課綱與國族認同： 探討1997 年國中教科書改革的影響

國立臺灣大學社會科學院經濟學系 碩士論文

Department of Economics College of Social Science National Taiwan University

Master Thesis

課綱與國族認同： 探討 1997 年國中教科書改革的影響 Curriculum and National Identity: Evidence from the

1997 Textbook Reform in Taiwan

陳煒林 Wei-Lin Chen

指導教授： 林明仁 博士 Advisor: Ming-Jen Lin, Ph.D.

中華民國 106 年 8 月

August, 2017

謝辭

這是一篇關於「認識台灣」這套教科書的文章，研究這個題目的兩年間，我才開始認識台灣。

謝謝林明仁老師鼓勵我將這個題目做成碩士論文，更在我研究陷入低潮時，鼓勵我 繼續。謝謝中研院的楊子霆老師，花了無數個下午與我討論這篇文章使用的計量方法和 結果詮釋。謝謝研究室 643 的大家，邱錦洋、沈暉智、林建勳、吳東懋、吳愷頡，跟大家一起 打屁酸人討論碩論的那段時光，非常開心。謝謝 Prof. Eik Swee, Prof. Martin Huber, Prof. Bei Qin，在研究的初期給予我意見。謝謝江淳芳老師、樊家忠老師、駱明慶老師、Prof.

Patrick Dejarnette, Prof. Gerard Roland, Prof. Hans-Joachim Voth, Prof. Noam Yutchman 在這篇文章成形後，給予我非常多寶貴的建議。最後，謝謝我的家人，謝謝爸爸 媽媽和姊姊一直相信我，讓我能夠在自己選擇的路上，堅持下去。

這是一篇歪打正著的文章，研究前，我對政治偏好絲毫沒有理解。對於教育體系如何影響 人們各式各樣的偏好，亦絲毫沒有想像。然而，現在一旦認真回想義務教育的九年間，那是一 個如何封閉且嚴格遵守秩序的社會，不禁讚嘆起，教育真是一部潛在的，嚴謹的洗腦機器。當 然，即使真有洗腦這回事，在人們長大後，出教育體系進社會，每天接觸（各式各樣的）資 訊，教育體系的洗腦還會有效嗎？被教育體系洗腦的人對社會中其他人的影響是什麼？這篇 文章的最後，提出的便是這樣的問題。

關於洗腦這個詞彙，我想補充解釋。通常洗腦帶有負面意涵，但當我使用這個詞的時候，

我心裡想的只是一種「偏好的建構」，是中性的。而如果認真回想，你可能會發現，幾乎所有的 偏好，某種程度上都是被洗腦出來的，可能是家人、老師、同儕、偉人，也可能是優質電視節目、

十大好書、百萬粉絲團。研究洗腦機器的過程中，我愈來愈珍惜言論自由，雖然我們不一定能 選擇，但言論自由至少讓我們被洗腦之後，有很多可能的長相。

中文摘要

教材能夠改變學生的國族認同嗎？我們檢視了一次引進大量台灣相關內容的國中教科書 改革，利用迴歸不連續法去除社會趨勢影響以及世代效果後，我們發現於國中時期讀新教科 書的學生更有可能持有較強的台灣人認同。這個現象在就讀學術體系的學生身上更為明顯。此 外，藉由分析學生故鄉的族群比例，我們發現如果學生原生故鄉的台灣人認同較低，則教科 書的效果較明顯。上述結果隱含兩個教科書影響學生國族認同的管道—背誦教材與鄰里間的社 會化。學生年歲較長後，是否曾讀過新教科書變得無法預測學生的國族認同差異。這份結果來 自以下事實：未曾讀過新教科書的學生長大後，他們的台灣人認同上升趨勢，比起讀過新教 科書的學生更為快速。究竟這份快速的上升趨勢來自於同儕互動中產生的新教科書外溢效果，

或是因為整體社會的台灣人認同上升，我們需要更多研究才能回答這個問題。

關鍵詞: 教科書改革、國族認同、長期效果

Abstract

Could education content casually affect students’ national identity? We ex- ploit the sharp junior high school textbook reform which introduced large amount of Taiwan-related materials, using regression discontinuity design to tease out the society trend and cohort effect. We find that students exposed to the new textbook are more likely to hold stronger Taiwanese identity. The effect is larger for students who entered academic education track and students endowed with less Taiwan-oriented ideology approximated by the ethnic distribution of the hometown, implying mechanisms of memorizing and socialization with neigh- borhoods. As the students aged, whether they study the new textbook or not could not predict significant national identity differences. We find that this in- significance emerges from the steeper increasing trend of Taiwanese identity in students not exposed to the new textbook. Whether the steeper increasing trend comes from spill-over effect of the new textbook through peer interaction or from the exposure to the also increasing Taiwanese identity trend of the whole society requires further examination.

Keywords: Textbook Reform, National Identity, Long Term Effect

Contents

謝辭 i

中文摘要 ii

英文摘要 iii

1 Introduction 1

2 Background and Textbook Analysis 6

2.1 The curriculum reform of

Knowing Taiwan series

. . . 6

2.2 Textbook Analysis . . . 7

3 Data 12 4 Main Results 15 4.1 Graphical Analysis . . . 15

4.2 Regression Discontinuity Design . . . 19

4.3 Assumptions of Regression Discontinuity Design . . . 21

4.4 Regression Analysis of the Effect of Knowing Taiwan series . . . 24

4.5 Robustness Check . . . 29

4.5.1 Bandwidth Choice . . . 29

4.5.2 Falsification Test. . . 31

5 Subgroup Analysis 33 5.1 Heterogeneity: Education Track . . . 33

5.2 Heterogeneity: Hometown Ethnicity Distribution . . . 40

7 Conclusion 52

A National Identity Distribution and Sample Selection 56

B Observable Characteristics 56

C Bandwidth Sensitivity 60

D Falsification Test 62

E Hoklo Ethnicity Distribution in Taiwan 63

List of Figures

1 Overall National Identity Trend in the Society . . . 17

2 National Identity Trend in Different Education Cohort . . . 17

3 Estimated Textbook Effects by Education Cohort . . . 18

4 Number of Observations at Month Level . . . 22

5 Short Run National Identity Variation at Birth Quarter Level . . . 25

6 Bandwidth Sensitivity. . . 30

7 National Identity Trend in Different Education Track . . . 35

8 Heterogeneity Textbook Effect in Education Track: Cohort Analysis 35 9 Heterogeneity Textbook Effect in Education Track: RD visulization 38 10 National Identity Trend in High/Low Hoklo Ratio Areas in 1990s. 43 11 National Identity Trend in High/Low Hoklo Ratio Areas . . . 43

12 Heterogeneity Textbook Effect in High/Low Hoklo Ratio Areas: Cohort Analysis . . . 44

13 Heterogeneity Textbook Effect in High/Low Hoklo Ratio Areas: RD visulization . . . 45

14 Estimated Textbook Effect by Education Cohort in the Long Run . 49 15 Long Run National Identity Variation at Birth Quarter Level. . . . 49

