ABSTRACT
Title of Dissertation: ESSAYS ON RACIAL AND
GENDER DISCRIMINATION IN
DEVELOPING COUNTRIES
Tomohiro Hara
Doctor of Philosophy, 2022
Dissertation Directed by: Professor Sebastian Galiani
Department of Economics
This dissertation investigates racial, ethnic, and gender issues mainly in developing coun-
tries. The first two chapters examine the relationship between mass-media and identity, and the
third chapter examines the labor market formation and gender wage difference in an emerging
modern industry.
The first chapter, titled, ?Building the Rainbow Nation through Mass Media: Television,
Cultural Diversity, and National Unity in post-Apartheid South Africa? examines whether na-
tional mass-media promotes cultural diversity, while preserving national unity. Rainbow Nation
was a national slogan to make cultural differences a source of strength in South Africa. Televi-
sion broadcasts in South Africa promote both cultural diversity and national unity, aligning with
that national slogan. I investigate the effects of television broadcasts on two outcomeslanguage
choices in elementary schools, which are associated with cultural identities, and voting shares of
political parties that espouse opposite views on national unity. I digitize locations and features
of television transmitters, leverage topographical and time variations that determine television
broadcast coverage over time, and, using a difference-in-differences approach, estimate the ef-
fects of broadcasts on the two outcomes listed above. I find that exposure to television broadcasts
increases the use of native languages in schools, which contributes to cultural diversity. Such
exposure also increases voting shares for the political parties that promote national unity.
The second chapter, ?Radio and Racism during Apartheid?, investigates the impacts of ra-
dio programs, which opposed racial segregation policies, on amplification of racism in South
Africa during Apartheid. During Apartheid, radio broadcasts were highly censored, and infor-
mation against racial segregation, such as international sanctions and riots in Black townships
were not broadcasted. However, due to political reasons, people could listen to uncensored ra-
dio in a part of the country. I digitize locations of radio transmitters and estimate its impacts
on voting behaviors which were highly associated with preferences on segregation policies for
the White population. Difference-in-differences analyses show that exposure to anti-Apartheid
radio reduced supports for a right wing-party which promoted racial segregation and increased
supports for a left-wing party which opposed the segregation.
The third chapter titled ?Labor Market Evolution in an Emerging Industry: Japanese Cotton
Spinning Industry before 1900? is a coauthored paper with Serguey Braguinsky. We utilize an
unusually rich historical data set on firms, establishments, workers and wages in the Japanese
cotton spinning industry in late 19th to explicitly link the evolution of the industry to labor market
outcomes in the long run. In the early periods, firms recruited workers with low wages that
resembles Harris and Todaro?s dual economy model. We then show that soon after experiencing
new entries, competitions to acquire skilled workers changed the nature of labor market toward
competition, and wages for both gender increased.
ESSAYS ON RACIAL AND GENDER DISCRIMINATION
IN DEVELOPING COUNTRIES
by
Tomohiro Hara
Dissertation submitted to the Faculty of the Graduate School of the
University of Maryland, College Park in partial fulfillment
of the requirements for the degree of
Doctor of Philosophy
2022
Advisory Committee:
Professor Sebastian Galiani, Chair/Advisor
Associate Professor Serguey Braguinsky
Associate Professor Jing Cai
Associate Professor Jessica Goldberg
Associate Professor Ethan Kaplan
Acknowledgments
I am grateful to my advisor, Professor Sebastian Galiani, for his dedicated advices and sug-
gestions during my career at Maryland. I benefitted and learnt a lot from working as an research
assistant and coauthoring with Professor Serguey Braguinsky. I also thank Professor Ethan Ka-
plan, Professor Jessica Goldeberg, and Professor Jing Cai for their advising, as well as readily
agreed to serve on my thesis committee. Interactions with professors and graduate students at
Maryland through seminars were also invaluable experiences. Among many, I especially benefit-
ted from Development Research Group organized by Professor Galiani and Professor Goldberg,
and Race Research Group organized by Professor Kaplan.
I also thank my wife, parents, brother and grand parents for their sincere supports and
encouragements. We have had difficulties to see each other face to face, especially during the
pandemic of COVID-19. Yet, I never felt that I am working alone, because of their warm encour-
agements.
Finally, I acknowledge financial assistance to support my academic life from Japan Stu-
dent Services Organization (JASSO) and the PINE Foundation, and research supports from the
University of Maryland.
ii
Table of Contents
Acknowledgements ii
Table of Contents iii
List of Tables vi
List of Figures viii
Chapter 1: Building the ?Rainbow Nation? through Mass Media: Television, Cultural
Diversity, and National Unity in Post-Apartheid South Africa 1
1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1.1 Related literature . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.2 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.2.1 South African political environment . . . . . . . . . . . . . . . . . . . . 9
1.2.2 Background of South African television broadcasts . . . . . . . . . . . . 12
1.2.3 Background of language choices in South African schools . . . . . . . . 17
1.3 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
1.3.1 Data: Television coverage . . . . . . . . . . . . . . . . . . . . . . . . . 19
1.3.2 Data: Annual Schools Survey . . . . . . . . . . . . . . . . . . . . . . . . 21
1.3.3 Data: Political attitudes . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
1.4 Identification strategies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
1.5 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
1.5.1 Results: languages of instruction . . . . . . . . . . . . . . . . . . . . . . 29
1.5.2 Results: Political attitudes . . . . . . . . . . . . . . . . . . . . . . . . . 40
1.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
Chapter 2: Radio and Racism During Apartheid 48
2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
2.2 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
2.2.1 Brief overview of Apartheid in South Africa . . . . . . . . . . . . . . . . 53
2.2.2 Radios under Apartheid . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
2.2.3 Politics in South Africa during Apartheid . . . . . . . . . . . . . . . . . 57
2.3 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
2.3.1 Exposure to radio . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
2.3.2 Voting outcomes during Apartheid . . . . . . . . . . . . . . . . . . . . . 62
2.4 Identification strategy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
iii
2.5 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
2.5.1 Testing identification assumptions . . . . . . . . . . . . . . . . . . . . . 67
2.5.2 Robustness check and discussion on potential concerns . . . . . . . . . . 69
2.6 Concluding remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
Chapter 3: Labor Market Evolution in an Emerging Industry: Japanese Cotton Spin-
ning Industry before 1900 75
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
3.2 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
3.2.1 Japan?s Economy before 1900 and the Cotton Spinning Industry . . . . . 80
3.3 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
3.3.1 Geographical distance . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
3.4 Labor market institutions and descriptive analyses . . . . . . . . . . . . . . . . . 86
3.4.1 The Beginning Stage . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
3.4.2 Large-scale entry and industry expansion . . . . . . . . . . . . . . . . . 91
3.5 Regression analyses . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
3.5.1 Model and Identification . . . . . . . . . . . . . . . . . . . . . . . . . . 97
3.5.2 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
3.5.3 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
3.6 Conclusion and further agenda . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
Appendix A: Appendices for Chapter 1 112
A.1 Subjective attitudes on national unity and ethnicity . . . . . . . . . . . . . . . . 112
A.2 Additional figures and tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
A.2.1 Figures: the construction of transmitters for each year . . . . . . . . . . . 117
A.2.2 Figures: Different assumptions on clustering standard errors . . . . . . . 118
A.2.3 Figures: Effect of TV access on language choices: standard two-way
fixed effect . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
A.2.4 Figure: Effect of TV access on language choices: adding covariant from
Census . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
A.2.5 Figures: Effect of TV access on language choices with different cutoffs . 121
A.2.6 Figures: Effect of TV access on voting outcomes . . . . . . . . . . . . . 123
A.2.7 Tables: difference-in-differences . . . . . . . . . . . . . . . . . . . . . . 123
Appendix B: Appendices for Chapter 2 128
B.1 Additional background information . . . . . . . . . . . . . . . . . . . . . . . . . 128
B.1.1 Detailed background of the Capital Radio 604 . . . . . . . . . . . . . . . 128
B.1.2 Other radio in independent Bantustan . . . . . . . . . . . . . . . . . . . 130
B.1.3 Potential radio exposure from other countries . . . . . . . . . . . . . . . 131
B.2 Engineering model of calculating AM radio propagation . . . . . . . . . . . . . 132
B.3 Additional Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135
B.3.1 Tables: results with different signal strength as treatment definitions . . . 135
B.3.2 Tables: results using data after 1970 . . . . . . . . . . . . . . . . . . . . 138
B.3.3 Tables: others . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141
B.4 Additional figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141
iv
Appendix C: Appendices for Chapter 3 144
C.1 Simple model of monopsony and peer effects . . . . . . . . . . . . . . . . . . . 156
C.2 Discussions on estimating dynamic panel model . . . . . . . . . . . . . . . . . . 159
Bibliography 161
v
List of Tables
1.1 Summary of historical background . . . . . . . . . . . . . . . . . . . . . . . . . 12
1.2 Summary statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
1.3 Event study: Effect of SABC exposure on language choices in elementary schools 31
1.4 Effect of SABC exposure on language choice . . . . . . . . . . . . . . . . . . . 39
1.5 Event Study: Effect of SABC exposure on language choices in elementary schools 43
2.1 Difference in differences: Effects of AM radio on voting behavior Treated if the
signal strength is higher than 60dBu . . . . . . . . . . . . . . . . . . . . . . . . 66
2.2 Difference in differences:Placebo test Treated if the signal strength is higher than
60dBu . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
2.3 Test of parallel trend assumption: Treated if the signal strength is higher than 60dBu 69
2.4 Robustness check: main results with state-year FE Treated if the signal strength
is higher than 60dBu . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
2.5 Difference in differences: only with open districts . . . . . . . . . . . . . . . . . 73
3.1 The lowest and the highest wages in 1896 by concentration . . . . . . . . . . . . 94
3.2 Summary statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
3.3 Peer effects from other firms in 20 km: two months lag regressors . . . . . . . . 105
3.4 Instrumental variable estimations on the effect of peers? capacity expansion . . . 106
A.1 Effects of television on subjective attitudes . . . . . . . . . . . . . . . . . . . . . 115
A.2 Difference-in-differences: Effect of SABC exposure on language choice . . . . . 126
A.3 Effect of SABC exposure on voting outcomes by political ward . . . . . . . . . . 127
B.1 Difference in differences: Effects of AM radio on voting behavior Treated if the
signal strength is higher than 50dBu . . . . . . . . . . . . . . . . . . . . . . . . 135
B.2 Difference in differences: Effects of AM radio on voting behavior Treated if the
signal strength is higher than 55dBu . . . . . . . . . . . . . . . . . . . . . . . . 136
B.3 Difference in differences: Effects of AM radio on voting behavior Treated if the
signal strength is higher than 65dBu . . . . . . . . . . . . . . . . . . . . . . . . 137
B.4 Difference in differences: Effects of AM radio on voting behavior Treated if the
signal strength is higher than 50dBu . . . . . . . . . . . . . . . . . . . . . . . . 138
B.5 Difference in differences: Effects of AM radio on voting behavior Treated if the
signal strength is higher than 55dBu . . . . . . . . . . . . . . . . . . . . . . . . 139
B.6 Difference in differences: Effects of AM radio on voting behavior Treated if the
signal strength is higher than 65dBu . . . . . . . . . . . . . . . . . . . . . . . . 140
B.7 Difference in differences: replication of Wilse-Samson (2013) . . . . . . . . . . 141
vi
C.1 First stage of the 2SLS regression dependent variable: number of male and female
workers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147
C.2 Peer effects from other firms in 20 km: six months lag regressors . . . . . . . . . 148
C.3 Peer effects from other firms in 20 km: twelve months lag regressors . . . . . . . 149
C.4 Peer effects from other firms in 50 km: two months lag regressors . . . . . . . . 150
C.5 Peer effects from other firms in 50 km: six months lag regressors . . . . . . . . . 151
C.6 Peer effects from other firms in 50 km: twelve months lag regressors . . . . . . . 152
C.7 Peers? capacity expansion on own wages: different specifications . . . . . . . . . 152
C.8 Peer effects from other firms in 20 km: two months lag regressors with different
criteria of the sample . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153
C.9 Peer effects from other firms in 20 km: two months lag regressors with imputing
missing variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 154
C.10 Peer effects from other firms in 20 km: two months lag regressors with imputing
missing variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155
vii
List of Figures
1.1 Votes by party after the end of Apartheid . . . . . . . . . . . . . . . . . . . . . . 11
1.2 Population density by municipality as of 2001 . . . . . . . . . . . . . . . . . . . 14
1.3 Coverage of SABC1, as of 2001 . . . . . . . . . . . . . . . . . . . . . . . . . . 16
1.4 Additional coverage of SABC1 between 2002-2010 . . . . . . . . . . . . . . . . 17
1.5 Construction of new transmitters for SABC . . . . . . . . . . . . . . . . . . . . 20
1.6 Number of schools by treatment cohort . . . . . . . . . . . . . . . . . . . . . . . 23
1.7 Distribution of schools by the proportion of students learning in native languages 24
1.8 Number of voting wards by treatment cohort . . . . . . . . . . . . . . . . . . . . 25
1.9 Event study estimate with year and school fixed effects . . . . . . . . . . . . . . 30
1.10 Event study estimate with province-year and school fixed effects . . . . . . . . . 32
1.11 Event study estimate with year and school fixed effects: number of students . . . 35
1.12 Outcome variables: an indicator taking 1 if the proportion of native language
users increased more than certain amount . . . . . . . . . . . . . . . . . . . . . 36
1.13 Outcome variables: an indicator taking 1 if the proportion of native language
users ever increased more than certain amount . . . . . . . . . . . . . . . . . . . 37
1.14 Effects of television exposure on vote shares for political parties . . . . . . . . . 41
1.15 Effects of television exposure on vote shares for political parties, with municipal-
ity fixed effects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
1.16 Effects of television exposure on vote share ANC: all states . . . . . . . . . . . . 44
1.17 Effects of television exposure on turnout rate . . . . . . . . . . . . . . . . . . . . 45
2.1 Political supports by year during Apartheid . . . . . . . . . . . . . . . . . . . . 58
2.2 Ground Conductivity in South Africa . . . . . . . . . . . . . . . . . . . . . . . . 60
2.3 Predicted AM radio coverage, using Sommersfeld-Norton model . . . . . . . . . 61
2.4 Proportion of parties by location and year . . . . . . . . . . . . . . . . . . . . . 65
2.5 Areas which allowed mining recruitment from Crush (1993) . . . . . . . . . . . 72
3.1 Competition and Growth of the Cotton Spinning Industry . . . . . . . . . . . . . 82
3.2 Example of a company report . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
3.3 Location of firms by years 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
3.4 Wage variation by location . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
A.1 Construction of new transmitters by year . . . . . . . . . . . . . . . . . . . . . . 117
A.2 Event study estimate with school-level cluster . . . . . . . . . . . . . . . . . . . 118
A.3 Effects of television exposure on vote shares for political parties, with municipal-
ity fixed effects with voting ward level clustering . . . . . . . . . . . . . . . . . 118
A.4 Event study estimate with year and school fixed effects, standard TWFE . . . . . 119
viii
A.5 Event study estimate with year and school fixed effects, additional covariant from
Census 2001, interacting with year . . . . . . . . . . . . . . . . . . . . . . . . . 120
A.6 Event study estimate with year and school fixed effects: cutoff 55dB?V/m . . . . 121
A.7 Event study estimate with year and school fixed effects: cutoff 65dB?V/m . . . . 121
A.8 Event study estimate with year and school fixed effects: cutoff 70dB?V/m . . . . 122
A.9 Event study estimate with year and ward FEs: votes for ANC: cutoff 55dB?V/m . 123
A.10 Event study estimate with year and ward FEs: votes for IFP: cutoff 55dB?V/m . . 123
A.11 Event study estimate with year and ward FEs: votes for ANC: cutoff 65dB?V/m . 124
A.12 Event study estimate with year and ward FEs: votes for IFP: cutoff 65dB?V/m . . 124
A.13 Event study estimate with year and ward FEs: votes for ANC: cutoff 70dB?V/m . 125
A.14 Event study estimate with year and ward FEs: votes for IFP: cutoff 70dB?V/m . . 125
B.1 Methods to take weighted average of ground conductivity . . . . . . . . . . . . . 134
B.2 Approximation of ground conductivity, taking into account the travel of waves . . 134
B.3 Proportion of parties by location and year . . . . . . . . . . . . . . . . . . . . . 142
B.4 Turnout by year for different by exposure to AM radio . . . . . . . . . . . . . . . 143
C.1 Proportion of the cotton spinning industry and the textile industry to GNP . . . . 144
C.2 Male and Female Workers?s average wages, 1883-84 and 1888-89, comparison
between Osaka and others . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145
C.3 Osaka Spinning mill female wage distribution around first capacity expansion . . 145
ix
Chapter 1: Building the ?Rainbow Nation? through Mass Media: Television,
Cultural Diversity, and National Unity in Post-Apartheid South Africa
The rainbow has come to be the symbol of our nation. We are turning the variety
of our languages and cultures, once used to divide us, into a source of strength
and richness. ? Speech by Nelson Mandela on National Reconciliation Day, 16
December 1995
1.1 Introduction
Many countries in the world are culturally, ethnically, or religiously diverse. Preserving
cultural identities for minorities in such societies involve challenges. On the one hand, many
countries have experienced conflicts between cultural diversity and the traditional idea of nation-
state (Baumann 2002): minorities have often been forced to be assimilated to majorities or ex-
cluded and marginalized from societies. On the other hand, some countries attempt to generate
united identity from diversity, such as E pluribus unum in the United States and ?Unity in Diver-
sity? in the European Union. Many countries, as well as the international society, nowadays try
to protect rights of minorities,1 however, there is no clear answer for policy makers regarding to
1For example, the General Assembly of the United Nations declared the Declaration on the Rights of Persons
Belonging to National or Ethnic, Religious and Linguistic Minorities in 1992.
