korea. KEDI Journal of Educational Policy,
1(1):5–24.

Kim, J.-H. and Chang, J. (2010). Do governmental regulations for cram schools decrease the
number of hours students spend on private tutoring? KEDI Journal of Educational Policy,
7(1):3–21.

Kim, S. (2016). Effects of government’s regulations on private education expenditure in korea.

Seoul Journal of Economics, 29(2).

Kingma, B. R. (1989). An accurate measurement of the crowd-out effect, income effect, and price

effect for charitable contributions. Journal of Political Economy, 97(5):1197–1207.

Kobakhidze, M. N. (2014). Corruption risks of private tutoring: case of georgia. Asia Pacific

Journal of Education, 34(4):455–475.

Kuan, P.-Y. and Lee, D.-Y. (2010). Effects of cram schooling on math performance in junior high:
A propensity score matching approach. Bulletin of Educational Research, 56(2):105–139. (In
Chinese).

Kuo, C.-Y. (2009). A study on the decision process of children’s english learning program. Master’s

thesis, National Chengchi University.

Kwo, O. and Bray, M. (2014). Understanding the nexus between mainstream schooling and private
supplementary tutoring: patterns and voices of hong kong secondary students. Asia Pacific
Journal of Education, 34(4):403–416.

Lau, Y. (2017). The dragon cohort of hong kong: Traditional beliefs, demographics, and education.

working paper.

Lee, A. (2013). 25 weirdest superstitions and rituals in sports. Bleacher Report. Website.

Lee, D.-Y. (2006). The effects of cram schooling on educational attainment in the secondary school
diversified admission plan. Journal of Education & Psychology, 29(3):489–516. (In Chinese).
Lee, J. and Paik, M. (2006). Sex preferences and fertility in south korea during the year of the

horse. Demography, 43(2):269–292.

Lee, M.-L., Liu, B.-C., and Wang, P. (1994). Education, human capital enhancement and eco-
nomic development: Comparison between korea and taiwan. Economics of Education Review,
13(4):275 – 288.

107

Leon, A. K. and Pfeifer, C. (2017). Religious activity, risk-taking preferences and financial be-
haviour: Empirical evidence from german survey data. Journal of Behavioral and Experimental
Economics, 69:99 – 107.

Levy, B. R., Chung, P. H., and Slade, M. D. (2011). Influence of valentine’s day and halloween on

birth timing. Social Science & Medicine, 73(8):1246 – 1248.

Li, Y. (1990). Further Analysis of Religious Problem, chapter 9, pages 135–152. Social Problem

in Taiwan. Taipei: Chu Liu Press.

Lin, C.-C. (2012). Why is it hard for you to not go to cram school? specialization and marketization

of cram school labor. (In Chinese).

Lin, C.-C. and Huang, T.-C. (2009). Cram culture. Journal of Social Sciences and Philosophy,

21(4):587–643. (In Chinese).

Lin, J.-S. and Tsai, T.-C. (2007). Research of the relation between shadow education and psycho-
logical well-being: Taking the students of the taitung junior high school for example. NTTU
Educational Research Journal, 18(1):73–106.

Lin, S.-H. (2015). How to create purple cow cram schools during the market decling birthrate.

Master’s thesis, Chaoyang University of Technology.

Lin, T.-S. (2001). The effect of family educational resources on tracking and educational status

attainment. The NCCU Journal of Sociology, 31:45–75. (In Chinese).

Lin, T.-S. and Chen, Y.-F. (2006). Cram school attendance and college entrance exam scores
of senior high school students in taiwan. Bulletin of Educational Research, 52(4):35–70. (In
Chinese).

Liu, J. (2006). The transition, effect, and stratification of cram schooling in taiwan. Bulletin of

Educational Research, 52(4):1–33. (In Chinese).

Liu, J. (2012). Does cram schooling matter? who goes to cram schools? evidence from taiwan.

International Journal of Educational Development, 32(1):46 – 52.

Liu, Y.-C. (2015). The years of dragon and tiger and human capital. Master’s thesis, National Chi

Nan University.

Lochner, L. (2011). Nonproduction benefits of education: Crime, health, and good citizenship.
In Hanushek, E. A., Machin, S., and Woessmann, L., editors, Handbook of The Economics of
Education, volume 4 of Handbook of the Economics of Education, pages 183 – 282. Elsevier.

Manza, L., Hilperts, K., Hindley, L., Marco, C., Santana, A., and Hawk, M. V. (2010). Exposure
to science is not enough: The influence of classroom experiences on belief in paranormal
phenomena. Teaching of Psychology, 37(3):165–171.