16 National Identity Variation by Treatment Status . . . 50

18 Observable Discontinuity . . . 59

19 Bandwidth Sensitivity with Demographic Controls . . . 60

20 National Identity Variation at Birth Month Level with Different Bandwidths. . . 61

21 National Identity Variation with Fake Birth Month Cut-Off . . . . 62

22 The Map with High/Low Hoklo Peole Proportion at Town Level . 63

List of Tables

1 Observables Continuity . . . 23

2 Estimated effects of the textbook reform in the Short Run . . . 28

3 Falsification Test of Fake Textbook Reform . . . 32

4 Heterogeneity Textbook Effect: Education Track . . . 39

5 Heterogeneity Textbook Effect: Hometown Ethnicity Distribution 46 6 Estimated effects of the textbook reform in the Long Run . . . 51

7 National Identity Distribution in Main Regression Sample . . . 56

8 Samples Dropped by Sample Selection . . . 56

9 Observables Continuity with Polynomials of Second Order . . . . 57

1 Introduction

The more homogeneity the people, the more easiness the management within a nation. Based on this intuition, state leaders have incentives to use education system as an instrument to cultivate national identity, an essential step toward nation building. Empirical evidence has shown that in recent 150 years, the mass education investment by the government appears in response to military threats, when governments acutely requires patriotic people as a preparation for future wars (Aghion et al., 2014). The causal effect underlying the intuition and the transmission mechanism of the effect of education on national identity formation, however, lacks detail scrutinization.

In this paper we investigate the causal effect of education content on national identity formation by examining the junior high school textbook reform in Tai- wan. In 1997, the National Institute for Compilation and Translation (Guoli

Bianyi Guan) published the Knowing Taiwan series social subjects textbook. The

history volume, in particular, adopted a new perspective of Taiwan’s history and provided abundant Taiwan-related content, both of which were absent from previous textbooks. We examine whether studying the

Knowing Taiwan series

causally induce students to hold stronger Taiwanese identity relative to weaker Taiwanese identity.

Using repeated cross section surveys (Taiwan Social Change Survey), we mea- sure the national identity by self reported identity groups and look into the iden- tity variation through both simple cohort comparison and regression discontinu- ity design (RDD) specifications. We find that students studying

Knowing Taiwan

series are on average 18 percentage points more likely to report themselves as

Taiwanese than Both (Taiwanese and Chinese) and Chinese when students were around their twenties, with a control group mean 0.6.

National identity trend in the society and cohort effects pose challenges to the identification of the education content effect. Specifically, the two effects interact with each other in the sense that students in different cohorts experience differ- ent society development. Further, the introduction of

Knowing Taiwan series may

coincide with other social and political events which influence people’s national identity. We first compare the national identity in education year cohort level in different survey years to rule out the effect of society atmosphere in different survey years. The national identity variation between the first cohort exposed to the new textbook and the last cohort studying the old textbook is regarded as the textbook effect. To further close out confounding factors that would af- fect students’ national identity

within an education year, we exploit the identity

variation in birth month level. The education system in Taiwan forces children who were born after 1 September to enter next education cohort; namely, chil- dren born in August would enter the education system earlier than those born in September for a year. People born in September 1984 would therefore be the first month cohort studying the

Knowing Taiwan series, while those born in Au-

gust 1984 would study the old textbooks. The running variable in our regression discontinuity design is thus birth month and the treatment cut-off is September 1984. Thanks to the shifting national identity trend in the society in 1990s in Taiwan, the textbook reform is not exogenous to the society development , but our regression discontinuity design is still valid since the reform is exogenous to the birth month decision of parents giving birth in 1984. Moreover, since people

born near two sides of the cut-off experience very similar society development, the RDD helps us rule out the effect of the society development along our respon- dents’ grown-up.

We examine the possible channel of the

Knowing Taiwan textbook effect. The

effect only appears in students who enter the academic education track when we divide observations by their final education level into vocational and academic tracks. Since students in academic track generally put more efforts on studying academic materials, this result implies that memorizing the education content is a possible channel. Another heterogeneity we investigate into is the prior famil- iarity of Taiwan-oriented ideology. We approximate the prior Taiwan-oriented ideology with hometown ethnic distribution. Four major ethnicities lived in Tai- wan: Hoklo, Mainlanders, Hakka, and Aborigines. In order generations, Hoklo people typically are much more likely to hold Taiwanese identity than other eth- nicities. We divide one’s living environment by the high/low level of Hoklo peo- ple proportion at township level. We find large textbook effect in people who lived in towns with less Hoklo people, whereas no effect is found in people who lived in towns with more Hoklo people. The result is in align with the prediction from “belief-based (bayesian learning) models”: people with weaker prior belief are more likely to be persuaded by new information (DellaVigna and Gentzkow, 2010).

We also study the long run effect of education content and find that people who studied the old textbook hold the Taiwanese identity as strong as those who studied new textbook. This comes from the fact that the increase of the propor- tion of former group of students who also hold Taiwanese identity is more rapid

comparing to students of the latter group in the long run.

Our paper is relevant to the political preference formation and the culture transmission literature. Previous literature has shown that political regime and media would affect people’s policy and party preference (Alesina and Fuchs-

Schundeln, 2007;DellaVigna and Kaplan, 2007) Recently, some researchers ex- amine the relationship between education policy and political institution pref- erence and ethnic identity. Clots-Figueras and Masella(2013) andFouka(2015) study the language usage promotion and prohibition in education system respec- tively. Clots-Figueras and Masella (2013) find that the change from single lan- guage (Spanish) to bilingual (Spanish and Catalan) education in Catalan makes students have stronger Catalan feelings, and this further leads to changes in political party preference in elections. Fouka (2015), on the other hand, find a backlash effect of German language prohibition in the U.S. education system during 1917 and 1923. She documents that children of German immigrants who experienced the language prohibition are more likely to marry with Germans, choose more German flavor first name for their child, and less likely to volun- teer in World War II. She ascribes the backlash effect to the substitute property between parents’ investment in children’s identity and school’s influence on chil- dren’s identity.

Our paper is most relevant toCantoni et al.(2015). They conducted a survey in Peking university to study the effect of the new curricula in China which was adopted in different years in different provinces ranging from 2004 to 2010. They find that the new curricula effectively shape students’ ideology toward the goal of the curricula; specifically, students exposed to the new curricula exhibit more

trust in government, more skepticism toward unconstrained democracy and un- constrained free markets. They do not, however, find curricular effect on ethnic identity. The goal of the curricula is to make students feel more integrated as a united identity. .

We add up evidence in the direction that other than language usage in edu- cation, education content as an education instrument or more broadly a kind of authoritative media, would influence the national identity. Based on the finding ofCantoni et al. (2015), we contribute to the literature by showing that the text- book effect exhibit heterogeneity in the dimension of the effort put into studying the education content and the prior familiarity with the education content. More importantly, the textbook effect might not be permanent and we thus have to be cautious about interpretation when we observe the effect of education content policy in the short run. We note that the effect we uncover may include all subse- quent behaviors such as the change of media consumption induced by studying new textbooks. We thus interpret the effect as a total impact derived from behav- iors entailed by studying

Knowing Taiwan textbooks.