1
balance multiculturalism and organizing unified nation.2
South Africa, after Apartheid, is one of the countries that confronts a challenge to balance
cultural diversity and national unity. After the experiences of strict racial segregation during
the period of Apartheid, where the white majority systematically segregated and discriminated
other racial groups, the new South African national government has tried to overcome these lega-
cies. ?Rainbow Nation? became the slogan of the country, aiming that different racial and ethnic
groups coexist, and making such differences a source of unity instead (Baines 1998). With a lot of
struggles, the country has tried to promote cultural diversity and national unity by depoliticizing
ethnicity by simultaneously implementing multiple policies, such as making 11 languages (two
of European origins, and nine of African origins) official languages, using multiple languages in
the national anthem, and letting schools choose languages of instruction (Barnett 1999).
In this chapter I explore whether national mass-media can promote cultural diversity while
preserving national unity. I focus on television broadcasts, which visibly convey the same in-
formation to a large number of audiences. Previous studies, such as Blouin and Mukand (2019)
and Russo (2021), show that television broadcasts promote cultural assimilation. By contrast, the
aim of the ?Rainbow Nation? in South Africa is distinct from cultural assimilation, because it
aims to sustain cultural diversity. Therefore, this chapter studies the impacts of mass-media in a
more complex situation, where the country promotes cultural diversity. I investigate impacts of
television broadcasts on two main outcomes that are associated with both cultural diversity and
national unity: choices of languages in elementary schools and voting shares of political parties
2Multicultural countries experience debates over various dimensions in some degrees, for instance, on choices
of languages in a parliament and courts, drawing internal boundaries within local regions, distribution of political
offices, and so on(Kymlicka 1995).
2
that expose opposite views on national unity.
In South Africa, free-to-air television broadcasts were dominated by South African Broad-
casting Corporation (SABC) owned by the national government.3 Aligning with the national
goal to pursue the ?Rainbow Nation,? SABC promoted programs to emphasize the co-existence
of multiple racial and cultural groups. Multiple languages are mixed intentionally within a pro-
gram to depict the cultural mixture of the country (Barnard 2006).
Television broadcasts in South Africa during Apartheid mainly targeted the urban popu-
lation, and the coverage expanded in rural areas at the end of 1990s and mid-2000s (Barnett
1999). I leverage on the construction of new television transmitters in 2000s as a source of ge-
ographical and time variations in the empirical analyses. To do so, I first digitize locations and
features of television transmitters in 2000s from governmental gazettes. Then, I employ engi-
neering models to calculate coverage of television broadcasts for each year and location. Finally,
I employ difference-in-differences empirical models to obtain the effect of exposure to television
broadcasts.
As a first outcome variable, I investigate the impact of television broadcasts on language
choices in elementary schools. In South Africa, the national government allows choices of lan-
guages by school level. Teachers, parents, and representatives from local communities, denom-
inated School Governing Body (SGB), have gatherings every year and decide languages to be
used in each school and grade.Languages of instruction in schools are important both in terms of
cultural and economic aspects. From the cultural aspect, Black South African parents frequently
demand English education due to negative images toward their own mother tongues which are
3In addition, e-tv became the first and only free-to-air television broadcasts in 1990s. Most e-tv transmitters were
built in 1990s, and there is no or little variations in time for my study periods. Therefore, location fixed effects in
my analysis absorbs all receivability for e-tv.
3
regarded as ?second class languages.? Such negative images are stemming from educational ex-
periences during Apartheid (NEPI 1992). Economically, English education is demanded due to
the expected higher return, though empirical evidence shows opposite (Taylor and von Fintel
2016). Television broadcasts with emphasis on cultural diversity are expected to reduce negative
images of native languages as ?second class languages.? Moreover, usage of native languages in
television programs provides role models of successful individuals who use their own languages.
Therefore, if the television broadcasts successfully convey their messages on cultural diversity, I
expect to observe, at schools, an increase in the usage of native languages, instead of English. I
use annual school-level data capturing the universe of schools in South Africa between 2002 to
2010 and estimate the impact of exposure to television broadcasts on language choices in schools
by dynamic difference-in-differences model proposed by Sun and Abraham (2020).
There are three important issues related to the identification strategy adopted in this study.
First, along with controlling for year fixed effects and location fixed effects, I also control for
the free-space signaling. Radio waves are blocked by topographical objects, such as hills and
mountains. On the other hand, the free-space signaling is the hypothetical radio coverage with-
out any topographical objects, determined solely by transmitters? features. By controlling both
location fixed effects and the free-space signaling, the remaining variation only comes from topo-
graphical features between transmitters and receivers. This approach has been taken by the recent
literature of economics to estimate causal impacts of radio and television broadcasts on various
socio-economic outcomes (see Olken (2009) and Yanagizawa-Drott (2014) as seminal papers us-
ing the same variation), and regarded as plausibly exogenous. However, even after controlling
by free-space signaling and location fixed effects, the treatment variation may be not exogenous,
if the government takes into account such topographical features when they decide transmitters
4
locations. 4 Therefore, as the second important point of the identification strategy, I test the
null hypothesis of parallel pre-trend in the difference-in-differences model. Third, I incorporate
treatment heterogeneity over timing of the treatment. Recent development in the literature of
econometrics on difference-in-differences show that estimated treatment effects through standard
two-way fixed effects are contaminated when treatment effects are heterogeneous over treatment
timing (Goodman-Bacon 2019, de Chaisemartin and d?Haultfoeuille 2019, Callaway and San-
tAnna 2020, and Sun and Abraham 2020). Therefore, I employ methods proposed by Sun and
Abraham 2020, which is able to account for event-study type of estimates with heterogeneous
treatment timings.
I find that exposure to television broadcasts increases the use of native languages in ele-
mentary schools. Usage of native languages is increased by approximately 3 percentage points
in three years after the construction of transmitters, and around 4 percentage points after five to
seven years. These are large effects, given that the baseline proportion of students who use native
languages is about 24 percent. I also find that this result is mainly driven by schools located in
municipalities which are ethnically heterogeneous. My results suggest that television programs
indeed encourage local communities to use native languages, instead of English, which contribute
to promote cultural diversity.
Promotion of cultural diversity through mass media may cause, however, unintended con-
sequences that the country lose unity, because each cultural group may want to diverge from
national policies by strengthening their own identities. As the goal of the government is to create
4Since a single television transmitter can deliver television signal by fairly long distance (more than 100km
with high voltage), it is basically impossible to consider entire topographical features for 360 degrees. However,
constructors may consider a few ?targeted villages? to avoid particular geographical objects, and such selections
may be not controlled by free-space signaling and the location fixed effects.
5
the ?Rainbow Nation,? it is crucial to examine whether mass media could also foster national
unity. Therefore, I examine effects of exposure to television broadcasts on voting outcomes,
which are associated both with ethnicity and national unity of South Africa.
I investigate effects of television broadcasts on vote shares of two political parties: African
National Congress (ANC) that promotes the ?Rainbow Nation? policy5, and Inkatha Freedom
Party (IFP) that is predominantly supported by one particular ethnic group (isiZulu) that initially
sought ethnic independence. I use voting-ward level data consolidated by De Kadt and Larreguy
(2018), which capture approximately 4,000 voting wards for each election. I expect to observe an
increase in the vote share of ANC if the broadcasts balance cultural diversity and national unity
objectives simultaneously, and a decrease in the vote share if the broadcasts emphasize cultural
diversity more than they intended. Similarly, I expect to observe an increase in the vote share of
IFP if the television broadcasts emphasize importance of ethnicity more than intended. Employ-
ing the same identification strategy used to study the choice of languages of instruction at school,
I find that the exposure to television broadcasts increases the vote share for ANC, while there
are little impacts on the vote share for IFP. I find that vote shares for ANC increased by as much
as 6.6 percentage points, four elections after exposure to television broadcasts. Results suggest
that messages from television broadcasts successfully increase support for the party which pro-
motes cultural diversity and national unity, but they do not promote voting for a party with ethnic
attachment.
5ANC has been a majority party since 1994, the first general election after the end of Apartheid.
6
1.1.1 Related literature
The main contribution of the study is to provide novel causal evidence on whether mes-
sages from mass media can promote cultural diversity, while preserving national unity. As a
broad picture, this chapter contributes to the growing literature of identity formation in eco-
nomics, stemming from Akerlof and Kranton (2000). A few studies demonstrate impacts of mass
media on the formation of cultural identities. Russo (2021) finds that radio, during early 20th
century, increased cultural assimilation by analyzing first names of immigrants in the United
States, and Blouin and Mukand (2019) show that messages from radio contributed to reduce
salience of ethnicity and increase inter-ethnic trust in the post-conflict environment in Rwanda.6
The direction of the effects, however, are toward cultural assimilation, which are against the cur-
rent international agenda to promote cultural diversity and ensure minority rights. In addition,
messages on cultural diversity from a national government can only be successful when it can
preserve national unity.7 My study, to the best of my knowledge, is the first to provide evidence
that messages through mass media can balance both.
Second, and related to the first contribution, this chapter relates to the growing literature
to understand identity and nation building. For example, Bazzi, Gaduh, Rothenberg and Wong
(2019) show that intergroup contact between other ethnic groups fosters creation of national unity,
measured by usage of a common language within households in Indonesia. While Bazzi et al.
(2019) analyze cultural assimilation as a measurement of national unity, this chapter examines
6In addition, there are some evidence that mass media impacts to form culture in broader context, such as fertility
choices (La Ferrara, Chong and Duryea 2012) and women?s status (Jensen and Oster 2009).
7There are some evidence that collective experiences can increase national unity (e.g. a victory of soccer game
by Depetris-Chauvin, Durante and Campante (2020)), but previous studies do not answer how to balance national
unity with cultural diversity.
7
diversification in culture and making such a diversification a source of unity in the nation building
context.
Third, this chapter is the first to provide causal evidence on whether mass media could
influence choices on languages of instruction in schools, which is collectively made by local
individuals involved in education. Languages of instruction are important both culturally and
economically. Culturally, forced changes in languages of instruction may cause cultural back-
lash. Clots-Figueras and Masella (2013) show that policy changes in languages of instruction
in Catalonia, from Catalan to bilingual (Catalan and Spanish), decreased cultural attachment to
Catalan, and Fouka (2020) shows cultural backlash by banning German in the U.S. after the
World War I for German students. Economically, languages of instruction may change returns to
education. Impacts are mixed in previous studies: Angrist and Lavy (1997) show usage of Ara-
bic instead of French decreased French ability in future, Ivlevs and King (2014) show decrease
in educational attainments for Russian minorities after changing 40% of the medium to Latvian
from 100% of Russian, and Taylor and von Fintel (2016) shows usage of mother tongues in early
grades improves future English abilities in South Africa, by improving learning efficiency. The
World Bank considers the current situation of the world problematic, that 37% of students in
low- and middle- income countries are required to use different languages (World Bank 2021),
which negatively affect their learning efficiencies. My paper does not directly estimate economic
impacts caused by changes in the learning languages, but instead, takes a step back and considers
how such language choices are made. In some countries such as Morocco studied by Angrist and
Lavy (1997) and Malawi, central governments are in charge of changing language policies on
education. In contrast, in South Africa, choices are up to each school and community, and I take
advantage of such a setting.
8
Fourth, I add evidence that mass media impacts political outcomes. This aligns with pre-
vious literature showing impacts of radio and television on political supports (DellaVigna and
Kaplan 2007, Di Tella, Galiani and Schargrodsky 2021), populism (Wang 2020), votes for fe-
male candidates (Okuyama 2019), genocide (Yanagizawa-Drott 2014), and conflicts (Armand,
Atwell and Gomes 2020). In terms of ethnic attitudes, my findings on ethnic party are the closest
to DellaVigna, Enikolopov, Mironova, Petrova and Zhuravskaya (2014), which find unintended
media exposure from a neighboring country (Serbia) increased negative attitudes toward Serbian
for Croatian citizens.