McLean, C. P. and Miller, N. A. (2010). Changes in critical thinking skills following a course on
science and pseudoscience: A quasi-experimental study. Teaching of Psychology, 37(2):85–90.

108

McMahon, W. W. (2004). The Social and External Benefits of Education.

Handbook on the Economics of Education, Chapters, chapter 6. Edward Elgar Publishing.

In International

Ministry of Education (1985). Republic of China Education Yearbook, 5th Edition. Ministry of

Education. (in Chinese).

Ministry of Education (1999). Investigative report on all forms of short-term supplementary classes
in taiwan region: Overview and statistics. Technical report, Ministry of Education. (in Chinese).

Ministry of Education (2001). Report on status of student learning and living in taiwan (high school

and below). Technical report, Ministry of Education. (in Chinese).

Ministry of Education (2010). Statistics on recent development in education:1994-1995 2009-2010

school years. (in Chinese).

Ministry of Education (2012). Republic of China Education Yearbook, 7th Edition. Ministry of

Education. (in Chinese).

Mocan, N. and Pogorelova, L. (2017). Compulsory schooling laws and formation of beliefs:
Education, religion and superstition. Journal of Economic Behavior & Organization, 142:509 –
539.

Mocan, N. H. and Yu, H. (2017). Can superstition create a self-fulfilling prophecy? school
outcomes of dragon children of china. Working Paper 23709, National Bureau of Economic
Research.

Musch, J. and Ehrenberg, K. (2002). Probability misjudgment, cognitive ability, and belief in the

paranormal. British Journal of Psychology, 93(2):169–177.

Nielsen, J. S., Bech, M., Christensen, K., Kiil, A., and Hvidt, N. C. (2017). Risk aversion and
religious behaviour: Analysis using a sample of danish twins. Economics & Human Biology,
26:21 – 29.

Nunn, N. and Sanchez de la Sierra, R. (2017). Why being wrong can be right: Magical warfare
technologies and the persistence of false beliefs. American Economic Review, 107(5):582–87.

Nye, J. and Xue, M. M. (2015). Raising dragons. GMU Working Paper in Economics.
Orenstein, A. (2002). Religion and paranormal belief. Journal for the Scientific Study of Religion,

41(2):301–311.

Oster, E. (2017). Unobservable selection and coefficient stability: Theory and evidence. Journal

of Business & Economic Statistics, pages 1–18.

Otis, L. P. and Alcock, J. E. (1982). Factors affecting extraordinary belief. The Journal of Social

Psychology, 118(1):77–85.

Parish, W. L. and Willis, R. J. (1993). Daughters, education, and family budgets taiwan experiences.

The Journal of Human Resources, 28(4):863–898.

109

Pasachoff, J. M., Cohen, R. J., and Pasachoff, N. W. (1970). Belief in the supernatural among

harvard and west african university students. Nature, 227(5261):971.

Pennycook, G., Cheyne, J. A., Seli, P., Koehler, D. J., and Fugelsang, J. A. (2012). Analytic

cognitive style predicts religious and paranormal belief. Cognition, 123(3):335 – 346.

Pop-Eleches, C. and Urquiola, M. (2013). Going to a better school: Effects and behavioral

responses. American Economic Review, 103(4):1289–1324.

Reiling, R. B. (2016). Does size matter? educational attainment and cohort size. Journal of Urban

Economics, 94:73 – 89.

Roe, C. A. (1999). Critical thinking and belief in the paranormal: A re-evaluation. British Journal

of Psychology, 90(1):85–98.

Rohlfs, C., Reed, A., and Yamada, H. (2010). Causal effects of sex preference on sex-blind and
sex-selective child avoidance and substitution across birth years: Evidence from the japanese
year of the fire horse. Journal of Development Economics, 92(1):82 – 95.

Rubinstein, M. A. (1991). The protestant community on modern Taiwan: Mission, seminary, and

church. ME Sharpe.

Ruiu, G. and Breschi, M. (2017). Superstitions surrounding the choice of wedding date in italy:
What has changed since the beginning of the economic development process to the present?
Journal of Population Research, 34(1):45–78.

Salter, C. A. and Routledge, L. M. (1971). Supernatural beliefs among graduate students at the

university of pennsylvania. Nature, 232(5308):278.

Senbet, D. and Huang, W.-C. (2012). Do dragons have better fate? revisited using the us data.

International Journal of Economics and Research, 3(5):101–118.