The paper proceeds as follows: in section 2 we discuss the background of the textbook reform and analyze the difference between the old society subject textbooks and

Knowing Taiwan series. We describe the data used in this paper

in section 3. Main results including graphical analysis and regression analysis based on regression discontinuity design are presented in section 4. Then we explored the potential mechanisms through subgroup analysis in section5. We explore long run effects in section6. Finally section7concludes.

2 Background and Textbook Analysis

2.1 The curriculum reform of Knowing Taiwan series

In 1994, the Ministry of Education of Republic of China announced the new cur- riculum for junior high school social subjects. The major change lay in the design of the first year content of social subjects. In earlier textbooks, Taiwan-related content accounted for only small portions in the text and was scattered through- out different volumes, The new curriculum of social subjects, including history, geography, and society, aimed to not only provide much more Taiwan-related documentation but also different angles on the history and social development of Taiwan. After three years of writing and editing, the National Institute for Com- pilation and Translation published the new social subjects textbook, known as the

Knowing Taiwan series, and students entering junior high school in Septem-

ber 1997 were expected to utilize the series.

The textbook reform aroused fierce debate among political parties on whether the textbook is “appropriate.” Political parties at that time could be divided into two groups. Parties including Kuomintang (KMT) and New Party reckoned the regime as “the successor of China”, while Democratic Progressive Party (DPP) ad- vocated “Taiwan Independence” and considered KMT government, which ruled Taiwan since 1945, as a foreign regime. The discussions at the time about whether the

Knowing Taiwan history volume should be adopted centered around three

perspectives in the textbook: the “relationship between Taiwan and Japan in his- tory,” the “relationship between Taiwan and China in history” and the “judg- ment of contemporary political events and politicians” (Wang,2001). According

to Wang (2001), in just two months from June to August in 1997, 341 articles (five articles every day on average) about

Knowing Taiwan series appeared in the

four main newspapers. Facing severe doubt on the content of the textbook, the editors of

Knowing Taiwan series made minor revisions and the textbooks were

still adopted in junior high school across Taiwan in September 1997.

The reform was comprehensive. Students across Taiwan who entered ju- nior high school between September 1997 and September 2000 would study the

Knowing Taiwan series. Though the major change of the new textbook lay in the

ones for the first grade of junior high school. Textbooks for the second and the third year were also adjusted. The senior high school/vocational school entering examinations for students born after Sep 1984 and the earlier education cohort are therefore based on different textbooks of all three years. This makes sure that earlier education cohorts are not exposed to the

Knowing Taiwan series since they

do not have to memorize the materials in the

Knowing Taiwan series to prepare

for the examination.¹

2.2 Textbook Analysis

The curriculum reform aroused politicians’ attention because it manifested the stark differences between the two imagined nationalities–the Chinese conscious- ness and the Taiwanese consciousness (Liu et al.,2005;Wang,2001). In particu- lar, the new history textbooks divert the angle from “China-oriented" in the ear- lier textbooks to “Taiwan-oriented". This is significantly reflected in the nested

1We define education cohort as students entering compulsory education system in the same year and label them with the year they enter the junior high school. For example, the 1997 education cohort are those who firstly studiedKnowing Taiwan series. They were born between September 1984 and August 1985. Likewise, the education year refers to the year in education system. The 1997 education year refers to the time period from Sep 1997 to Aug 1998.

variation of the two dimensions between the two versions of the textbooks that being the amount of content of Taiwan history and the context of Taiwan history.² The latter could be further divided into two parts: the chronological narrative or- der and the usage of following terms “our country,” “Taiwan,” and “China.”

Under old curriculum, junior high school students studied the National His- tory for a year and a half and World History for another year and a half, whereas under new curriculum, students studied history about Taiwan in the first year (the

Knowing Taiwan history volume), National History in the second year, and

World history in the third year. In other words, the content of the National His- tory and the World History in the old version is condensed in order to add new materials about Taiwan.

In terms of time, teachers utilizing new textbooks might spend twelve times more on Taiwan history than before. Under old curriculum, only one chapter and a section discuss Taiwan ³–the third section entitled “The rebellion of Koxinga against Qing Dynasty and the development of Taiwan” in fifteenth chapter in the second volume, and the twenty-fifth chapter entitled “The achievement and vision of a base for revival” in the third volume. Assuming that teachers spend the same time on each chapter and section in a volume, we approximate that teachers would spend less than fifth of a semester on Taiwan history. In contrast, the

Knowing Taiwan history volume was designed to cover two semesters. The Knowing Taiwan history volume has 116 pages about Taiwan. The older textbooks

contained only sixteen pages.

2Taiwan history is referred to as history about Taiwan “island,” instead of the one about Re- public of China

3In the old textbook series, National History consisted of twenty five chapters in three vol- umes, each volume for a semester

The notable increment of Taiwan history showed that the editors put more emphasis on it. As the third point of editors’ preface in the

Knowing Taiwan

writes:

This book aims to introduce students to

know the history about how an- cestors of different ethnic groups made developments in Taiwan. By learn-

ing this, students will cultivate cooperation spirit, patriotic feelings, and worldwide horizons. Also, this will augment students’ under- standing of

Taiwanese cultural assets, and make them appreciate and

treasure it.⁴

The intention of making students acquainted with Taiwan development is never seen in the old version. We contrast this by recording the editors’ preface in the National History in the old version below:

The National History describes

the evolution of Chinese nationality, the change of the territory, and the development of politics, society, economics, and culture. In particular, it stresses the long history and the blending

of the culture of nationality in order to strengthen the patriotic feel- ings and cooperation spirits, and to know the nation’s traditions, the position and the responsibility of national people.

In the new version of the National History textbooks, approximately the same words appear in the editors’ preface except that “the evolution of Chinese na- tionality, the change of the territory” part is deleted.

Obviously, the editors of the

Knowing Taiwan history volume did not incor-

porate the

development in Taiwan and the Taiwanese cultural assets in the context

4Emphasis of the quote from the textbooks in this section is added by the author

of

the evolution of Chinese nationality, the change of the territory, and the develop- ment of politics, society, economics, and culture. It is this change of perspective that

highlights the battle between “Taiwanese consciousness”and “Chinese conscious- ness.”

That editors in the new version viewed Taiwan history as a distinct entity different from China history is as mentioned before denoted with the narrative order and the usage of “our country,” “Taiwan,” and “China.”

In their first grade, junior high students who studied the old textbooks started learning the history of “our country” with the statement that the earliest hu- man beings lived in “our country” is

Homo erectus pekinensis in Paleolithic age.

The “common ancestor” of

Chinese nationality is Huang Di. The first dynasty of

“our county” is Xia dynasty. The history of “our country” then developed un- der sequential dynasties, from Xia to Qin, Tang all the way to Qing. Between the dynasty of Ming and Qing in this straightforward development line rests the first appearance of “Taiwan,” identified by editors as a base for Koxinga to fight against Qing regime.