The rest of this chapter is structured as follows. In Section 3.2, I summarize background
information, including South African political environment, South African mass media and its
contents, and language choices in elementary schools. In Section 3.3, I explain data I use for the
chapter. I then report my identification strategies in Section 1.4 and report my results in Section
1.5. Section 2.6 concludes the chapter.
1.2 Background
1.2.1 South African political environment
During Apartheid, South African population was separated by race and ethnicity. Apartheid
refers to the set of policies regarding racial segregation of the White population against other
races, under the National Party?s regime between 1948 to 1990s8. One of the main policies im-
8Segregation and discrimination existed throughout the history of South Africa, including before the election
in 1948. After the independence of the Union of South African of 1910, the White national government enforced
multiple laws to segregate other races. This includes the South Africa Act of 1910 which ensured the monopoly of
political control by the White and the Native Land Act of 1913 that restricted the Black population to live in limited
areas. New laws after the National Party became the majority party include the Prohibition of Mixed Marriage Act
of 1949 and the Population Registration Act of 1950, which strictly classified citizens into four races (White, Black,
9
plemented under Apartheid was ?separate development,? meaning ?separate national units on the
basis of language and culture? (Rogers 1980, p.63). A systematic policy on separate development
emerged after Hendrik Verwoerd became Prime Minister in 1958. Not only races, but also ethnic
groups within the Black population were separated. The government allocated small portions of
land to each ethnic group, which were generally remote, rural, and arid, and called these lands
homeland or Bantustan. The African National Congress (hereafter ANC), which was an illegal
organization playing the central role of the resistance against Apartheid regime, strongly opposed
these policies, and pursued a ?single multi-national society.?
Moreover, the government used the national broadcasting system to promote ethnic division
by broadcasting programs in each ethnic language, which was called Radio Bantu. According to
Hayman and Tomaselli (1989) ?Radio Bantu sets itself the task of inducing the majority of Black
South Africans to accept their ?homelands? status and to view it as independence and develop-
ment? (pp. 100-101). Not only news, but all music also had to be related to the culture of each
ethnic group, to be consistent with the policy of separate growth (Hamm 1991).
After the election in 1989, the National Party decided to allow all races to have voting
rights. African National Congress (ANC), which was previously illegal and classified as a terror-
ist group, also became a legal political party. After 5 years of a transition period, South Africa
had its first all-race election. ANC became the majority party achieving more than 60% of votes
and promoted rapid reconciliation between races. Including the quote in the beginning of this
chapter by Nelson Mandela, the party constantly delivered messages to achieve national identity
Coloured and Indian) See Kaplan, McLaughlin, Marvin, Nelson, Rowland and Whitakar (1971) and Posel (1987)
for detailed explanations on fundamental policies before and during Apartheid.
10
with cultural diversity, holding equal foundation of citizenship (Baines 1998).9 Figure 1.1 shows
the support for political parties by year for national elections after the end of Apartheid (Data
are from IEC 1994?2019). As shown in the figure, ANC has gained the majority of seats in all
elections.
Figure 1.1: Votes by party after the end of Apartheid
Among multiple parties, Inkatha Freedom Party (IFP) has a unique feature. Large majority
of supporters are from the Zulu ethnic group, largely concentrated in the former KwaZulu Ban-
tustan. The first leader of the party, Mangosuthu Buthelezi, was a local leader of the homeland.
While ANC sought multi-race and ethnic cooperation, IFP aimed for ethnic independence. They
initially opposed to register in the 1994 general election to pursue an autonomous and sovereign
Zulu Kingdom (which was not proposed in the end and their leaders decided to participate in the
9As explained in the introduction, promotion of cultural diversity may contradict with the traditional idea of
nation-state. ANC also understand the fact. Indeed, they commented in the discussion document as follows: ?... it
should be noted that some of the identities associated with culture or ethnicity or religion can in fact be contradictory
to the building of a new nation that is based on principles of equity.? (ANC 1997, pp32)
11
election). Since then, most supporters belong to the Zulu ethnicity from Kwa-Zulu Natal province
(KZN): in KZN, they won more than 50% of votes in the 1994 election, while they got small vote
shares in other states (see Piper (2005) for more discussion). IFP no longer pursued ethnic inde-
pendence after the first general election, but the party consistently has strong connection to local
ethnic chiefs (Beall, Mkhize and Vawda 2005). n the empirical analyses, I consider supports for
IFP as a proxy of the preferences toward own ethnic group (instead of multi-culturalism by ANC)
in KZN province.
Table 1.1: Summary of historical background
During Apartheid After Apartheid
Systematic racial segregation Rapid reconciliation
- Separate development - ANC became the majority party
Politics - Homeland (Bantustan) - ?Rainbow Nation? policy
- African National Congress (ANC) - Inkatha Freedom Party (IFP)
was illegal was one of main opponent
Promotion of racial cleavage Multi-cultural programs
- Relatively late introduction - Rural areas had poor coverage
Mass Media of television broadcasts until 2000
- Strict censorship - Construction of transmitters in 2000s
- Ethnic and racial independence - Programs promote diversity
Table 1.1 summarizes the background information on South African history and mass me-
dia, which is discussed in the next section.
1.2.2 Background of South African television broadcasts
South African Broadcasting Corporation (hereafter SABC) is a dominant broadcaster in
charge of television programs in South Africa, owned by the national government. There are
some commercial paid-in TV programs too, but these are not the focus of this study since I focus
on rural households, where such paid-in programs do not have much coverage. Due to the strong
12
media censorship in South Africa, the country introduced television broadcasts relatively late in
1976. The censorship of mass-media was due to the promotion of ethnic and racial segregation.
After the end of Apartheid, there was a rapid reform on the contents of broadcasting, which
aimed to promote cultural diversity and generating unity from diversity relying on the slogan
?One nation, many cultures? (Barnett 1999). SABC1s (one of the leading programs) jingle at
this period, ?Simunye - We are one,? represents the message to balance cultural diversity and
national unity (Baines 1998).10 Literally, Simunye means ?we are one? in isiZulu, which is one
of the languages used in SABC1. Barnett also mentioned that before 2000s, ?rural areas are
particularly poorly served (by mass media), and to a considerable extent television is an urban
media form? (Barnett 1999, pp. 279). Indeed, SABC1 had 129 transmitting stations in 1995.
The number was increased to 140 in 2000 and to 169 in 2008. I exploit these variations in the
construction of new transmitters, especially in rural areas.
SABC has three programs. SABC1 and SABC2 mainly target the youth African popula-
tion. Importantly, these two programs use both English and Black languages. According to the
Annual report of SABC in 2008, approximately 70% of programs were in English, while the other
30% were in other languages. SABC1 uses Nguni languages (isiXhosa, isiZulu, isiNdebele, and
siSwati), and SABC2 uses seSotho, seTswana, xiTsonga, TshiVenda and sePedi in addition to
English. SABC 3 mainly uses English with some Afrikaans, mainly broadcasting imported pro-
grams from English speaking countries (i.e. the United States and the United Kingdom). Every
day, SABC1 and SABC2 broadcast news in Black African languages for an hour.11 In addition,
10Private sector used a similar concept for their advertisement: for instance, Castle Lager, the largest brewery in
South Africa, had a slogan ?One Beer, One Nation? (Baines 1998).
11For instance, SABC1 broadcast news in isiNdebele, and siSwati between 6pm to 6:30pm, and in isiXhosa
between 7pm to 7:30pm for every weekday, as of the first week of April in 2007. Though it?s not comprehensive,
one can find information on programs for each channel at TVSA, the online industry publication.
13
Barnard (2006) shows that many programs are multilingual to achieve the national goal of pro-
moting cultural diversity and mutual understandings. Several presenters use different languages
in news and quiz shows, while soap operas frequently mix multiple languages (Barnard 2006).12
I visually present expansion of the coverage of television broadcasts for SABC1 using
multiple figures. To begin with, Figure 1.2 shows the population density measured by the number
of people per km2 in 2001 for each municipality.13 The figure roughly approximates demands
for television broadcasts by region. One can observe high population density in the east, north
and south parts of the country, while there is relatively scarce population in the west side of the
country.14
Figure 1.2: Population density by municipality as of 2001
12Muvhango, the long-lasting soap opera started in 1996 was originally in TshiVenda on SABC2. The drama
has been translated into multiple languages and achieved 4 million viewership on average (Ndlovu 2012). Several
characters speak multiple languages in a sitcom 102 Paradise Complex, and a soap opera Isidingo. In these programs,
there are sometimes no subtitles for multiple languages: Barnard (2006) analyzes the situation the situation as
?an assumption of an already multilingual viewership or a viewership not loyal to conventional conceptions of
understanding and mastery but resigned or accepting or delighting in the knowledge that understanding is always
partial.? (Barnard 2006, pp49).
13Population density is calculated by the 10% sample of the census in 2001.
14Most areas in the west side of the country is called Karoo. This area is arid and incompatible with farming.
14
Figure 1.3 shows the coverage of television broadcasts as of 2001. Signal strength is the
indicator of whether people can watch television broadcasts. The higher the value is, the higher
the chance that people can watch television without noise. It is usually said that having a signal
more than 60dB?V/m is sufficient to watch TV.15 Therefore, people in areas in pink and yellow
colors in each map have a high chance to watch television broadcasts, while areas in green, blue
and white colors have low chances. Though wide range of areas were covered by TV in 2001,
some areas with high population density, especially in the east and the northwest side of the
country, had low coverage of television broadcasts. Therefore, the government constructed new
television transmitters in these areas. Figure 1.4 shows the coverage of the newly constructed
transmitters between 2002 and 2010. Figure A.1 in the appendix further shows the construction
of new transmitters every two years to illustrate the time variation of the constructions. In the
empirical analysis, I exploit this time and geographical variation of the construction of these
transmitters.
There are two features regarding the exposure to television broadcasts that I could not fully
capture in my analysis, due to lack of data. First, it is the exposure to satellite television. DStv
started its operation in 1995, and it has been the only satellite television operator in South Africa
throughout my sample period. Due to the features of satellite television, people living in areas
without coverage of terrestrial television can watch DStv. Since DStv also broadcasts SABC
channels, I may underestimate the effects of television broadcasts on various outcomes if some
15Governmental agencies in each country provide guidelines for recommended signal strength. Federal Commu-
nications Commission in the United States defines ?Grade B? criteria, which is ?a field strength that is strong enough,
in the absence of man-made noise or interference from other stations, to provide at least 90% of the time a television
picture that the median observer would classify as ?acceptable?? (FCC 2000, pp2). FCC notes that 64dB?V/m is
required to the frequency band for South African television (470-890MHz). For other countries, Japan recommends
60dB?V/m (ICCJ 2007), South Africa recommends 65dB?V/m(ICASA 2008b), International Telecommunication
Union (ITU) recommends 58dB?V/m in rural areas(ITU 2002).
15
Figure 1.3: Coverage of SABC1, as of 2001
people in control areas watch SABC through the satellite. While I could not rule out such an
underestimation, it should not be significant since only approximately 10% of households have
access to satellite television during the sample period (News24 2005).16
Second is the transition of the propagation technique from analogue to digital. South Africa
started the transition in 2008 and completed its transition in 2011. The national government
announced the original plan in 2005, and then had the concrete plan in August 2008 (Department
of Communication 2008). Generally, digital propagation has wider coverage given the hight
and electricity of transmitters, so that more people are able to watch television with the digital
propagation. . I do not capture this change in the digital propagation in my empirical analysis.
This is because tuners owned by receivers should be replaced to watch digital television, and I
do not have good data on the replacement. My main results do not change if I do not include
samples after 2008.
16The subscription of DStv in 2005 exceeded 1 million (News24 2005). The number of households, according to
the census in 2011 was approximately 14.45 million.
16
Figure 1.4: Additional coverage of SABC1 between 2002-2010
1.2.3 Background of language choices in South African schools
In South Africa, there are 11 official languages: two of them are of European origins (En-
glish and Afrikaans) which were the only official languages during Apartheid, and nine of them
are Bantu African languages. The South African government allows choices of languages by
school level (South African Schools Act (Act 84 of 1996)). Teachers, parents, and representatives
from local communities have gatherings, which are called School Governing Bodies (SGB),17 and
decide languages to be used in each grade.The Revised National Curriculum Statement (RNCS)
of 2002, which stipulates the curriculum for elementary schools, states that ?the home language
should continue to be used in conjunction with the additional language for as long as possible?
(DOE 2002). In practice, students in the early stage (1st-3rd grade in elementary schools) use
17In addition, School Governing Bodies have a right to decide level of school fees to supplement government
funding (Chisholm (2012) and also have some discussions on school discipline and religious policies (Msila 2007).
17
mother tongues as languages of instruction, and then switch to English.18 This policy is not
compulsory, but the national and provincial governments promote such a transition (Taylor and
von Fintel 2016)19. Yet, many schools do not follow this rule, depending on the discussions
at School Governing Bodies, and use English from grade 1, or keep using native languages at
grade 4. Throughout the study periods, 24% of students use native languages (i.e. not English,
Afrikaans). According to the data analyzed by Taylor and von Fintel (2016), switching languages
by grade-level sometimes occur. Approximately 6% of schools switch from native languages to
English, and 14.3% switch from English to native languages for some grades.
Demand and supply for English and native language education are complex. Parents usu-
ally demand English education. NEEDU (2013), the governmental report on literacy education
in South Africa, indicates there are increasing demand for English or Afrikaans as the learning
language in early grades. This may be due to the high (expected) returns to education in English,
which is estimated between 40 to 80% more relative to education in native languages as of 2000
(Levinsohn 2007). On the other hand, Taylor and Vinjevold (1999) show that learning in mother
tongues often helps to enhance learning efficiency for Black students, as most teachers in rural
areas are Black and their mother tongues are not English. In addition, Taylor and von Fintel
(2016) show that English education in early stages of education decrease academic attainment in
later grades, because of the low learning efficiency. Yet, bad images of native languages asso-
18Learning additional first language in the 1st grade became necessary in the new curriculum as of 2012 (Cur-
riculum and Assessment Policy Statement: CAPS). This additional language was, in most cases, English for Black
students. This was not the mandatory for the earlier curriculum (Du Plessis and Marais 2015.)
19For instance, Teachers? Guide for Revised National Curriculum Statement (RNCS) of 2002 discusses the diffi-
culties on learning mathematics for students who speak indigenous languages, as ?there are very poorly developed
Mathematical lexicons for most of South Africas indigenous languages? (DOE 2003, pp.32). The Teachers? Guide
also implicitly assume that students switch their languages in Grade 4, as it states ?Another source of difficulty aris-
ing from language is the transition that some learners experience in Grade 4 when they must move from learning in
their mother tongue to learning in a language of instruction that is not their mother tongue (DOE 2003, pp.32)?
18
ciated with inferior eduction during Apartheid push parents to demand English education (NEPI
1992). Furthermore, in rural schools, students have few opportunities to use textbooks because
of the scarcity (Taylor and Moyane 2005, Setati, Adler, Reed and Bapoo 2002). Consequently,
Setati et al. (2002) describe that learning in English in rural schools is close to learning a for-
eign language instead of learning of an additional language. From the supply side of education,
government advocates for using native languages both because they want to promote the national
goal of cultural diversity, as well as improve learning efficiency in early grades (NEEDU 2013).