Shimizutani, S. and Yamada, H. (2014). Long-term consequences of birth in an ‘unlucky’ year:
evidence from japanese women born in 1966. Applied Economics Letters, 21(16):1174–1178.
Shuai, J.-J., Chen, H., and bin Chen, Y. (2017). Characteristics and behavior analysis of online
fortune telling users-a data mining approach. Survey Research-Method and Application, 37:97–
154.

Sim, N. (2015). Astronomics in action: The graduate earnings premium and the dragon effect in

singapore. Economic Inquiry, 53(2):922–939.

Siniver, E. and Yaniv, G. (2015). Kissing the mezuzah and cognitive performance: Is there an

observable benefit? Journal of Economic Behavior & Organization, 117:40 – 46.

Spohr, C. A. (2003). Formal schooling and workforce participation in a rapidly developing economy:
evidence from “compulsory” junior high school in taiwan. Journal of Development Economics,
70(2):291 – 327.

110

Stark, E. (2012). Enhancing and assessing critical thinking in a psychological research methods

course. Teaching of Psychology, 39(2):107–112.

Sun, C.-S. and Hwang, Y.-J. (1996). Shadow education, cultural capital and educational attainment.

The Taiwanese Journal of Sociology, 19:95–139. (In Chinese).

Sun, T.-h. (1989). A review of Fertility Control Policies in Taiwan Area, ROC, volume Social
Phenomena in Taiwan - An Analysis, pages p.275–310. Sun Yat-Sen Institute for Social Sciences
and Philosophy, Academia Sinica. In Chinese.

Ting, S.-T. (2005). A study on the emergence and transition of taipei’s cram schools for college

entrance exams. Master’s thesis, National Taiwan Normal University. (in Chinese).

Tsai, W.-J., Liu, J.-T., Chou, S.-Y., and Grossman, M. (2011). Intergeneration transfer of human
capital: Results from a natural experiment in taiwan. Working Paper 16876, National Bureau of
Economic Research.

Tsai, W.-J., Liu, J.-T., Chou, S.-Y., and Thornton, R. (2009). Does educational expansion encourage
female workforce participation? a study of the 1968 reform in taiwan. Economics of Education
Review, 28(6):750 – 758.

Tseng, C.-H. (2013). Analyzing regional differences of elementary cram school in buying behavior-

an example of taichung cram school. Master’s thesis, Feng Chia University.

Vere, J. P. (2008). Dragon children: Identifying the causal effect of the first child on female
labour supply with the chinese lunar calendar. Oxford Bulletin of Economics and statistics,
70(3):303–325.

Vyse, S. (2014). Believing in magic: The psychology of superstition. Oxford University Press.
Welch, F. (1979). Effects of cohort size on earnings: The baby boom babies’ financial bust. Journal

of Political Economy, 87(5):S65–S97.

Wilder, S. (2014). Effects of parental involvement on academic achievement: a meta-synthesis.

Educational Review, 66(3):377–397.

Wilson, J. A. (2018). Reducing pseudoscientific and paranormal beliefs in university students

through a course in science and critical thinking. Science & Education, 27(1):183–210.

Wolfradt, U., Oubaid, V., Straube, E. R., Bischoff, N., and Mischo, J. (1999). Thinking styles,
schizotypal traits and anomalous experiences. Personality and Individual Differences, 27(5):821
– 830.

Wong, K.-F. and Yung, L. (2005). Do dragons have better fate? Economic Inquiry, 43(3):689–697.
Yamada, H. (2013). Superstition effects versus cohort effects: is it bad luck to be born in the year

of the fire horse in japan? Review of Economics of the Household, 11(2):259–283.

Yi, C.-C. and Wu, C.-I. (2004). Teen life around the world: Introduction to teen life in taiwan. Teen

life around the world, pages 223–241.

111

Yin, C.-C., Tao, H.-L., and Hung, C.-Y. (2012). The effects of cram school on the performance in
the college entrance examination in taiwan. Jing Ji Lun Wen Cong Kan, 40(1):73. (In Chinese).
Yip, P. S., Lee, J., and Cheung, Y. (2002). The influence of the chinese zodiac on fertility in hong

kong {SAR}. Social Science & Medicine, 55(10):1803 – 1812.

Yuan, C. and Zhang, L. (2015). Public education spending and private substitution in urban china.

Journal of Development Economics, 115:124 – 139.

Za’Rour, G. I. (1972). Superstitions among certain groups of lebanese arab students in beirut.

Journal of Cross-Cultural Psychology, 3(3):273–282.

Zhan, S. (2014). The private tutoring industry in taiwan: government policies and their implemen-

tation. Asia Pacific Journal of Education, 34(4):492–504.

112