Despite just a verb, it is worth noted that Koxinga is written as “recovering”

Taiwan from the Dutch. The usage of the verb explicitly demonstrates the ideol- ogy behind the old textbook, showing that the editor reckons the ruling power of the Dutch in 17^th century as a “foreign regime.” Simultaneously, this implicitly claimed Taiwan as the territory of “our country”before the ruling of the Dutch.

At the end of the line is another appearance of Taiwan, recognized as the revival base under the development of Republic of China. In the new curricu- lum, however, this straight line of “our country” is decomposed into two separate

strands: Taiwan and China.

The word “our country” is rarely used in the

Knowing Taiwan history vol-

ume and the National history textbooks under the new curriculum. “Taiwan”

and “China” filled the blanks. More precisely, “our country” only appears in de- scription about Republic of China. Following the divided usage of terms, Taiwan history stands out not as part of China history, buy as an individual entity in the

Knowing Taiwan history textbook, and the growing background of Taiwan is

introduced:

the history about how ancestors of different ethnic groups made develop- ments in Taiwan.

Developments in Taiwan are stated in order. Cultures in prehistory era and the aborigines are stated as the first group of people living in Taiwan. What fol- lows are the trade activities among pirates from China and Japan in 16^thcentury, and the governing of the Dutch and the Spain. Ruling of Koxinga, Qing dynasty, Japan, and finally the Republic of China are all documented as developments of different groups. On the other hand, the National history textbook under the new curriculum maintains most of the content and the narrative order of the old textbook.

The battle between Taiwanese consciousness and Chinese consciousness would never emerge in the old curriculum since “Taiwan” is almost neglected and “our country” is the only imagined nation. Only when “Taiwan” and “China” are ex- plicitly separated do readers have the chance to distinguish concepts between them. The

Knowing Taiwan history volume may cultivate Taiwanese identity in

two ways. First, priming effect arises from students reading the word “Taiwan”

more often. Second, the distinction made by describing Taiwan history and China

history separately may provide students different materials to instill in the two imagined groups and hence helps them differentiate between Taiwanese and Chi- nese.

3 Data

The TSCS and the measure of national identity.

The Taiwan Social Change Survey (TSCS) provides repeated cross-section rep- resentative observations across Taiwan. It holds two surveys of different topics every year and covers observations aged above 18, each with sample size about 1800 to 2200. The first education cohort exposed to the

Knowing Taiwan reform,

who were born after Sep 1984, is firstly surveyed in 2003. To balance the observa- tions before and after the reform in regression analysis, we hence include surveys held after 2003 which contain the national identity questions and enough back- ground information, including 2003, 2004, 2005, 2010, 2012, 2013 and 2014 waves.

The national identity is measured by the following question:

• In our society, somebody call themselves Taiwanese, somebody call them- selves Chinese, and somebody call themselves Both. Do you consider your- self as Taiwanese, Chinese, or both?

We create the dummy variable

Identity by assigning one to respondents answer-

ing Taiwanese and zero to those answering Chinese and Both. We drop the ob- servations who answers "Other." In our main regression sample, only 3.84% of respondents answer "Chinese," indicating that in this generation, very few people

hold pure Chinese identity. We hence adopt the above categorization to differen- tiate the inclination of holding stronger or lesser Taiwanese identity.⁵

Since the measurement of the national identity is based on the self-reported response, the natural question appears: could this measurement truly reflect re- spondents’ national identity? One possible explanation of the change of

Identity

(if it is observed) is that previous students were afraid of responding himself as Taiwanese. The new textbooks provide not the Taiwanese identity, but the mes- sage that saying oneself is Taiwanese is not a taboo anymore. We provide two counterarguments of this explanation. First, the simple mean of

Identity of the

control group in our main analysis sample is 0.6. When over half of one’s peer would identify themselves as Taiwanese, it is hard to believe that Taiwanese iden- tity is a taboo for him. Second, the change of

Identity should be spot in different

subgroups under this explanation, but in section5we would find that this is not the case.

The Advantage of TSCS.

Two features of TSCS make it suitable for our analysis. First, TSCS records the birth year

and birth month of respondents. This allows us to identify cor-

rect education cohort and makes feasible the regression discontinuity design on month level, letting us look into finer variation between birth months instead of yearly cohorts. The second feature of TSCS is the representative sampling and rich demographic variables. The representative sampling allows us to first esti- mate the treatment effect in the whole population and then further investigate the mechanism through subgroup analysis.

We include the respondent’s gender, their parents’ education level and their

5We document detailed proportion of respondents answering different options in AppendixA

parents’ ethnicity in the analysis. The rich demographic variables help us con- struct different subgroups. In particular, we put emphasis on the education track and the ethnic group distribution in respondents’ hometown. For the self- education subgroup, we split the sample into academic track and vocational track. We construct

Edu-Academic dummy variable by assigning one to the re-

spondents whose final education attainment is senior high school or university (academic track); zero to those with junior high school, senior vocational school, college and tech university (vocational track).

The proxy of hometown ethnicity distribution.

One of the questions in TSCS reads: “Where do you live longest before 15 years old?” The response is in township (zip code) level. We regard the response as where the respondents live when in junior high school, and regard it as his hometown.

The hometown information is combined with ethnicity township level data to approximate how many Hoklo people surrounded the respondents in his daily life when in junior high school. The ethnicity data is from National Hakka Pop- ulation Basic Information Survey Research conducted in 2004 with sample size 37693, about 100 in a town. We use the response of the question: “You con- sider yourself as?” Six options include: (1) Taiwan Hakka, (2) Mainland Hakka, (3) Hoklo, (4) Mainlanders, (5) Aborigines, (6) Foreigners. The respondents could only pick one answer to this question. The proportion of people answering Hoklo, for example, would be regarded as the proportion of Hoklo people in the town.

We then estimate the population median of Hoklo people ratio in Taiwan using Hoklo people proportion in each town and the population of the town in 2004 as

weights (population data from the Ministry of Interior). The median is 77.1%.⁶ Finally we construct the dummy variable

Home-Hoklo-Ratio by assigning one to

the respondents whose hometown’s hoklo people proportion is

less than 77.1%,

and assign zero otherwise. If the respondent lived in a town with

Home-Hoklo- Ratio as one, he is less likely to meet a Hoklo person in his daily life comparing to

an average person in Taiwan.

Sample Selection Rule.

We drop the respondents who reports himself born outside Taiwan and the respondents who reports himself live longest outside Taiwan before 15, since we could not make sure that these observations entered junior high school in Taiwan and hence were exposed to the textbook reform. We also drop respondents who did not report their parents’ ethnicity and education level, which are required demographic information we need for the regression analysis. Finally, we drop respondents whose answer of the national identity question is “Other.” These selection rules drop 19 (4.5%) of the main regression sample (the main regression sample size is 417).⁷ The main results in this paper are not influenced by the sample selection.