1.3 Data
The data for the study comes from multiple data sources. First is the television coverage
from Government Gazettes, which is the main explanatory variable in my empirical analysis. The
outcomes studied are gathered from school-level data on languages of instruction and election
outcomes.
1.3.1 Data: Television coverage
I construct a panel data of the locations of television transmitters, that are exploited from
the policy documents by Independent Communication Authority of South Africa (ICASA) stored
in the Government Gazette No. 23695 (ICASA 2002) and No. 32728 (ICASA 2008a).
I use Cloud RF, the online commercial interface to predict television coverage20. The in-
terface originally installs topographical data on the global surface by 30m?30m. I apply an
engineering model, the Longly-Rice model, to compute the following two variables. The first
is the actual signal strength of the television broadcasts, and the other is the free-space signal-
20Alex Farrant at Cloud RF helped me to implement my analysis.
19
ing. The actual signal strength is determined by (i) transmitters features, such as electric powers
and heights, and (ii) topographical features between a transmitter and a receiver. Topographical
features are crucial, because radio waves are blocked by physical objects, such as hills and moun-
tains. The free-space signaling is the hypothetical coverage of television, which only considers
(i) transmitters features, and ignores (ii) topographical features.
Figure 1.5 shows the trend of the construction of transmitters by program. Figure 1.5-(a)
shows the number of new transmitters by year by program, and Figure 1.5-(b)shows the cumula-
tive numbers of transmitters. SABC1 and SABC2 did not experience an increase in the number of
transmitters in 1990s. Expansions on the construction of transmitters, targeting rural population,
started in 2000s. As a result, SABC1 and SABC2 experienced increases in the number of trans-
mitters during the 2000s. In each year in 2000s, around 5 transmitters were constructed for each
program. In contrast, SABC3 did not experience large increases in the number of transmitters in
2000s.
(a) New transmitters over time (b) Cumulative numbers of transmitters
Figure 1.5: Construction of new transmitters for SABC
20
1.3.2 Data: Annual Schools Survey
The data on language of instruction comes from the Annual Schools Survey of South Africa
(ASS). This is an annual survey conducted by the Department of Basic Education on the first
Thursday of March every year. Each school reports information regarding teachers and students,
such as the number of students enrolled in each grade, the number of teachers, and more impor-
tantly, the number of students by the use of languages. I merge this data to the South African
School Master List, which is also provided by the Department of Basic Education to obtain ge-
ographical locations. Approximately 20% of the samples were not included in my analysis, due
to either (1) missing reports or (2) missing geographical location data. It should be noted that the
choices of languages in ASS are reported by principals. Therefore, there may be some measure-
ment errors in the outcome variable. This would not affect consistency of the estimator, as far as
the errors are occurring at random (see Wooldridge 2010).
Table 1.2 shows summary statistics for the year 2002. Panel A shows the features of treated
schools. This includes schools which did not have television signals at the initial period of the pe-
riod studied (i.e. 2002), but experienced construction of a television transmitter close to their lo-
cations, and received television signals at some periods before 2010. Panel B shows the statistics
of the control group, which have never received the television signal stronger than 60 dB?V/m
throughout the period studied. Panel C shows the statistics of schools which are not used in the
analyses. Most of these schools are located in urban areas (i.e. major cities, such as Cape Town,
Johannesburg and Durban) which receive television signals throughout the sample periods. No-
tably, all of the reported indicators, the number of students, the size of schools, and the size of
local municipalities are all similar between treated and control schools. By contrast, the school
21
size and the population of municipalities are much larger for the schools not used in the analyses.
Table 1.2: Summary statistics
Mean SD Min Max N
Panel A: Treated group
Number of students 342.48 267.93 11 3000 2000
Proportion of students using native languages 0.24 0.24 0 1 2000
Total population in municipality 19,995.69 20,979.99 329 269223 2000
Panel B: Control group
Number of students 304.52 272.61 11 2674 3202
Proportion of students using native languages 0.24 0.25 0 1 3202
Total population in municipality 17,526.20 18,808.25 53 269223 3199
Panel C: Not used in the analysis
Number of students 439.07 334.57 11 3186 11064
Proportion of students using native languages 0.22 0.26 0 1 11064
Total population in municipality 64,183.53 87,282.48 60 269223 11058
Treated group in Panel C represents schools which are exposed to television (stronger than 60 dbu) between
2002-2010
Figure 1.6 shows the number of schools by treatment cohort. 3,202 schools have not re-
ceived television signal throughout the sample periods. Approximately 2,000 schools experi-
enced construction of a new transmitter in neighboring areas, so that they were able to receive
television signal at some point during the sample period. Since sizes of transmitters and the
number of new transmitters vary over time and space, the number of affected schools also vary
over time and space. There is only a relatively small number of treated schools which gained
television signal in 2009 and 2010, because the number new transmitters in this period were very
small.
Finally, Figure 1.7 shows distributions of the proportion of students using native languages,
22
Figure 1.6: Number of schools by treatment cohort
as of 2002. In both two graphs, the white bars with black outlines represent schools in the
control group. The red bars in Figure 1.7-(a) represents schools that are not used in my analyses.
Compared to the red bars, control group has smaller fraction of schools which only use European
origin languages (i.e. the left tail of the distribution). This is because urban schools which
are excluded from my analyses are more likely to be all-English or all-Afrikaans with higher
proportion of non-Black students. The green bars in 1.7-(b) represent treated schools. Not only
the average value reported in Table 1.2, distribution on the proportion of students using native
languages look very similar between control and treated schools in 2001 in the pre-treatment
period.
23
(a) Control group and full observations (b) Control group and treated group
Figure 1.7: Distribution of schools by the proportion of students learning in native languages
1.3.3 Data: Political attitudes
I obtain vote shares for political parties by political ward for municipality elections (2000,
2006, and 2011) and for the National General Elections (2004 and 2009).21 I use the digitized
data by De Kadt and Larreguy (2018). There are roughly 4,000 voting wards for each election,
with 19,576 observations over 5 election periods. I use variables constructed by De Kadt and
Larreguy (2018), on the vote shares for ANC and IFP as dependent variables. Figure 1.8 shows
the timing of treatments for voting wards. Each cohort experiences the new exposure to television
broadcasts between the time of election (labeled as cohorts in the graph) and the next election.
For example, the 2004 cohort experienced exposure to television between 2004 and 2006. Again,
because of the small scale of new transmitters in the late 2000s, the number of newly exposed
voting wards is very small in the last period, 2009.
21It should be noted that the municipality elections are the single-member winner-takes-all election by ward level,
while the National General Elections are the proportional representation. While there are differences in the voting
system by year, vote shares by party are similar year by year, because South African elections are party-centric,
rather than individual-centric. This approach is taken by previous studies (De Kadt and Larreguy 2018). Including
year fixed effects, I can control differences in the voting system.
24
Figure 1.8: Number of voting wards by treatment cohort
In addition to above main data sources, I complement the data by obtaining the ethnic
composition of each municipality from census. I use the 10% sample data from the 2001 census
to compute the proportion of languages used in each municipality.
1.4 Identification strategies
To identify impacts of exposure to television on changes in the proportion of native lan-
guage users in elementary schools, I estimate the following regression as a main specification:
?7
PropNative = ? 1l{t = t?st k s + k}+ ?FreeSpacest + ?s + ?t + ?st, (1.1)
k=?5
25
where PropNativest is the proportion of student using native languages (i.e. non-European lan-
guages) as languages of instruction22 for a school s at time t. 1l{t = t?s + k} is an indicator
variable taking 1 if year t is equal to the time of treatment (t?) +k.23 I define the treatment as
schools which receive television signals stronger than 60dB?V/m.24 Dynamic treatment effects
are preferable to be estimated, rather than, say, the common effect in a difference-in-differences
framework, because exposure to television may not influence choices instantly. It may take a
few years to buy a television, or long exposure to television may be needed to change individ-
uals? attitudes. ?s and ?t are school and year fixed effects. Standard errors are clustered at the
municipality level as of 2001, where there are approximately 200 municipalities. This is be-
cause the outcome variable may be correlated at that level. Changes in the proportion of students
using native languages are likely to be driven by switching languages of instruction in grade
level, determined by School Governing Bodies (SGB). SGB usually has some members from
local community, and these community members sometimes obtain information from regional
workshops on the importance of language choices (Webb, Lafon and Pare 2010).25
Similarly, I estimate impact of television exposure on vote shares for two political parties
using following model:
22Note that this is not a binary variable, since within school, some students may use English and others may use
native languages as languages of instruction.
23As shown in Figure 1.6, the number of treated schools are very small in the later years. Event study estimates for
leads which are far from the treatment timing use these small number of samples, which generate large confidence
interval. Therefore, I aggregate periods k < ?5 as if they were treated in k = ?5.
24This is an arbitrary cutoff suggested by the engineering literature explained in Section 1.2.2, and I show the
robustness of the results in Figure A.6 to A.8 in the Appendix. See Section 1.2.2 for more discussion on how I
choose this cutoff.
25The choice of clustering unit of my analyses is not as straightforward as in other studies clustering by the
treatment unit. This is because my original treatment (signal strength of the television broadcasts) is continuous.
I choose a reasonable unit of clustering, that error terms may correlate due to unobservable reasons. This is an
arbitrary choice: an alternative way to cluster observations is to cluster by individual level (i.e. school level in the
language choice analyses, and voting-ward level in the political analyses. In Appendix A.2.2, I present additional
results of clustering error terms by school level and voting ward level, instead of the municipality level. These results
show much smaller standard errors. In that sense, my main results are more conservative.
26
?4
V oteSharewt = ?k1l{t = t?w + k}+ ?FreeSpacewt + ?w + ?t + ?wt, (1.2)
k=?2
where V oteSharewt is a vote share for two political parties in voting ward w at year t. Instead of
school fixed effects in the previous equation, I include ?w, voting ward fixed effects. Fixed effects
are to control for differences in the local-level time invariant error terms and differences in the
election systems discussed in Section 1.3.3. Note that the data is not annual, but only available
in years when elections are held.
There are three important issues in my identification strategy. First, I control for the free-
space signaling (FreeSpacest). As explained in Section 1.3.1, coverage of television broadcasts
is determined by transmitters features and topographical features between transmitters and re-
ceivers. The free-space signaling is the hypothetical coverage of television, solely determined
by transmitters? features. By controlling for the free-space signaling, along with the location
fixed effects (i.e. school or voting-ward fixed effects)26 the remaining variation comes from to-
pographical features. This variation is considered as plausibly exogenous, and multiple seminal
works in this area in economics, such as Yanagizawa-Drott (2014) and DellaVigna et al. (2014),
that use cross-sectional variation of radio coverage as a causing force, adopt this approach. Note
also that in previous studies using cross sectional variation, such as Yanagizawa-Drott (2014), re-
searchers usually control for local topographical variations of receivers. In my analyse, location
fixed effects absorb such time invariant effects.
Even after controlling by free-space signaling and location fixed effects, the treatment vari-
26 Also note that FreeSpacest varies by time, because it is affected by newly constructed transmitters.
27
ation may not be strictly exogenous if the government considers such topographical features when
they decide transmitters location. However, since television transmitters can send radio waves
for more than 100km (with enough electricity), it is unrealistic to believe that policy makers can
consider entire receivers around transmitters. Yet, there may remain some strategic choices of
transmitters? locations. Therefore, a second important issue in my identification strategy is to test
the null hypothesis of parallel pre-trend implicit in the time effects assumption in equation 1.1
and 1.2.
The third important point of my identification strategy is to consider heterogeneity in treat-
ment effect over timing of the treatment. Recent development in the literature of economet-
rics on difference-in-differences show that estimated treatment effects through standard two-way
fixed effects are contaminated when treatment effects are heterogeneous over treatment timing
(de Chaisemartin and d?Haultfoeuille 2019, Goodman-Bacon 2019, Callaway and SantAnna
2020, Sun and Abraham 2020). There are multiple reasons for why the treatment effect could
vary by treatment cohort, such as differences in television programs, political environment, or
income levels. Therefore, I employ a method proposed by Sun and Abraham (2020), which
is able to estimate leads and lags with treatment heterogeneity over treatment timing. In plain
words, an estimator which is free of contamination, due to the heterogeneity over time, can be
obtained through running regressions cohort-by-cohort, and taking a weighted average of them.
For each regression, I only include a specific cohort that is treated at a particular time period and
never-treated units as a control group.27 Weights are given by the share of each cohort in total.
The effect of television on language choices is ambiguous. On the one hand, due to the aims
of SABC and the South African government to promote ethnic diversity, programs may increase
27I use the stata command eventstudyinteract, which is available at Liyang Sun?s webpage.
28
usage of native languages. In this case, we expect to observe ?k > 0 for k > 0, because the
television would play a role to mitigate negative images with respect to the use of native languages
and encourage students to learn in their own languages. On the other hand, since English contents
in SABC1 and SABC2 also increase exposure to English, television broadcasts may stimulate
students to use English, because television provides role models to use English. In such a case, we
expect to observe ?k < 0 for k > 0. For the same reasons, effects of television on vote shares for
various parties are ambiguous. Although the messages from television promote cultural diversity
and national unity at the same time, which is in favor of ANC?s views, unbalanced messages to
promote cultural diversity may increase the vote shares for IFP by emphasizing importance of
cultural identities.
1.5 Results
1.5.1 Results: languages of instruction
First, I analyze the effect of television exposure on language choices in elementary schools.
The proportion of students who use native languages (i.e. non-White languages) is the outcome
variable. Figure 1.9 shows the main result by estimating equation 1.1. The analysis uses schools
which were exposed to television between 2002 to 2010 and schools which were never exposed
to television in the study periods. The number of observations used in this analysis is 52,023.
Results are reported in column (1) of Table 1.3.
Figure 1.9 shows that exposure to television broadcasts contributes to the use native lan-
guages, instead of English. First, there is no clear trend before the exposure to television, and the
coefficients in pre-intervention are not statistically distinguishable from zero (individually and
29
Figure 1.9: Event study estimate with year and school fixed effects
jointly).28 Second, the point estimates are increasing after the treatment is introduced. Some
periods (t = t?s +1, t = t
?
s +3, t = t
?
s +4) are statistically significant at the 5% level, and all point
estimates are significant and large, given that the proportion of students using native languages
is approximately 24% across schools in 2002, as shown in Table 1.2. The effect of television
increases usage of native languages by approximately 3 percentage points three years after the
construction of transmitters, and around 4 percentage points after five years. The confidence in-
terval becomes larger in the later periods due to the small number of treated observations used to
estimate the coefficients.
Figure 1.10 shows the result estimating equation 1.1 with using province-year fixed effects,
instead of year fixed effects. The results are mainly unchanged from Figure 1.9. Results are also
reported in Table 1.3.