4 Main Results

4.1 Graphical Analysis

Our objective is to investigate the relationship between studying

Knowing Taiwan

series and holding stronger Taiwanese Identity. The design of the textbook reform

6The mean using the same data and weight is 73.2%

7We provide detail numbers of dropped observations in each step in AppendixAfor all sam- ples we use in this paper.

and the repeated cross section survey data bring us two challenges: the swiftly altering national identity trend in the society and the potential factors affecting national identity formation within an education cohort.

Younger cohorts are more likely to be influenced by more recent society atmo- sphere, which is more Taiwanese-oriented. An obvious increase of the proportion of people responding himself as Taiwanese in the last two decades is spot in Fig- ure1, which depicts the national identity trend derived from observations across all ages and measured by National Cheng Chi University Election Study Center.⁸ To control for this channel and the survey year fixed effect, we stratify the obser- vations in education-cohort-survey-year level. We include samples with educa- tion cohorts ranging from 1992 to 1998 and then investigate the national identity variation at education cohort level. The textbook reform changes the treatment status (studying certain textbooks) on an education cohort basis. Figure2plots the simple mean of

Identity in each cell. The sample size in each cell is on average

41.

Two noticeable facts could be seen in Figure2. First, we observe a jump with roughly 0.2 magnitude between the 1996 and 1997 education cohort, the first cohort studying

Knowing Taiwan series. To examine if the jump is statistically

significant, we regress

Identity on education cohort and survey year dummies us-

ing 1996 education cohort as the reference group. We plot the point estimates and 90% confidence intervals of education cohort dummies in Figure 3. The point estimates of previous education cohorts are not significantly different from 1996 education cohort. The 1997 education cohort dummy is significantly dif-

8We do not use the data from National Cheng Chi University Election Study Center due to data accessibility and the lack of birth moth records in the data.

0204060

1990 1995 2000 2005 2010 2015

Survey Year

Taiwanese Both Taiwanese and Chinese

Chinese Non Response

Figure 1: Overall National Identity Trend in the Society

Notes: Cheng Chi University Election Study Center conducts telephone survey on adults aged 20 above every year. The sample size ranges from 1,200 to 34,000; on average it is about 13000. The question used to measure the national identity adopted by Cheng Chi University Election Study Center is the same as the one in TSCS.

.5.55.6.65.7.75Taiwanese Identity

1993 1994 1995 1996 1997 1998

Education Cohort

Survey Year 2004 Survey Year 2005

Figure 2: National Identity Trend in Different Education Cohort

Notes: Data is collapsed at education cohort-survey year level from 2004 and 2005 wave of TSCS.

Dots represent the simple mean ofIdentity at each cell. We do not include the cell mean derived from 2003 survey year since the sample size for the treatment cohort is too small (less than 20).

For the same reason, we drop the mean of 1998 education cohort surveyed in 2004.

-0.100.000.100.200.30Taiwanese Identity

1993 1994 1995 1996 1997 1998

Education Cohort

Point estimate 95% CI

Figure 3: Estimated Textbook Effects by Education Cohort

Notes: We pool 2003, 2004, 2005 TSCS waves. Point estimates derived from regression of Identity on survey year dummies , hometown dummies and education cohort dummies (1996 as reference group). Standard errors are clustered at birth month level.

ferent from the 1996 education cohort, with magnitude 17.3 percentage points.

The insignificance of 1998 education cohort dummy may arise from the small sample size of the 1998 education cohort. This is the first evidence showing that the treatment, studying

Knowing Taiwan series, causes students to hold stronger

Taiwanese identity.

The second important observation from Fig 2 is that the survey year fixed effect could not capture the upward time trend of national identity fully. This could be seen by observing that the difference between two lines from two sur- vey years are not the same across education cohorts. To avoid this problem, we further restrict our sample to fewer education cohorts in the rest of our analysis:

observations who were born between Sep 1982 and Aug 1986 (four education co- horts, two studied

Knowing Taiwan series). We first analyze these cohorts when

they are relatively young, aging from18 to 23 and surveyed from 2003 to 2005

(henceforth short run). To examine if the textbook effect is persistent, we exam- ine the same education cohorts surveyed from 2010 to 2014 , with age range 24 to 32 (henceforth long run).

4.2 Regression Discontinuity Design

The society trend problem prompts us to compare observations who experience almost the same society development along their grown-up. Regression Discon- tinuity Design (RDD) enables such task. At first glance, we should conduct RDD on education cohort basis since the treatment status varied at that level. Nev- ertheless, people in the same education cohort may experience different events possibly altering their national identity, for instance, the voting history.

Elections in Taiwan are generally held in December, January, and March. The eligible age is 20. In some elections, people born in the first half of education cohort are eligible, while people in the second half are not. Students who were the first cohort exposed to the textbook reform offers one example. The sixth legislative election was held on Dec 11, 2004, splitting the education cohort into two groups: people who have voting right (born before Dec 11, 1984) and people who do not (born after Dec 11, 1984). Students who were the last cohort studying textbook of the old version provides another example. The election splitting their vote history is the president election in Mar 20, 2004. Since politicians in Taiwan debate fiercely on the national identity issue in elections, the different "first vote"

experience may in turn affect people’s national identity formation. Bearing in mind such differences embedded with respondents within an education cohort,

we adopt the standard RDD specification at the birth month level:

Identity

_ijt

= α

₀

+ α

₁

T extbookExposure

_i

+ f ( m ) + γX

_i

+ η

_j

+ δt + ϵ

_ijt (1)

where

Identity

_ijt indicates the dummy variable defined in the data section of individual

i, with hometown j interviewed at time t. The variable T extbookExposure

indicates whether the respondent was exposed to the textbook reform and takes 1 if the respondent reports himself born after Sep 1984, 0 otherwise. We use birth month as our running variable and recenter the birth month at Sep 1984, the first month cohort affected by the reform. Following standard practice, we include

f ( m )

, a smooth function of birth month, to control for the impact of so- ciety trend. In our case,

f ( m )

is first order or second order polynomials fully interacted with

TextbookExposure. Standard errors are clustered at birth month

(m) level. The month level RDD provides a stronger identification than the co- hort comparison in Fig3since people around the cut-off should experience very similar society atmosphere. The only disparity around the cut-off is the entering year of compulsory education, which subsequently includes our treatment, the studying of

Knowing Taiwan series in junior high school.

To increase the precision of our estimates, In all specifications, we include survey year fixed effect (δt) to single out overall effect of the society trend in each survey year. we also include hometown fixed effect (ηj), which help us control for regional factors possibly influencing the national identity formation, e.g., the local support for certain political party. In some specifications, we include de- mographic variables (X_i) which might influence the national identity formation,

The core interest is

α

₁, which indicates whether studying

Knowing Taiwan series

would causally make one more likely to hold stronger Taiwanese identity.

4.3 Assumptions of Regression Discontinuity Design

The validity of the RDD rests in the assumption that no precise control exist at the cut-off. Specifically, if the “quarter of birth” is endogenous in Taiwan, the assumptions of the regression discontinuity design may be violated. Fan et al.

(2014) shows that contrast to United States, high socioeconomic status mothers have similar birth pattern with low socioeconomic status mothers, assuring us that birth seasonality is not a concern in our paper. We then check the smoothness of the number of observations and conduct standard practices of regression of observables to validate our empirical strategy.