28F (7, 195) = 1.75 when I cluster sample at the municipality level, and F (7, 6229) = 1.18 when I cluster sample
at the school level.
30
Table 1.3: Event study: Effect of SABC ex-
posure on language choices in elementary
schools
(1) (2)
t = t?s ? 6 0.0280 0.0346
(0.0243) (0.0244)
t = t?s ? 5 -0.0131 -0.0100
(0.0179) (0.0169)
t = t?s ? 4 0.0134 0.0031
(0.0144) (0.0134)
t = t?s ? 3 0.0039 -0.0048
(0.0114) (0.0103)
t = t?s ? 2 0.0165 0.0161*
(0.0086) (0.0080)
t = t?s ? 1 0.0159 0.0187*
(0.0096) (0.0092)
t = t?s + 1 0.0154** 0.0129 **
(0.0052) (0.0051)
t = t?s + 2 0.0154 0.0106
(0.0082) (0.0075)
t = t?s + 3 0.0312 * 0.0182
(0.0137) (0.0108)
t = t?s + 4 0.0263 * 0.0193
(0.0134) (0.0114)
t = t?s + 5 0.0377 0.0441 *
(0.0222) (0.0176)
t = t?s + 6 0.0375 0.0411*
(0.0217) (0.0173)
t = t?s + 7 0.0452 0.0252
(0.0280) (0.0240)
Year FE Yes No
Province-Year FE No Yes
School FE Yes Yes
R2 0.601 0.610
N 52,023 52,023
? p < 0.05, ?? p < 0.01, ??? p < 0.001. Stan-
dard errors are clustered by local-municipality
level. All specifications control for free-space
signaling. Each explanatory variable takes 1 if a
school is exposed to television at time t?s and time
t is equal to t? ?s + k where ts is the time of con-
struction of new transmitter affecting the school
s. A school is classified as treated if signal at
school is stronger than 60 dBu.
31
Figure 1.10: Event study estimate with province-year and school fixed effects
1.5.1.1 Robustness
To check the robustness of my results, I report multiple different specifications for my
findings. First, I show the results using the standard two-way fixed effect model, instead of the
method proposed by Sun and Abraham (2020). The result is reported in Figure A.4. Results are
very similar to what I obtained using the method proposed by Sun and Abraham (2020).
Second, I test whether my results are robust by changing cutoffs in the definition of the
treatment variable. I pick up a cutoff of television exposure to construct the treatment variable
(i.e. more than 60dB?V/m) from previous engineering literature, but of course, this cutoff is
arbitrary, since the original variable of the signal strength is continuous. Some individuals with
relatively low signal coverage may still watch television with bad resolution. I estimate the same
specification with Figure 1.9 with different cutoffs in Figure A.6 to Figure A.8. Results are
generally unchanged by changing cutoff, with less significant effects for lower cutoffs. This is
consistent with the fact that by lowering the cutoff, treated areas include more areas with lower
32
chances of receiving television signal.
Third, I include additional covariant generated from the 2001 census. Unfortunately, school
data from ASS do not include other characteristics regarding schools, such as the number of
teachers, school budgets, and so on.29 To complement this situation, I include local characteristics
as of 2001, interacted with year dummies by municipality level. The data include average age,
average years of education, average income, and gender. The result is reported in Figure A.5,
which shows a very similar pattern as in the main result.
Fourth, I report the results with the traditional difference-in-difference framework estimat-
ing the following equation:
PropNativest = ?Signalst + ?FreeSpacest + ?s + ?t + ?st. (1.3)
Signalst is either the binary of continuous measure of signal strength of television. Results are
reported in Table A.2. Results with the binary Singalst is reported in columns (1), (4) and (7),
and results with continuous Singalst is reported in columns (2), (5), and (8). Panel A shows
results with clustering standard errors at the school level, and Panel B shows those clustering at
the municipality level. Columns (4) to (6) are the results estimating equation 1.3, while columns
(1) to (3) exclude FreeSpacest, and columns (7) and (9) add additional controls from the census
interacted with year dummies. Results are similar to the event study analysis, which are smaller
than the later periods of the event study. This may be because of the increasing trend of the
treatment effect over time, as shown in the main result. Since initial periods have small treatment
29ASS originally asks these questions in the survey. However, as the data is old, the Department of Basic Education
could not find the location of data at this point. I keep contacting the department, and will add information on these
features once I get an access to data.
33
effects, the treatment effect on average in difference-in-differences framework becomes smaller.
To overcome this, I also report the results estimating the following regression in columns (3), (6),
and (9):
PropNativest = ?Signalst ? TimeTrendst + ?FreeSpacest + ?s + ?t + ?st, (1.4)
where TimeTrendst captures the linear time trend, relative to the first year of exposure to televi-
sion. Consistent with the main event study specification, estimated coefficients are positive and
significant for the all specifications in Panel A, and insignificant at the 5% level but significant in
10% for the results in Panel B.
Fifth, I report whether the results are driven by migrations or changes in the dropout rate
due to the television, which can become confounders in my main results. After exposure to
television, parents (or students) may want to move from one school to another to obtain the
education that they prefer. If parents or students are encouraged to learn in native languages
after exposure to television, students may move from an English teaching school to a school with
native languages, for instance. Different direction of migration may be also possible. Also, if
television programs encourage (or discourage) students to study more, dropout rate may change
by exposure to television.30 If dropout occurs more in upper grades, which are more likely to use
English as a language of instruction, then the main results may be observed without changes in the
language of instruction for particular grades. To test whether this type of migration or dropout
drive the previous results, I run equation 1.1 with a different outcome variable, the number of
total students. The result is reported in Figure 1.11. I do not observe any significant effect of the
30It should be noted that the dropout rate in South African elementary school is not that high: in 2007-2008, there
were only 1 to 2% of students dropped out for each grade in elementary school (DBE 2011).
34
Figure 1.11: Event study estimate with year and school fixed effects: number of students
television exposure on the total number of students.
If changes in the language of education for some grades drive the main result, the change
in PropNativest should not be a few percentage points, but rather large. Since South African
elementary schools have seven grades, the proportion should be approximately 14.2% changes if
one grade switches from one language to another. To verify this point, I construct an indicator
variable on whether the proportion of students using native language changes by more than 15
and 25 percentage points, and make these variables as dependent variable of equation 1.1. Figure
1.12 shows the results. Again, I do not observe any significant effects in the pre-trend periods.
I observe constantly positive, but insignificant effects except the first year after exposure. This
is because the outcome variable is temporary: Once schools change languages of instruction,
then the indicator variable takes 1 at that period. However, the indicator takes zero after such
a period if there are no further changes in the languages of instruction. Overall, Figure 1.12 is
supportive evidence that the main results is driven by changes in languages by grade-level, rather
than migrations or dropouts.
35
(a) More than 15% changes (b) More than 25% changes
Figure 1.12: Outcome variables: an indicator taking 1 if the proportion of native language users
increased more than certain amount
To complement above shortcoming, I additionally examine whether there are big changes
in PropNativest after exposure to television. I construct a binary variable that is equal to 1 if
PropNativest ever changed by more than 15 and 25 percentage points before. That is, once this
variable takes 1 for period t, the variable takes 1 for all periods t? > t for a same school s. Results
are show in Figure 1.13. All pre-periods are not distinguishable from zero, and all post-treatment
periods are positive and significant. This also indicates that the main result is driven by changes
in school policy to switch languages of instruction.
1.5.1.2 Potential mechanisms
In this subsection, I discuss potential mechanisms which drive the changes in the languages
of instruction. I divide samples into multiple different categories and see whether the main results
are driven by some specific groups.
To test this, I divide samples into those located in municipalities which are ethnically ho-
mogeneous and those located in municipalities which are ethnically heterogeneous and estimate
36
(a) More than 15% changes (b) More than 25% changes
Figure 1.13: Outcome variables: an indicator taking 1 if the proportion of native language users
ever increased more than certain amount
equation 1.3 for various specifications. Table 1.4 shows the result. Columns (1) to (4) report the
results with the sub-sample of schools located in low ethnic concentration, and columns (5) to
(8) report the results with schools located in high ethnic concentration. Results Columns (1), (3),
(5), and (7) report the results using continuous measure of the signal strength in dB?V/m, and
other columns report the results with binary measure of the signal strength, where the indicator
takes 1 if the signal is stronger than 60dB?V/m.
In columns (1) and (2), I define areas with low ethnic concentration as municipalities where
less than 80% of the population is the sum of the proportion of ethnic groups which are the same
language groups broadcast in SABC. In other words, I define two large ethnic groups: the first
group with the languages used in SABC1 (isiXhosa, isiZulu, isiNdebele, and siSwati) and the
second group with the languages used in SABC2 (seSotho, seTswana, xiTsonga, TshiVenda and
sePedi). Columns (5) and (6) report the results with high ethnic concentration, where the sum
of the proportion of ethnic groups for the same classification as in columns (1) and (2) exceed
80%. Note that the proportion of particular ethnic groups is calculated by municipality level as
37
of 2001, which is the date before the data for language choices are available. In columns (3) and
(4), I define areas with low ethnic concentration as municipalities where less than 80% of the
population is one particular ethnic group. Similarly, I define areas with high ethnic concentra-
tion as municipalities where more than 80% of the population is one particular ethnic group in
columns (7) and (8). Note that due to the homeland policy during Apartheid (see Section 1.2.1
for the homeland policy), many municipalities are highly ethnically concentrated.
Results indicate that my main results, increase in the use of native languages in schools by
exposure to television, is mainly driven by schools located in the municipalities with low ethnic
concentration. Positive and economically relevant coefficients are reported only in the columns
(1) to (4). Note that the coefficients in the columns (1) and (2) are statistically insignificant at
5%, but significant in 10% level with t = 0.068 and t = 0.059 respectively. In contrast, I do not
observe any statistically significant effects for the areas with high ethnic concentration. These
results suggest that television might empower ethnic groups that do not have enough share in
local labor market, who were used to having little chance to be exposed to their own languages
outside their home.
38
39
Table 1.4: Effect of SABC exposure on language choice
Areas with low ethnic concentration Areas with high ethnic concentration
(1) (2) (3) (4) (5) (6) (7) (8)
Signal (dBu) of SABC 0.00046 0.000024 0.0000056 0.000057
(0.000024) (0.0000148) (0.0000040) (0.000409)
Signal stronger than 60dbu 0.0217?? 0.0251?? 0.000056 0.0013
(001085) (0.0077) (0.00992) (0.00639)
school FE Yes Yes Yes Yes Yes Yes Yes Yes
Year FE Yes Yes Yes Yes Yes Yes Yes Yes
Dep.Var.Mean 0.248 0.248 0.273 0.273 0.327 0.327 0.328 0.328
R2 0.751 0.751 0.699 0.698 0.607 0.607 0.610 0.610
N 13862 13862 20814 20814 50052 50052 43100 43100
? p < 0.05, ?? p < 0.01, ??? p < 0.001. Standard errors are clusterred by local-municipality level.
1.5.2 Results: Political attitudes
1.5.2.1 Election results
This subsection shows results on political attitudes. I examine the effect of television expo-
sure on voting outcomes, which are associated with ethnicity and national unity of South Africa.
Promotion of cultural diversity, shown as the main result in the previous subsection, may cause
unintended results that the country may lose their unity, because each cultural group may want to
diverge from national policies by strengthening their own identities. As the goal of the ?Rainbow
Nation? is to make cultural diversity the source of national unity, it is crucial to examine whether
the national unity is promoted along with cultural diversity. In my empirical framework, I expect
to observe an increasing vote share for a party which promotes the ?Rainbow Nation? policy
(i.e. ANC) if the broadcasts can balance cultural diversity and national unity simultaneously, and
decreasing vote share for the party if the broadcasts emphasize cultural diversity more than they
intended. Similarly, if cultural diversity is emphasized more than intended, I expect to observe
that the vote share for IFP increases through exposure to television broadcasts. For a main anal-
ysis, I concentrate on voting wards that have some vote shares for IFP, because IFP is the only
party which has strong attachment to a particular ethnic group.
Figure 1.14 reports the estimate coefficients and their confidence interval at 95% of equa-
tion 1.2, where the outcome variable is the vote share of ANC in the panel (a) and IFP for the
panel(b). 1,456 observations are used, after I drop voting wards with no or very small IFP shares.
I find that the increasing trend of vote shares after exposure to television broadcasts, though the
effects are insignificant until third election for ANC. Their vote share increased by 6.5 percentage
40
points in the four elections after exposure to television broadcasts. In contrast, I do not observe
any significant impacts on the vote share of IFP. The results indicate that television can success-
fully achieve its aim of balancing cultural diversity and national unity. Note that the programs
in SABC is often criticized by opponent parties and other mass-media, because they play a pro-
paganda role to spread ANC?s opinions broadcast news which are in favor of the views of ANC
(See, for example, Booysen 2011 and Plaut 2014). Even if this is the case, the economically and
statistically insignificant effects on the vote share of IFP suggests that exposure to television do
not stimulate favoritism for own ethnic group. Regression results are also reported in columns
(3) and (5) of Table 1.5.
(a) Vote share for ANC (b) Vote share for IFP
Figure 1.14: Effects of television exposure on vote shares for political parties
Note that the voting wards are fairly similar across elections, but there are some differences
in size and boundaries. Some districts split into multiple districts, and some other districts merge
when the number of voters are not that large. By including voting ward fixed effects in Figure
1.14, some districts are excluded from analyses if there is only one observation for a particular
ward. Therefore, in Figure 1.15, I report the results using the municipality fixed effects, instead
41
of the voting ward fixed effects. This increases the number of observations by approximately
10%. Results are very similar to those in Figure 1.14, both in terms of magnitudes and statistical
significance. Again, I report regression results in Table 1.5, columns (4) and (6) respectively.
(a) Vote share for ANC (b) Vote share for IFP
Figure 1.15: Effects of television exposure on vote shares for political parties, with municipality
fixed effects
Finally, I show results for ANC, using all states (i.e. not restricting voting wards having
some vote share by IFP) in Figure 1.16. There are 4,663 observations in total, included in the
political wards exposed to the television within the study periods and wards which do not have
any signal throughout the periods. Panel (a) shows the result using voting ward fixed effects, and
Panel (b) shows the results with municipality fixed effects instead of voting ward fixed effects.
As the number of observations significantly increase, estimate for t = t?s + 3 now becomes
statistically significant, while the coefficients are very similar to previous figures. Regression
results are reported in columns (1) and (2) in Table 1.5.
42
Table 1.5: Event Study: Effect of SABC exposure on language choices in elementary schools
Full observations Only IFP exist
Dep. var: Vote Share of ANC Dep. var: Vote Share of IFP
(1) (2) (3) (4) (5) (6)
t = t?s ? 2 -0.0051 -0.0004 0.00560 -0.0122 -0.0082 0.0056
(0.0279) (0.0212) (0.0172) (0.0160) (0.0253)
t = t?s ? 1 -0.0014 -0.0223 0.02159 0.0293 0.0195 0.0274
(0.0158) (0.0125) (0.0319) (0.0251) (0.0141) (0.0182)
t = t?s + 1 0.0086 -0.0008 -0.0035 -0.0009 0.0060 0.0141
(0.0155) (0.0133) (0 .0165) (0.0156) (0.0103) (0.0110)
t = t?s + 2 0.0275 0.0104 0.0190 0.0202 -0.0040 -0.0070
(0.0144) (0.0117) (0.0221) (0.0186) (0.0179) (0.0134)
t = t?s + 3 0.0591** 0.0514*** 0.0387 0.03171 -0.0004 0.0079
(0.0198) (0.0122) (0.0350) (0.0215) (0.0167) (0.0103)
t = t?s + 4 0.0737* 0.0722*** 0.0656* 0.0469* -0.0462 -0.0477
(0.0299) (0.0169) (0.0275) (0.0183) (0.0371) (0.0250)
Year FE Yes Yes Yes Yes Yes Yes
Political-ward FE Yes No Yes No Yes No
Municipality FE No Yes No Yes No Yes
R2 0.889 0.700 0.940 0.794 0.958 0.896
N 4,663 5,075 1,456 1,597 1,456 1,597
? p < 0.05, ?? p < 0.01, ??? p < 0.001. Standard errors are clustered by local-municipality level. All
specifications control for free-space signaling. Each explanatory variable takes 1 if a voting ward is exposed
to television at time t?s and time t is equal to t
?
s + k where t
?
s is the time of construction of new transmitter
affecting the ward s. A voting ward is classified as treated if signal at school is stronger than 60 dBu.
43
(a) Ward fixed effects (b) Municipality fixed effects
Figure 1.16: Effects of television exposure on vote share ANC: all states
1.5.2.2 Robustness
To check the robustness of the results, I again run same regressions with different cutoffs
of the treatments. The results are reported in Figure A.9 to A.14 in Appendix. Again, results
are fairly similar across different cutoffs, while the point estimates become smaller by lowering
cutoffs, as the lower cutoffs include areas with low signal coverage.
Second, I test whether the results are driven by changes in turnout rate. Vote shares would
change both due to the changes in turnout rate, as well as party switching behaviors by active
voters. To identify which of them caused the result, I estimate equation 1.2 with the dependent
variable of turnout rate, instead of the vote share. Figure 1.17 shows the result. Voters turnout
does not change after exposure to television broadcasts, indicating that the results are likely to be
driven by party switching behaviors.
In addition, I report results with traditional difference-in-differences model with equation
1.3. Results are reported in Table A.3, which are similar to the event study results. Similar to
Table 1.5, I report results with full available observations in columns (1) and (2), and results with
44
Figure 1.17: Effects of television exposure on turnout rate
voting words having non-zero voting shares for IFP in columns (3) to (6). Dependent variables
are the vote share for ANC in columns (1) to (4), and IFP in columns (5) and (6).
1.6 Conclusion
In this chapter, I examine impacts of exposure to television broadcasts on outcomes which
are associated with cultural diversity and national unity. While many countries and the interna-
tional society try to protect rights of the minority, balancing the idea of cultural diversity and
national unity involves challenges. I focus on television in South Africa, which tries to pro-
mote the national slogan of ?Rainbow Nation? to recover from the strict racial segregation during
Apartheid. Following this ?Rainbow Nation? policy, the government promotes cultural diversity
while emphasizing national unity and harmonization of different racial and ethnic groups.
To estimate impact of television broadcasts on cultural diversity and national unity, I first
digitize locations and features of television transmitters. I employ an engineering model to cal-
culate coverage of television and examine how exposure to television impacts two outcome vari-
45
ables: language choices in elementary schools and supports for a political party which promote
?rainbow nation? policy and a political party which is predominantly supported by single ethnic
group.
My findings show that successful delivery of the messages can promote cultural diver-
sity and national unity simultaneously. First, by investigating language choices in elementary
schools, where language is collectively chosen by individuals involved in local education, I find
that exposure to television increases usage of native languages in schools. The effect of tele-
vision increases usage of native languages by approximately 3 percentage points in three years
after the construction of transmitters, and around 4 percentage points after five to seven years.
In South Africa, parents demand English education instead of native languages because of the
negative images stemming from policies during Apartheid (NEPI 1992). My results suggest that
the television which repeatedly broadcast ideas of cultural diversity encourage parents and local
communities to use native languages instead of English.
Second, I examine whether exposure to television changes supports for political parties.
Emphasizing cultural diversity on television programs may generate unintended outcomes of
having cleavage between racial and ethnic groups. As an outcome variable, I first investigate
impacts on supports for the African National Congress (ANC), which promote the ?Rainbow
Nation? policy. I find that vote shares for ANC increased approximately 6.5 percentage points
in four elections after exposure to television broadcasts. On the other hand, I do not find any
significant effects on the vote share for Inkatha Freedom Party (IFP), which is predominantly
supported by one ethnic group with strong connections to local chiefs. Results suggest that
messages from television successfully increase support for the party which promotes cultural
46
diversity and national unity, but they do not generate further cleavage between ethnic groups.31
While my quantitative results are encouraging for policy makers who confront trade-offs
between diversity and unity, I should note that usage of mass media by governments should be
carefully monitored in general. Indeed, as the majority party, ANC, uses SABC as a tool to spread
their ideas, other media and political parties often criticize that there is too much influences of
ANC on SABC (Plaut 2014). Recent quantitative evidence shows that mass media could be used
to spread propaganda. In extreme cases, radio promotes supports for Nazis (Adena, Enikolopov,
Petrova, Santarosa and Zhuravskaya 2015) and amplify the genocide in Rwanda (Yanagizawa-
Drott 2014). My results do not support such forms of governmental involvements to mass media.
31To complement above observations, I also examine subjective attitudes toward ethnicity and nationality, using
geocoded data from Afrobarometer. However, due to the small number of observations with self-reported, unin-
centivized questions, I do not find any evidence on the changes in subjective attitudes. Results are reported and
discussed in Appendix A.1.
47
Chapter 2: Radio and Racism During Apartheid
2.1 Introduction
Race and ethnicity are used to categorize people and generate cleavages in some societies.
Understanding how people form or change their attitudes towards people with different identities
is essential to build peaceful relationships in such societies. Yet, most previous studies are either
correlational or in short-run laboratory experiments, and little is known about causal mechanisms
especially in the long-run. In this chapter, I focus on radio as one of the important channels
which has a potential to spread opinions to broad audience, and investigate its impact on racism
and ethnocentrism.
South Africa, during Apartheid, is an ideal setting to study the impact of radio as a tool
for both nurturing and repressing racism and ethnocentrism. Radio waves in South Africa were
monopolized by the central government, and the broadcasted programs had an explicit aim to jus-
tify racial segregation and ethnic polarization. That is, the radio aimed to promote racism for the
White population and promote ethnocentrism for the Black population. In addition, there were
two radio stations independent from the government within the country to counter Apartheid.
Utilizing geographical and time variation of exposure to radio during Apartheid, I examine the
impacts of exposure to anti-Apartheid radio on the voting behavior of White population for po-
litical parties which opposed to racial segregation.
48
Under the Apartheid regime in South Africa since 1948, White minorities (with approx-
imately 20% of the total population) systematically segregated other races. ?Separate develop-
ment? of each race and ethnic group was one of the central policies. The government considered
Black South Africans as ?separate national units on the basis of language and culture? (Rogers
1980, p.63). The government allocated small portions of lands as ?homelands? (or Bantustan) of
the Black population, and each Black individual belonged to one of these homelands associated
with their own ethnic identities. The government strictly restricted mobility both between ?White
areas and homelands?, and within ?White areas? (Pass Laws and Group Areas Act of 1950). The
African National Congress (hereafter ANC), which was an illegal organization playing the cen-
tral role of the resistance against Apartheid, strongly opposed to these acts, and pursued a ?single
multi-national society?.
Radio played an important role to promote the separate development policy. First of all,
formation of anti-Apartheid opinion through public media was strictly prohibited. Radio waves
were monopolized by the South Africa Broadcasting Corporation (hereafter SABC), which was
owned by the national government. Contents were heavily overseen by the government. Racial
segregation was justified, and issues negatively affecting the Apartheid regime were targets of
the censorship. For instance, uprisings in Black townships and international sanctions on racial
discrimination were not allowed to be broadcasted (Johnston 2019).
The monopoly of the radio waves by SABC lasted until 1979. In the 1970s and 80s, 4 out
of 10 homelands (Transkei, Venda, Ciskei and Bophuthatswana) declared ?independence? from
South Africa. From the viewpoint of the separate development policy, the central government
of South Africa supported homelands? independence. These states owned their own legislatures.
49
The central government could no longer oversee public media in homelands, due to their inde-
pendence. Consequently, two radio stations were established in 1979 and 1980 to avoid South
African regulations: Capital Radio 604 in Transkei and Radio 702 in Bophuthatswana. Since
they used medium waves (known as AM radio) which was suitable for long-distance propaga-
tion, people outside the homelands were also able to listen to these programs, and stations aimed
to do so. As the program director of Capital Radio 604 mentioned in its documentary (Johnston
2019), these independent stations explicitly aimed to convey unbiased news without influences
of the central government. Members of ANC also showed up to Capital Radio 604 to have open
discussion on anti-Apartheid opinions.1
Exploiting the geographical variation of the coverage of radio programs and the time vari-
ation before and after the establishment of independent radio stations, I examine effects of the
radio exposure on voting behaviors by White citizens during Apartheid as an indirect measure of
racism. I digitize the signal strength of radio waves utilizing engineering models (the simplified
version of Sommersfeld-Norton model (Trainotti 1990). Then, I rely on difference-in-differences
strategy to estimate the effects of radio exposure on vote shares for right-wing parties (HNP: Her-
stigte Nasionale Party before 1982 and CP: Conservative Party after then) and left-wing parties
(PP: Progressive Party before 1977, PFP: Progressive Federal Party between 1978 and 1982, and
DP: Democratic Party after 1982), which promoted and opposed to racial segregation relative to
the majority party (National Party). While vote share is an incomplete measure of racism, racial
issues were important in party politics throughout the history of Apartheid (Botha 1996). Espe-
1There were other possibilities of exposure to other media. I discuss TV, short-wave radio, and radio stations
outside the country in Appendix B.1.2.
50
cially for the election after the establishment of independent radio stations in 1981, the central
policy discussion was whether the country reforms Apartheid and mitigates racial segregation
(Charney 1982). Therefore, even if election results capture non-racial factors (such as labor mar-
ket policies, etc.), estimated results should also capture attitudes toward racial issues. I found
exposure to radio increased the voting share of anti-apartheid party by percentage points, and
decreased the share of pro-apartheid party by percentage points, respectively.
The main contribution of this chapter is to obtain causal estimates of mass-media, using
panel data, on changes in racial preferences in a relatively long periods. Contributions can be
divided into three streams of literature. The first contribution is on the growing literature on
the formation of racism and ethnocentrism. While the consequences of discrimination are well-
studied in various circumstances (see Bertrand and Duflo 2017 for survey), systematic under-
standing on the formation and amplification of such attitudes are still limited. As mentioned in
Depetris-Chauvin et al. (2020), most previous studies only report correlations, rather than causal
effects (e.g. Eifert, Miguel and Posner 2010, Robinson 2014, and Green 2018). Among recent
works uncovering causal process of formulating preferences towards in- and out-group members,
one of the most well-discussed channels is direct contacts. To name a few studies, Abel (2019)
finds an effect of forced resettlements and following interactions in South African policies during
Apartheid, Corno, La Ferrara and Burns (2018) investigates communications at dormitories in a
South African university, and Mousa (2019) studies playing soccer among different religions to
formulate preferences towards in-group and out-group members in Iraq. However, research on
the formation of inter-group preferences without direct interactions is limited, especially outside
laboratories. DellaVigna et al. (2014) (discussed in the next paragraph) and Depetris-Chauvin
et al. (2020) are some of a few papers studying the causal process of building national identities
51
without direct interactions: DellaVigna et al. (2014) investigate collective experiences of watch-
ing national football teams? victory.2 While their study analyzes the short-run impacts both in
terms of occasion of events (i.e. football games) and realization of outcomes (i.e. surveys just
after games), my project adds evidence on the mid- and the long-run by analyzing continuous
exposure of radio and its impacts.
Second, this chapter contributes to the literature on the role of mass media, namely tele-
vision, on preferences and behavior. Previous studies show impacts of exposure to radio on
genocide (Yanagizawa-Drott 2014), salience of ethnicity (Blouin and Mukand 2019), mitigating
conflicts (Armand et al. 2020), political knowledge (Stro?mberg 2004), voting for female can-
didates (Okuyama 2019), vote buying (Green and Vasudevan 2016), immigrant?s assimilations
(Russo 2021), anti-fascism resistances (Gagliarducci, Onorato, Sobbrio and Tabellini 2018), pop-
ulism (Wang 2020) and the emergence of a dictatorial regime (Adena et al. 2015).3 Among all,
DellaVigna et al. (2014) is the closest to my interest: they exploit exposure to radio from a con-
flicting country (Croatians receiving Serbian radio), and estimate impacts on nationalism as an
unintended consequence. Instead of exposure from rivals, radio stations aim to strengthen racial
and ethnic identities in my study. In addition, DellaVigna et al. (2014) relies on cross-sectional
variation in exposures of radio, while my study use a panel data.
2In gender context, Beaman, Chattopadhyay, Duflo, Pande and Topalova (2009) examine exposure to power-
ful female leaders on electoral gains in India, and Banerjee, Duflo, Imbert and Pande (2013) examine effects of
others? experiences. In addition, literature in social psychology emphasize several mechanisms to reduce biases
without direct interactions, which are well-summarized in Paluck and Green (2009). This body of literature in-
cludes consciousness-raising treatments (e.g. Dasgupta and Greenwald 2001) and perspective takings (e.g. Todd,
Bodenhausen, Richeson and Galinsky 2011) Lai, Marini, Lehr, Cerruti, Shin, Joy-Gaba, Ho, Teachman, Wojcik,
Koleva et al. (2014) compares 18 potential interventions which are believed to reduce biases. Yet, most research in
social psychology are conducted in laboratory setting and measuring the short-run outcomes. My contribution of
this chapter is to provide impacts on the long-run outcomes with socio-economic outcomes.
3In addition to radio, there are multiple studies examining impacts of other mass-media on various political and
economic outcomes, such as voting and political supports (DellaVigna and Kaplan 2007 and Di Tella et al. 2021),
women?s status (Jensen and Oster 2009), teen?s childbearing (Kearney and Levine 2015), and educational and labor
market outcomes (Kearney and Levine 2019).