The Smoothness of Observation Frequency.

We use survey data in this paper, so we could not detect if the population around the cut-off exhibit discontinuity. Instead, we plot the number of obser- vations in each month cell to see if the survey design give more weight to one side of the birth month cut-off. As Figure 4shows, discontinuity is absent. The discontinuity at the fourth month after the cut-off arises due to the fact that TSCS only surveyed adults aged 18 above, so people who were born in 1985 would not be surveyed until 2004.

Smoothness of Observables Mean.

To test if the distributions of the observables are smooth around the cut-off, we specifically run regression on observables related to national identity using specifications as equation1, with first order birth month polynomial. Regression

05101520

-20 -10 0 10 20

Birth Month

Figure 4: Number of Observations at Month Level

Notes: I pool data from 2003, 2004, 2005 TSCS and include education cohorts from 1995 to 1998.

Each dot represent the number of observations in that birth month. The running variable (birth month) is recentered at Sep 1984, the first month cohort expose to the textbook reform.

results are shown in Table1. The counterpart graphs are in AppendixB.

Home-Hoklo-Ratio exhibits significant discontinuity with size 0.146 in Table

1, showing that it is significantly more likely to observe a respondent who had lived in a town with less Hoklo people on the right hand side of the cut-off. We argue that this discontinuity comes from sampling error. Since we have few samples in each birth month cell, it is likely that due to sampling error the survey data we used display different characteristics from the population . One evidence sup- porting our argument is that when we run the same regressions in the long run sample, no significant jump is found. We include the

Home-Hoklo-Ratio variable

into some of the specifications to lessen the potential bias. For other observables, we do not find discontinuities in Table1.

T able 1: Observ a bles C on tin uity V ARIABLES g ender F ather Mother F ather Mother F ather Mother Self Self Hometown Ethnicity Ethnicity Ed u Lev el Ed u Lev el Ed u Y ears Ed u Y ears Ed u Lev el Ed u Y ears Hokl o Ra tio

TextbookExposure

0.0398 -9.30e-05 0.0316 -0.0747 -0.0265 -0.0583 -0.0242 -0.0316 -0.0641 0.146* (0.0713) (0.0777) (0.0569) (0.0897) (0.0837) (0.605) (0.447) (0.0767) (0.251) (0.0736)

Month

0.000844 0.00222 0.000127 0.00110 -0.000559 0.0268 0.0225 0.00316 -0.0275* -0.00417 (0.00369) (0.00401) (0.00403) (0.00411) (0.00351) (0.0326) (0.0224) (0.00307) (0.0152) (0.00403)

TextbookExposure

-0.00649 -0.00732 -0.00472 0.00641 0.00247 -0.0246 -0.0207 -0.00724 0.0376 -0.00242 ×

Month

(0.00664) (0.00643) (0.00517) (0.00770) (0.00692) (0.0485) (0.0384) (0.00609) (0.0247) (0.00538) C onstan t 0.459*** 0.260*** 0.174*** 0.542*** 0.453*** 10.76*** 9.955*** 0.564*** 13.90*** 0.379*** (0.0582) (0.0519) (0.0459) (0.0636) (0.0382) (0.527) (0.238) (0.0538) (0.171) (0.0647) Observ ations 417 417 414 417 417 417 417 417 417 417 R -squared 0.002 0.004 0.003 0.004 0.001 0.004 0.003 0.003 0.012 0.007

Notes:Wepooldatafrom2003,2004,and2005TSCSwavesandincludeeducationcohortsfrom1995to1998..Werunregressionwithspecification1.Firstorder polynomialisincludedandwedonotincludedemographics.TextbookExposureis1ifthebirthmonthoftherespondentisafterSep1984,0otherwise,indicatingif onestudiestheKnowingTaiwanseries.MonthisthebirthmonthrecenteredatSep1984.Standarderrorsareclusteredatbirthmonthlevelinparentheses***p<0.01, **p<0.05,*p<0.1 Theconstructionoftheindependentvariablesarethefollowing:Gender:femalebeing1,malebeing0.Father/Motherethnicity:Hoklofathers/mothersbeing 0,otherwise1.Father/Mothereducationlevel:father/motherwithnoeducation,elementaryschool,andjuniorhighschooleducationlevelbeing0,otherwise1. Father/Mother/Selfeducationyearsisdiscretevariablerangingfrom0to16.Selfeducationlevel:respondentswithfinaleducationattainmentasseniorhighschool anduniversitybeing1,otherwise0.HometownHokloRatioisthedummyvariableHome-Hoklo-Ratio,seeconstructioninsection3Theproxyofhometownethnicity distribution.

4.4 Regression Analysis of the E ffect of Knowing Taiwan series

We focus on the Taiwanese identity variation measured by the dummy variable

Identity around the birth month cut-off. We show smoothed figures to examine

discontinuities around cut-off and present the estimates derived from equation 1. Unlike standard RDD, plotting simple mean of

Identity in each birth month

cell could not tell apart the survey year fixed effect and the treatment effect, lead- ing to misinterpretation. Also, since the sample size of each birth month cell is too small (9 observations on average), all of the following graphs are in birth quarter level to visualize more informative variation. Aiming to control for the survey year fixed effect, in Figure5, we followShigeoka(2014) and plot the cell mean at birth quarter level of the residual of

Identity after being regressed on

survey year dummies. The lines in Figure 5 represent fitted regressions of the cell mean dots using first order polynomials interacted with the dummy variable

T extbookExposure.

Although we do not observe a distinct jump of the dots just around the cut-off, the fitted line suggests that the discontinuity of

Identity is roughly 20% around

the cut-off. By comparing the near four dots (an education cohort) on both sides of the cut-off, we notice that the proportion of people holding stronger Taiwanese identity significantly increases. It should be noted that the interpretation of the last two dots on the right hand side of the cut-off, which represent the latter half of 1998 education cohort, requires further caution, since the observations in each dot is only 11.

Table 2shows the regression results of estimation of specification 1. Home- town and survey year fixed effects are included in all regressions and we find

-.2-.10.1.2

-10 -5 0 5 10

Birth Quarter

Figure 5: Short Run National Identity Variation at Birth Quarter Level

Notes: I pool data from 2003, 2004, 2005 TSCS and include education cohorts from 1995 to 1998.

We first regressIdentity on survey year dummies and then collapse the residuals at birth quarter level to derive the dots. Fitted lines are from regression of the dots on a first order polynomial of birth quarter interacted withKnowingT aiwan dummy variable.

qualitatively similar estimates of the treatment effect across different specifica- tions. Column [1] reports the regression results with first order polynomials con- trolling the time trend. The estimates of

T extbookExposure is 0.18 and signif-

icant, signifying that studying

Knowing Taiwan series on average increase one’s

probability of reporting himself as Taiwanese by 0.18, a much more stronger Tai- wanese identity comparing to 0.6, the mean of

Identity on the left side of the

cut-off (the control group).⁹

In column [2], we add the demographic variables to increase the precision of the estimates and lessen the potential bias due to observables’ discontinuity.