52
Finally, this project also contributes to understanding the structure of Apartheid and its
legacies. Various works have discussed legacies on education (e.g. Van der Berg 2007 ), migra-
tion (e.g. Reed 2013) and health (e.g. Dinkelman 2017). There is a small number of qualitative
works on highlighting the importance of radio at the time of Apartheid and their legacies (Hamm
1991, Gunner 2002, Lekgoathi 2009 and Lekgoathi 2012). To the best of my knowledge, my
work is the first to present quantitative evidence on the roles of radio under Apartheid.
2.2 Background
2.2.1 Brief overview of Apartheid in South Africa
Apartheid refers to the set of policies regarding racial segregation by White population
against other races in South Africa, under the National Party?s regime between 1948 to 1990s.4
One of the main policy implemented under Apartheid was ?separate development?, meaning
?separate national units on the basis of language and culture? (Rogers 1980, p.63). Systematic
separate development emerged after Hendrik Verwoerd became Prime Minister in 1958, which
became the foundation of the justification of separated radio programs by ethnic groups (see
Section 2.2.2).
Under the separate development, the White government emphasized the importance of each
ethnic group. The government allocated small portion of lands to each ethnic group, which were
4Segregation and discrimination existed throughout the history of South Africa, including before the election
in 1948. After the independence of the Union of South African of 1910, the White national government enforced
multiple laws to segregate other races. This includes the South Africa Act of 1910 which ensured the monopoly of
political control by the White and the Native Land Act of 1913 that restricted the Black population to live in limited
areas. New laws after the National Party became the majority party include the Prohibition of Mixed Marriage Act
of 1949 and the Population Registration Act of 1950, which strictly classified citizens into four races (White, Black,
Coloured and Indian) See Kaplan et al. (1971) and Posel (1987) for detailed explanations on fundamental policies
before and during Apartheid.
53
generally remote, rural, and arid, and called these lands homeland or Bantustan. Every Black pop-
ulation belonged to one of these homelands depending on their ethnicity, and they were required
to ?go back? to their homeland, unless they had legal permission to stay in ?White areas?. Prac-
tically, the policy functioned to control population in urban areas and ensure provision of cheap
labor in each industry (Crush 1993), and millions of Black South Africans were forced to migrate
to their homelands due to the policy (Abel 2019). Due to the aim of ?separate growth?, the central
government pushed homelands to be self-governing and independent. Four out of ten homelands
declared independence in the 1970s and 80s (Transkei, Venda, Ciskei and Bophuthatswana).
2.2.2 Radios under Apartheid
Radio played an important role as a mass media under Apartheid. The government pro-
hibited television until 1975, and radio was the main source of information for South African
citizens. South Africa Broadcast Corporation (SABC) monopolized the broadcasting right, both
for radio and TV, which is discussed in detail in Section 2.2.2.1. Programs provided by SABC
were highly overseen by the government. However, independence of homelands allowed some
freedom of broadcasting in these areas. Therefore, two radio stations, independent from SABC,
were established in the homelands. I explain the details of these independent stations in It in
Section 2.2.2.2. Existence and roles of other media is discussed in Appendix B.1.2.
2.2.2.1 South Africa Broadcast Corporation: SABC
Radio programs broadcasted by SABC were exclusively through FM waves. Contents
were overseen by the central government, and issues negatively affecting the Apartheid regime
54
were not broadcasted, such as uprisings in Black townships and international sanctions (Johnston
2019). The censorship of radio programs stemmed from the Entertainments (Censorship) Act of
1931 to restricts films and other public media, which initially aimed to eradicate intermingling
of White and non-White population. The regulation became stricter when the National Party
became the majority in South Africa after 1948: they established eleven new laws to regulate
public media to maintain racial segregation.5 One of the interviewees who organized the Capital
Radio?s news program in Johnston (2019) stated that, ?the news that South Africa was hearing
at the time was not just censored by the government, but it was created by the government?.
In addition to the censorship to the news contents, popular and rock music were also regulated
(Hamm 1985). The government indicated that a ?high proportion of it [i.e. pop and rock music]
is morally unacceptable? (Annual Report 1967, p32).
SABC constructed transmitters across the nation in 1960s and 1970s. They provide pro-
grams in English and Afrikaans (two official languages used by the White population). In addi-
tion, SABC also developed programs targeted to the Black population, which followed the aim of
?separate development? to stimulate ethnic identities. As Hamm (1991) states, ?Black and white,
listen to its own radio service, theorized and programmed in accordance with state ideology.?
2.2.2.2 Independent radio stations
As mentioned in the previous sections, some homeland became independent in the 1970s
and 80s. Most contents in this subsection are based on the autographic documentary of the radio
station by Johnston (2019) or online interview to the engineers of the Capital Radio.
In the independent charter of Transkei, the first independent homeland, there was no men-
5See Kaplan et al. (1971) for the details of each legal restriction.
55
tion to how the newly independent states handle broadcasting. The National Party (the majority
party throughout the Apartheid) assumed that SABC could still operate in these territories. Re-
alizing the lack of a legal framework on broadcasting in Transkei, people who had gripe against
South African media censorship took advantage of the situation and decided to open radio sta-
tions in such independent Bantustan. Consequently, Capital Radio 604 in Transkei and Radio 702
in Bopthatswana became the only independent radio stations in South Africa under Apartheid
regime.
Capital Radio 604 initially targeted White elites, who did not have chances to be exposed to
non-censored news. News, music, and other talk programs were the main contents of the station.
News teams tried to tell the real news ?in an impartial and unbiased way?. For instance, when
the station discussed ANC, they avoid using ?terrorists?, which were used by South African gov-
ernment, and not ?freedom fighters?, which was used by ANC themselves, but used ?guerilla?
to make their news as neutral as possible. They also sometimes interviewed exiled ANC mem-
bers (e.g. Oliver Tambo in Stockholm), and avoided any sexism and racism throughout their
programs. Music programs were also distinctive from SABC?s program. Capital Radio tried to
play attractive songs including pops, jazz, R&B and rock, which were sometimes prohibited to
be broadcasted in SABC. Finally, there were some talk shows that promoted interracial under-
standings. For instance, a show called ?Capital School? got letters from students from different
races (who went to different schools under the Apartheid system) and introduced their school
activities. A listener described ?we never really interacted with other people, but when you lis-
ten to the radio station, everybody interacted together? in Johnston (2019)?s documentary. More
background information on Capital Radio is in Appendix B.1.1.
Since the program was prohibited by the government, official data on the actual listenership
56
does not exist. Advertising and Press Annual of Southern Africa, the complete list of mass media
in Southern African countries documenting advertisement costs, would give us some information
on listnership (Advertising and of Southern Africa 1980). According to the report, the spot rate
for Capital Radio was comparable to the regional radio stations operated by SABC. The spot
advertisement rates for ?Prime 1 Class? programs at the Capital radio was R33, R50, and R100
per 30 seconds for 6am to 9am on Monday-Friday, 8am to 2pm on Saturday, and 8 am to 3pm
on Sunday respectively.6 For a similar schedule, Radio Port Natal, the regional radio operated
by SABC around Durban area (which is a subset of Capital Radio?s coverage) charged R10 and
R15 for 6am to 8:30am on Monday-Friday, 8:30am to noon on Saturday and Sunday respectively.
Radio Good Hope, the regional radio operated by SABC around Capetown charged R17 and R26,
and Radio Highveld served around Johannesburg and Pretoria charged R50 and R88 for 6am to
8:30am on Monday-Friday, 8:30am to noon on Saturday and Sunday respectively.
Radio 702 was established in Bopthatswana, which is very close to the capital of South
Africa, Pretoria. The station was mainly aimed to play music, and it is said to be less political
compared to 604 (Drewett 2004). In my main analysis, therefore, I focus on the effect of exposure
on Capital Radio 604.
2.2.3 Politics in South Africa during Apartheid
South African politics during Apartheid was monopolized by the White population. Other
racial groups (Black, Coloured and Asians) did not have voting rights, and policies were deter-
mined by the White population. Racial issues were important in party politics throughout the his-
6R1 (one South African Rand) on April 1st, 1980 was approximately USD1.23.
57
tory of Apartheid ( Charney 1982, O?meara 1996 and Botha 1996).7 South African politics during
Apartheid is characterized by three different major groups: the National Party (NP), right-wing
parties, and left-wing parties. Throughout the history of Apartheid, NP had been the majority
party. Right-wing parties (Reformed National Party (HNP, 1969-1981) and Conservative Party
(CP: 1982-)) consistently claimed to promote stricter racial segregation aims, and left-wing par-
ties (Liberal Party (from 1953-1958), Progressive Party (1958-1976), Progressive Federal Party
(PFP: 1977-1988), and Democratic Party(DP: 1989-)) consistently claimed to weaken racial seg-
regation policies. Finally, United Party (UP: -1977) and New Republic Party (NRP: 1977-) were
modest minority parties that did not oppose to white minority rule but opposed to strict separation
through Apartheid. There Figure 2.1 shows the voting shares by party and year after 1960.
Figure 2.1: Political supports by year during Apartheid
7 Wilse-Samson (2013) interpret Apartheid as two strategic concerns: the long-run worry about being overpow-
ered by the native majority and politically powerful mining and farm sectors with their competing demands for black
labor.
58
In 1978, NP decided to reform policies when Willem Botha became the Prime Minister
to weaken strict racial segregation to counter international sanctions. They planned to legalize
interracial marriages and multiracial political parties. The Prime Minister Botha also visited
Black township to learn issues in Black communities. As a consequence, UP and NRP lost their
political power and integrated to NP, while supports to right-wing parties significantly increased
in the 1981 election.8 After the election in 1989, the National Party decided to allow all races to
have voting rights. African National Congress (ANC), which was previously illegal and classified
as a terrorist group, also became a legal political party.
2.3 Data
2.3.1 Exposure to radio
To obtain coverage of radio, I need to follow different strategies to measure AM and FM
radio respectively. For FM radio (very high frequency waves), topographical variation and power
of transmitters are the main sources to determine coverage (see Section 1.3.1 in Chapter 1 for
more explanations). On the other hand, AM radio (medium wave) is not affected by topographical
features. Instead, ground conductivity, in addition to electric power of transmitters, is a key factor
to determine distance of transmission. I follow Russo (2021) to predict coverage of AM radio,
and use the Longly-Rice model, which is widely used in the economic literature for FM radio(see
among others, Yanagizawa-Drott 2014).
Medium Wave radio uses lower frequencies than FM radio: in Europe, Asia and Africa,
AM radio are assigned frequencies between 531kHz to 1602kHz. There are two representative
8Establishment of the Tricameral Parliament is a good example of each party: PFP opposed to the referendum
because of the exclusion of Blacks, and CP opposed to it because of the inclusion of Indians and Coloured.
59
ways of propagations: ground-wave and sky-wave. Ground-wave propagation is highly affected
by the condition of soils, which is called ground conductivity. Sea water has the best ground con-
ductivity, agricultural land and rich soil (such as forest with moisture) have good conductivity,
while dry land and ice have bad conductivity.9 To calculate AM radio wave, I use the simpli-
fied version of Sommersfeld-Norton model, Trainotti 1990, firstly used in economic research
by Russo 2021. The detailed model that I employ is in Appendix B.2. Here, I summarize my
procedure.
First, I collect the data on radio transmitters from interviews to the engineers of Capital
Radio 604.10 I then digitized data on ground conductivities, acquired from CCIR (1992).11 The
original map from CCIR and the digitized map is shown in Figure 2.2.
(b) digitalized map
(a) Original map
Figure 2.2: Ground Conductivity in South Africa
Second, I use a simplified version of Sommersfeld-Norton model (Trainotti 1990) to calcu-
9See CCIR (1992) for differences in ground conductivity by soil qualities.
10I thank a lot to Mr. Dave Cherry and Mr. Craig Johnston for sharing their data and experiences.
11CCIR stands for Consultative Committee on International Radio, which was an international organization to
manage radio waves. Currently, the most of the functions of CCIR moves to ITU, international Telegraph Union.
60
late predicted areas of the radio coverage from the independent radio stations. Figure 2.3 shows
the predicted areas.12 I exploit the variation of this ground-wave propagation in my main analy-
ses.
Figure 2.3: Predicted AM radio coverage, using Sommersfeld-Norton model
Sky-wave propagation is only available in the nighttime, and also highly affected by season
and sunspots. I do not measure and estimate impact through sky-wave propagation and concen-
trate my analyses on ground-wave propagation. This is because the availability of medium wave
radio is unstable for sky-wave propagation, which fluctuate a lot day-by-day depending on solar
activities. In addition, the sky-wave could propagate much further distance than the ground-wave.
If my outcome variables were available at the daily or monthly level, I could use time variation
in sky-wave propagation. However, since my main outcome variable (political behavior during
12As AM radio is not affected by topographical features but the ground conductivity (which does not have lo-
cal geographical variation), prediction from the engineering model becomes similar to distance to the transmitter.
Therefore, my main results become similar to those that I define the treatment variation as distance to transmitters
(See Figure B.3). However, it is not obvious to determine the distance measure of treatment without having some
prediction model of the coverage of radio. Therefore, having engineering model to predict the coverage is still useful.
61
Apartheid) is only available once in 4-5 years, radio coverage of the sky-wave propagation is
averaged out throughout the country. Therefore, my empirical results could be underestimated if
people also listened to the radio through sky-wave propagation, because those who could not lis-
ten to the ground-wave (in ?control regions?) could potentially listen to the radio during limited
times through sky-waves.
2.3.2 Voting outcomes during Apartheid
In terms of the racism among the White population, I rely on electoral results before and
after establishments of the independent radio (Capital Radio 604). Using electoral-district level
data from 1961 to 1981 for national elections, I estimate the impacts of exposure to different types
of radio on election results by difference-in-differences. I use the digitized data on the election
outcomes by Wilse-Samson (2013).13 The data contains electoral outcomes by voting districts.
Each voting district has approximately 10,000 voters, and there are around 180 districts in total.
As explained in Section 2.2.3, there are two main opponents to the National Party, one to its from
right and the other to its left. I construct the share of right-wing votes and left-wing votes, and I
also analyze the absolute number of voters.
13I thank much Laurence Wilse-Samson for generously sharing the data with me. He digitized Government
Gazette to identify locations of voting districts and vote shares for each party. Note that elections during Apartheid
were available only for White population.
62
2.4 Identification strategy
To identify impacts of exposure to radio on White racism under the Apartheid regime, I
estimate the following equation:
V otest = ?AreaExposeds ? Postt + ?s + ?t + ?st, (2.1)
where V otest is a variable representing voting outcomes of various political parties for loca-
tion s at time t. The first main outcomes are the vote shares for PFP (left-wing party opposed to
Apartheid) and HNP (right-wing party which opposed to policy reform of NP, the majority party).
In addition, I also estimate effects on the actual number of voters for each party. AreaExposed
takes 1 if a district is covered by the anti-Apartheid radio (i.e. radio exposure from the indepen-
dent radio station), and Postt takes 1 if time t is after the construction of transmitters. In the
main analysis, I assume that a location s has strong enough signal if the predicted field strength
is larger than 60dBm (decibels-milliwatts). Results with other cutoffs are presented in the Ap-
pendix B.3.1 with similar results.14 ?s is location-specific fixed effect, ?t is a time fixed effect,
and ?st is an idiosyncratic error term.