The demographic variables include gender, parents’ education level, parents’ eth- nicity and the

Home-Hoklo-Ratio. Parents’ ethnicity and education level capture

the family influence on respondents’ national identity. Empirical literature has

9The reference mean 0.6 is the simple mean ofIdentity including samples born between Sep 1983 and Aug 1984. For samples born between Mar 1984 and Aug 1984, the mean is 0.65.

shown that parents would invest in strengthening offspring’s national identity to be the same as their own identity (Fouka,2015). In older generation in Taiwan, ethnicity largely determines one’s national identity; for example, Mainlanders are much more likely to hold Chinese identity than other ethnic groups.¹⁰

The magnitude and the significance of the treatment effect in column [2] are similar to column [1]. Column [3] and [4] reports the results with second order polynomials as controls. Estimate of the treatment effect in column [4] is a little bit smaller than the one in column [3]. We note that once taking into account of the size of the standard error of the

T extbookExposure estimates in column [3]

and [4], the estimates are similar statistically speaking. Further, the fact that the second order polynomial term added in column [3] and [4] are not significant and the small differences of the estimates across columns assures us that the estimates are not sensitive to the polynomial specification.

The magnitude of the estimates are also consistent with Figure3and Figure5.

We thus confirm that studying

Knowing Taiwan series do significantly strengthen

one’s Taiwanese identity, and our confident and conservative measure of the treat- ment effect is 18 percentage points.

Persuasion Rate.

We provide the persuasion rate calculated by the formula used in media eco- nomics (DellaVigna and Gentzkow,2010):

100×

y

_t− yc

e

_t− ec × 1 1− yc

,

10We define Mainlanders as people with at least one of their parents being Mainlanders; the rest, Others. Combining 2003, 2004, 2005 TSCS waves and keep the samples born between 1950 and 1964, we find that proportion of Mainlanders reporting himself as Taiwanese, Chinese, Both are 25.9%, 10.6%, 61.4% respectively, while the counterpart of Others are 67.2%, 2.7%, 28.7%

where

e

i denotes the share of the group

i receiving the message (the education

content in our case),

y

_i denotes the share of group

i adopting the behavior (re-

porting stronger Taiwanese identity in our case). The subscript

t and c means

treatment and control group. The persuasion rate measures to what degree treat- ment persuade people into adopting the behavior, scaled by the share of people who receives messages and the space for control group “to be persuaded” (1−yc).

In our case, since all students born after September 1984 are exposed to the new textbook,

e

_t− ec

=

1− 0

=

1. The persuasion rate reported in the first column in Table2is calculated by: 100×^0.179₁ ×_1−0.608¹

=

45.6. This 45% persuasion rate is quite high comparing to the persuasion rate found in media economics literature, which are barely higher than 20%. Our estimates, however, is in align with the persuasion rate found inCantoni et al.(2015). In their paper, more than a quar- ter of the persuasion rate is higher than 20%, and the highest one is 41%. The high persuasion rate is not that surprising after taking into account of the degree of the exposure: students have to at least study the

Knowing Taiwan series for a

year and they also have to memorize the materials for the high school admission examinations. The exposure is much greater than specific TV or radio program.

Table 2: Estimated effects of the textbook reform in the Short Run Dep Variable: Taiwanese Identity

[1] [2] [3] [4]

T extbookExposure

0.179** 0.176** 0.243** 0.213*

(0.0757) (0.0770) (0.0982) (0.110)

Month

0.000314 0.000225 0.00687 0.00917

(0.00491) (0.00503) (0.0172) (0.0179)

T extbookExposure

-0.00614 -0.00520 -0.0473** -0.0422**

×Month (0.00755) (0.00748) (0.0191) (0.0197)

Month

² 0.000256 0.000352

(0.000771) (0.000798)

T extbookExposure

0.00144 0.00102

×Month² (0.00107) (0.00107)

Constant 0.603*** 0.488*** 0.619*** 0.517***

(0.0929) (0.110) (0.107) (0.127)

Control Group Mean 0.608 0.608 0.608 0.608

Persuasion Rate 45.6 44.8 61.9 54.3

Observations 417 417 417 417

R-squared 0.110 0.160 0.118 0.165

Survey Year & Hometown FE Yes Yes Yes Yes

Demographic No Yes No Yes

Notes: We pool data from 2003, 2004, and 2005 TSCS waves. We run regression with speci- fication1. All columns include hometown fixed effect and survey year fixed effect. Column [1] and [2] include first-order polynomials. Column [2] further include demographic con- trol including gender, parents’ education level (category variable: elementary school, ju- nior high school, senior high school, senior vocational school, college, university, military), parents ethnicity(category variable: Hoklo, Mainlanders, Hakka, Other), andHome-Hoklo- Ratio dummy variable. Column [3] and [4] include second order polynomials. Column [3] does not include demographics while column [4] does. T extbookExposure is 1 if the birth month of the respondent is after Sep 1984, 0 otherwise, indicating if one studies the Knowing Taiwan series. Month is the birth month recentered at Sep 1984. Month²means square term ofMonth. Standard errors are clustered at birth month level in parentheses

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

4.5 Robustness Check

We validate the robustness of the main results in two ways. First we discuss the sensitivity of the results to the choice of birth month bandwidth, then we do the falsification test by changing the birth month cut-off to previous years.

4.5.1 Bandwidth Choice

The results in Table 2 are qualitatively similar to a wide range of bandwidths.

In Figure6, we plot the point estimates and the corresponding 95% confidence intervals derived from regression specifying linear polynomials controlling for trend (same specification as in column [1] in Table2) with bandwidths ranging from two education cohorts to one education cohort on each side of the cut-off.¹¹ The magnitudes of the point estimate are similar as we narrow down the birth month window, showing that the results in Table2are not sensitive to the band- width choice.

The change of the size and significance of the estimates appears once we re- strict our sample to less than 30 months. Investigating into the graphs present- ing

Identity variation at month level (Figure

20 in Appendix C), we find that the decrease of the magnitude is due to the fact that birth month cohort of−15 and−14 exhibit much more higher Identity mean (survey year fixed effects con- trolled) than later month cohorts. The left-hand side fitted line of

Identity mean

are therefore steeper when we exclude−14 and −15 birth month cohorts, leading to a decrease of the size of the jump implied by the difference of the fitted line around the cut-off. It should be noted that the fitted line of the right hand side

11We provide the results controlling for demographic variables in Fig19in AppendixC

0.000.100.200.300.40

24 26 28 30 32 34 36 38 40 42 44 46 48

bandwidth

Point estimate 95% CI

Figure 6: Bandwidth Sensitivity

Notes: We run regressions as column [1] in Table2with different bandwidths. The bandwidth 48 means we include 24 birth months on each side of the cut-off, i.e., two education cohorts. The solid line represent the point estimates of theT extbookExposure dummy variable and the dotted line represents the corresponding 95% confidence interval derived from standard errors clustered at birth-month level.

(the “treatment group”) barely changes when we confine samples to narrower bandwidths.

The fact that the estimate of the treatment effect decreases when we investi- gate the behavior within two education cohorts leads to concerns about the over- estimate of the size of the treatment effect. We argue that this is not a threat to the interpretation of our main results in Table2. Observing the 1996 education cohort in Panel C in Figure20, we could still see that most of the points are below 0 (survey year fixed effects controlled), while most of the dots on the right hand side (the treated group) locate between 0.1 and 0.2; the difference fits our main regression results.

4.5.2 Falsification Test

Another possible explanation of the jump other than the

Knowing Taiwan reform

is mental age effect: people who were born on the exact left hand side of the birth month cut-off are more mentally mature than those on the right hand side since they enter school system earlier and thus, at any given time, they may have more working experience and experience dealing with others (they leave the educa- tion system earlier). Being that case, we should observe similar jumps in every yearly cohorts. To examine this speculation, we run regression of equation1with moving fake reform windows.

We reckon 1996, 1995 and 1994 as fake textbook reform education year and thus treat September 1983, 1982 and1981 as fake birth month cut-off. We then replicate the results in Table2for each fake textbook reform using the same TSCS waves in 2003, 2004, and 2005. Note that we only include two education cohorts on each side of the fake birth month cut-off to make falsification results compa- rable to our main results.

Table 3 shows the results of the falsification regressions. Signs of

FakeText-

bookExposure are generally inconsistent across different columns within a fake

reform and nearly no significant result is found except the fake reform in 1996 education year. The magnitudes also largely fluctuate across columns within a panel. Furthermore, we plot the quarterly mean of

Identity (survey year fixed

effects controlled) for each fake reform in AppendixD. No obvious discontinuity around the fake birth month cut-off is displayed. Both the regression results and the visual evidence suggest that mental age effect could not explain the jump we see in Figure5and the significant treatment effects in Table2.

Table 3: Falsification Test of Fake Textbook Reform Dep. Variable:

Identity

[1] [2] [3] [4]

Panel A: September 1983 as Birth Month Cut-Off

FakeT extbookExposure

-0.0545 -0.0449 -0.249** -0.204*

(0.0713) (0.0804) (0.106) (0.117)

Observations 479 479 479 479

R-squared 0.073 0.135 0.083 0.143

Panel B: September 1982 as Birth Month Cut-Off

FakeT extbookExposure

0.0261 -0.00500 0.186 0.187 (0.0986) (0.0994) (0.139) (0.131)

Observations 500 500 500 500

R-squared 0.067 0.131 0.074 0.140

Panel C: September 1981 as Birth Month Cut-Off

FakeT extbookExposure

-0.0116 0.0150 -0.0201 -0.0589 (0.0714) (0.0736) (0.0950) (0.0899)

Observations 509 509 509 509

R-squared 0.058 0.172 0.060 0.173

Linear Trend Yes No Yes No

Quadratic Trend No Yes No Yes

Survey FE Yes Yes Yes Yes

Hometown FE Yes Yes Yes Yes

Demographic No Yes No Yes

Notes: We pool 2003, 2004, 2005 TSCS data. Panel A includes 1994-1997 education cohorts; Panel B, 1993-1996; Panel C, 1992-1995. We define FakeTextbookExposure as respondents born after Sep 1983, 1982 and 1981.

Specifications in each column are the same as in corresponding columns in Table2. Standard errors are clustered at birth month level in parentheses.

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

5 Subgroup Analysis

In this section we investigate the heterogeneity of textbook effect along two di- mensions: education track and the ethnicity distribution of one’s hometown.

We present four piece of evidence in each subgroup analysis, from raw data examination to the regression result of RDD specification, as in section4.1 and section4.4. First we look at the proportion of respondents holding stronger Tai- wanese identity in each subgroup in different survey years. We then run a sim- ple regression of

Identity to see if different subgroup exhibit different Taiwanese

identity pattern on year cohort basis. The specification is as follows:

Identity

_ikjt

= β

₀

+ β

₁

Z

_i

+ β

_2k

Z

_i× ζk

+ ζ

_k

+ η

_j

+ δ

_t

+ ϵ

_ijt, (2)

where

Z

_i is a dummy variable indicating the subgroup division (Edu− Academic or

Home

− Hoklo − Ratio); ζk are education cohort dummies;

η

_j and

δ

_tare home- town and survey year fixed effects. Our interest focuses on β2k. After year cohort analysis, we zoom in and conduct regression discontinuity analysis within each subgroup.

5.1 Heterogeneity: Education Track

One dimension to break up the potential treatment effect heterogeneity is to look into the degree of exposure to the new textbooks. Students who pay more efforts on studying the textbooks are associated with higher treatment intensity in the sense that they may memorize more Taiwan-related texts. The ideal proxy of the intensity is the grade of the social subject in the admission examination of senior

high school/senior vocational school.¹² We do not observe this in the TSCS data, but we could distinguish the rough high/low level of efforts students devoted to general academic subjects in junior high school by their choice of the education track.

After compulsory education, students in Taiwan are divided into two educa- tional tracks: the academic and the vocational tracks. The choice of the track is endogenous and would be highly correlated with the efforts student put in study- ing when in junior high school. For students who had been motivated to pursue more academic knowledge, they would put a great deal of efforts into studying the textbooks so that they could be selected into better senior high schools. On the other hand, common wisdom suggests that parents in Taiwan would encour- age students who lack motivation but are adept at obtaining excellent grades (for example, memorize the material more quickly than an average person) to opt for academic instead of vocational track. Consequently, the division of the education track imply the latent exposure to the

Knowing Taiwan series.

We categorize observations into two groups. Observations with senior high school or university as their final education attainment are labelled with aca- demic track (Edu− Academic equals to 1). Other observations, with junior high school, senior vocational school, college and tech university as final education attainment, are labelled with vocational track (Edu− Academic equals to 0).

In Figure 7, we plot the simple mean of

Identity at education cohort level in

2004 and 2005 survey year. The trend of

Identity before the textbook reform is

not parallel, but we observe that in both 2004 and 2005 waves, students in aca-

12Junior high school graduates, no matter which education track they wish to proceed, took the same national examination and use the scores from that examination to apply for senior high or

.2.4.6.81Taiwanese Identity

1995 1996 1997 1998

Education Cohort

Academic Vocational

2004

.2.4.6.81Taiwanese Identity

1995 1996 1997 1998

Education Cohort

Academic Vocational

2005

Figure 7: National Identity Trend in Different Education Track

Notes: Data is collapsed at education-cohort-survey year level for observations in two education tracks from 2004 and 2005 waves of TSCS. Dots represent the simple mean ofIdentity at each cell.

-0.200.000.200.400.60Taiwanese Identity

1995 1996 1997 1998

Education Cohort

Point estimate 95% CI

Figure 8: Heterogeneity Textbook Effect in Education Track: Cohort Analysis

Notes: We pool 2003, 2004 and 2005 TSCS waves. Point estimates and confidence intervals of β2k

are derived from regression with specification2reckoning 1996 education cohort as reference group. Standard errors are clustered at birth month level.