Equation (2.1) is a standard generalized difference in differences model. The causal effect
of radio exposure is identified under the exogenous arrival of the treatment. Areas with radio
exposure would have displayed changed voting pattern like the one in areas without them in the
absence of the access to the radio station studied. Tests to verify the existence of the parallel
14As Trainotti (1990) points out, urban areas require the highest electric power (88dBm), rural areas require
smaller electric power (54dBm), and residential areas are in between the middle. Since it is not obvious how to
assign urban, residential and rural areas are in, I select arbitrary cutoff.
63
pre-trend are presented in Section 2.5.1. In addition, since the exposure of radio is concentrated
in the East Coast of the country, estimated results may confound shocks which correlated with
treatment. While it is impossible to examine all possible events, I examine two likely confounding
shock in Section 2.5.2 and show these shocks were not the causes of my results.
2.5 Results
Figure 2.4 graphically plot the vote share for each party, comparing treated and untreated
voting districts. Shaded areas correspond to the 95% confidence interval. Black solid line corre-
sponds to the treated areas (i.e. areas with coverage of radio), and red dashed line corresponds
to untreated areas. Black vertical line (in year of 1979) is the time Capital Radio 604 started
operating. Graphs show that treated areas are more liberal than untreated areas on average, but
the voting trend is parallel in pre-treatment periods. The trend diverges when areas close to the
transmitters were exposed by the radio: vote shares for the left-wing party increased more in the
treated areas than the control areas, and vote shares for right-wing parties did not increase in the
treat areas as much as the control areas.
Table 2.1 shows the results estimating Equation (2.1). As outcome variables, Panel A
presents results with proportion with each political party Panel B presents the results with the
actual number of voters, and Panel C presents the results with the log of actual number of voters.
Each column shows a result with left-wing party, right-wing party, middle-left wing party and
combined left-wing parties for Column (1) to (4) respectively. I present the results with different
cutoffs for the radio coverage in Appendix B.3.1. Note that the number of observations are
smaller in column (2), because there were no right-wing party in the election of 1961.
64
(a) proportion of left-wing party (b) proportion of right-wing party
(c) proportion of middle-left party (d) proportion of left wing parties, combined
Figure 2.4: Proportion of parties by location and year
According to these simple difference-in-differences analysis, exposure to radio has effects
on political behaviors associated with racial preferences at the most left and right wings: as we
see in the Panel A, (1) the vote share for the left wing party increases by 9.7 percentage points for
the left wing party and (2) decreases by -9.7 percentage points for the right wing parties, but no
effects on mid-left party (which had similar political agenda as the majority party but slightly left-
side ideology) and combination of left-wing and mid-left parties. The results in Panel B basically
confirm results in Panel A, although I also observe a statistically significant positive effect on
the combination of left-wing and mid-left parties in column (4). Taking log of the number of
65
Table 2.1: Difference in differences: Effects of AM radio on voting behavior
Treated if the signal strength is higher than 60dBu
(1) (2) (3) (4)
Left-wing Right-wing Middle-Left All-left
Panel A: Results with proportion of votes
Areas >60dBu? 1981 0.0965?? -0.0968??? -0.0766 0.0199
(0.0314) (0.0201) (0.0461) (0.0440)
Location FE Yes Yes Yes Yes
Year FE Yes Yes Yes Yes
R2 0.700 0.667 0.738 0.833
N 801 651 801 801
Panel B: Results with the actual number of voters
Areas >60dBu? 1981 917.5?? -789.7??? 746.8 1664.3??
(315.8) (184.9) (479.8) (526.6)
Location FE Yes Yes Yes Yes
Year FE Yes Yes Yes Yes
R2 0.726 0.672 0.657 0.768
N 801 651 801 801
Panel C: Results with the log of actual number of voters
Areas >60dBu? 1981 1.554? -0.0428 3.312??? 2.830??
(0.700) (0.768) (0.930) (0.948)
Location FE Yes Yes Yes Yes
Year FE Yes Yes Yes Yes
R2 0.667 0.643 0.593 0.587
N 801 651 801 801
? p < 0.05, ?? p < 0.01, ??? p < 0.001 Dependent variables are the vote shares for
each type of political party. Areas > 60dBu ? 1981 takes one if the voting district
has strong enough signal and year is 1981.
voters by party in Panel C, the effect of the radio on the right wing party diminished. This is due
to the underlying differences in the voting share for the right-wing party: the right wing party
obtained less votes in the treated areas before the construction of radio towers (see Figure 2.4,
so that while we can observe the absolute changes in Panel B, we do not observe any substantial
differences in the log-scale. The number of observations is smaller for column (2), because the
right-wing party did not exist in the first election (1961) in my sample. I restrict the sample to
after 1970 so for column (1), (3) and (4), and results are unchanged.1516 In addition to the results I
15Results are in Appendix B.3.2.
16Previous literature such as DellaVigna et al. (2014), Adena et al. (2015) and Wang (2020) calculate persuasion
rate of the radio programs. That is, how much of the audience changed their attitudes due to the exposure of radio.
66
present here, I also test whether the voter turnout increases by the exposure to the radio. Running
same regression as in Equation 2.1 with the outcome variable of voter turnout rate, I find the
turnout increased by 7.3 percentage points, which is statistically significant in 1 percent level.
The regression excludes the data from 1960, because the data on voter turnout is noisy and there
is some difference in turnover rate in 1960. After excluding 1960 data, the turnout rate seems
well balanced in pre-treatment periods. The trend of turnover rate is shown in Figure B.4. It is
not possible to identify characteristics of the marginal new voters due to the limitation of the data.
Yet, it is reasonable to believe that some of the new voters, who decided to vote after exposures
to the radio, were more likely to vote for the left-wing party and middle-left wing party, and this
contributed to the main reported in Table 2.1.
2.5.1 Testing identification assumptions
Since the Equation (2.1) is a two-way fixed effect model, existence of parallel trends be-
tween treated and control areas in outcome variables is necessary to identify ?. To test the exis-
tence of parallel pre-trends, I firstly run Placebo tests, by restricting samples to the pre-treatment
periods (i.e. excluding observations for 1981, and use the observations before then). I assume
1977 and 1974 are the placebo timing of the construction of radio transmitters, fixing the loca-
tions of treated areas, and run same regressions in Equation (2.1), shown in Table 2.1. Results of
the Placebo tests are shown in Table 2.2. Panel A shows the results assuming the construction of
transmitters occurred in 1977, and Panel B shows the results assuming the construction occurred
in 1974. That is, the treatment variable AntiRadiost takes 1 for treated areas in 1977 in Panel
It requires public survey data on the listenership of radio. Unfortunately, I do not have reliable public survey data on
the listenership of Capital Radio, and therefore I am not able to estimate the persuasion rate.
67
A, and in 1974 and 1977 in Panel B. Overall, results do not show any significant effects both in
terms of magnitudes and standard errors.
Table 2.2: Difference in differences:Placebo test Treated if the signal strength is
higher than 60dBu
(1) (2) (3) (4)
Left-wing Right-wing Middle-Left All-left
Panel A: Placebo exposure in 1977
Pracebo Exposed in 1977 0.0344 -0.000765 0.00459 0.0390
(0.0273) (0.00640) (0.0521) (0.0488)
Location FE Yes Yes Yes Yes
Year FE Yes Yes Yes Yes
R2 0.711 0.865 0.754 0.835
N 636 486 636 636
Panel B: Placebo exposure in 1974
Pracebo Exposed in 1974 0.0122 0.00395 0.0295 0.0417
(0.0241) (0.00659) (0.0460) (0.0431)
Location FE Yes Yes Yes Yes
Year FE Yes Yes Yes Yes
R2 0.710 0.865 0.755 0.835
N 636 486 636 636
? p < 0.05, ?? p < 0.01, ??? p < 0.001. All regressions use pre-treatment periods only
(i.e. excluding observations from 1981). Panel A assumes the placebo treatment timing
as 1977, and Panel B assumes the placebo treatment timing as 1974, fixing the actual
treated districts.
I also test the parallel trend by estimating the following equation, which is the event-study
regression, ?
V otest = ?tExposeAreas ? Y eart + ?s + ?t + ?st, (2.2)
t
where ExposeAreas is the time invariant dummy taking 1 if a voting district has a potential ex-
posure to radio. If the parallel trend holds, ?ts should be equal to zero for all the pre-treatment pe-
riod, as ?s absorbs the time invariant locational differences in political preferences, and ?t absorb
the secular time effects. Results in Table 2.3 confirms the existence of parallel trends between
68
groups: the coefficients of ?t are not large in absolute value and never statistically significant.
Table 2.3: Test of parallel trend assumption: Treated if the signal strength is higher
than 60dBu
(1) (2) (3) (4)
Left-wing Right-wing Middle-Left All-left
Dummy of 1974 ? treated 0.00660 0.00602 0.0579 0.0645
(0.0619) (0.0299) (0.0674) (0.0984)
Dummy of 1977 ? treated 0.0259 0.00361 0.0343 0.0602
(0.0619) (0.0299) (0.0674) (0.0984)
Dummy of 1981 ? treated 0.101 -0.0937?? -0.0357 0.0648
(0.0619) (0.0299) (0.0674) (0.0984)
F(2, 643) 0.0958 0.0205 0.372 0.265
R2 0.137 0.308 0.396 0.212
N 651 651 651 651
? p < 0.05, ?? p < 0.01, ??? p < 0.001. F-test examines Dummy of 1974 ? treated and
Dummy of 1977 ? treated are jointly Zero.
2.5.2 Robustness check and discussion on potential concerns
While the parallel pre-trend seems to hold for estimating equation (2.1), the coefficient of
the interest (?) may be biased if the time variant error term (?st) correlates with the treatment.
That is, the estimated coefficient may be driven by other shocks which occurred at the same time
of the construction of transmitters. I test now the two main potential confounding shocks as
follows.
Time variant province-level policy differences: First, as shown in Figure 2.3, the loca-
tion of transmitters are concentrated to the east coast of the country, which are mainly in Natal
province and the north-east part of Cape province. Any policy differences between provinces,
which may be time variant, may influence my main results. As there were only four provinces
during Apartheid (Cape, Natal, Orange Free State and Transvaal), I include province-by-year
69
fixed effect to equation (2.1). Results are presented in Table 2.4. All estimated coefficients are
very similar to what those found in the main specification. Therefore, I claim that my main
findings are at least not driven by the time variant province-level differences.17
Table 2.4: Robustness check: main results with state-year FE Treated if the
signal strength is higher than 60dBu
(1) (2) (3) (4)
Left-wing Right-wing Middle-Left All-left
Areas >60dBu? 1981 0.0917? -0.0637?? -0.0924 -0.000682
(0.0389) (0.0242) (0.0574) (0.0553)
Location FE Yes Yes Yes Yes
Year FE Yes Yes Yes Yes
Province-Year FE Yes Yes Yes Yes
R2 0.714 0.697 0.746 0.836
N 800 651 800 800
? p < 0.05, ?? p < 0.01, ??? p < 0.001.Dependent variables are the vote shares for
each type of political party. Areas > 60dBu ? 1981 takes one if the voting district
has strong enough signal and year is 1981.
Differential Labor demand by location: One remarkable changes that the previous liter-
ature found around the time that some homelands became independent (so that independent radio
stations were constructed) was a significant changes in the labor demand, especially for Black
labors in mining sectors. During 1970s, South Africa experienced significant decline in min-
ing workers from neighboring countries. At the time, the mining sectors relied a lot on foreign
mining workers, but they switched to domestic workers due to the following negative labor sup-
ply shocks. First, President Banda of Malawi decided to stop sending workers to South Africa
in 1976, following a plane accident shipping Malawian workers. Second, South African min-
ing organization failed to recruit mining workers from Mozambique due to the independence of
17There is one electorial district called ?hospitaal? in 1961 election, which I could not identify its geographical
location, and therefore I could not identify the province for this district. One observation is dropped in Table 2.4
compared to the main analyses in Table 2.1. Since there are approximately 180 electoral districts, it is impossible to
include smaller geographical unit to control location-year fixed effects. For example, there are approximately 200
local municipalities in South Africa.
70
Mozambique from Portugal after 1975. These significant decline in labor supply from neighbor-
ing countries are well documented in Crush (1993).
Wilse-Samson (2013) uses recruitment policies in South Africa, that the recruitment for
mining workers were only permitted in certain districts, and identify impact of labor supply
shocks on voting behaviors. In rural areas where mining recruitment were permitted, there was
a tension between agricultural sectors that was the main sector hiring black workers and mining
sectors which newly demand workers to substitute foreign labor. Figure 2.5 shows the areas
which permitted recruitment for mining sector, according to Crush (1993). Large portion of the
areas which have radio coverage in Figure 2.3 is overlapping with the areas which permitted
mining recruitment in the East Coast areas. Moreover, the timing of the labor supply shock is
similar to the timing of the construction of transmitters: the labor supply shock mainly occurred
between 1976-1979.
To address this potential confound, I restrict the sample with the areas which permitted
mining recruitment. This significantly reduce the number of observations, but could allow us to
see if the labor supply shock was the potential confounding factor or not. First, Table B.7 is a
modified replication of Wilse-Samson (2013): I consider areas allowing the mining recruitment
are the treated regions, and areas prohibiting the recruitment as the control. While Wilse-Samson
considers the outcomes in 1977 and 1981 were treated, I modified the treatment timing only
for 1981 to make a comparison with my original findings. It is noteworthy that the modified
replication of Wilse-Samson looks very similar to my results, both in terms of magnitudes and
direction of the effects. This keeps the concern that my results are driven by other factors, not the
exposure to radio itself.
Therefore, I restrict the sample for areas which permitted the mining recruitment. That is, I
71
Figure 2.5: Areas which allowed mining recruitment from Crush (1993)
72
compare areas with and without the radio exposure only in areas allowing the mining recruitment.
The results are presented in Table 2.5. While the number of observation declined a lot due to the
trimming of the sample, results are very similar to the main results. This indicates that at least
the labor supply shock is not the confounding factor to determine my main results.
Table 2.5: Difference in differences: only with open districts
(1) (2) (3) (4)
Left-wing Right-wing Middle-Left All-left
Areas >60dBu? 1981 0.0930? -0.0703??? -0.0711 0.0219
(0.0375) (0.0201) (0.0599) (0.0532)
Location FE Yes Yes Yes Yes
Year FE Yes Yes Yes Yes
R2 0.750 0.679 0.744 0.846
N 321 261 321 321
? p < 0.05, ?? p < 0.01, ??? p < 0.001 Dependent variables are the vote shares for
each type of political party. Areas > 60dBu ? 1981 takes one if the voting district
has strong enough signal and year is 1981.
2.6 Concluding remarks
In this chapter, I investigate the impact of mass media on racism and ethnocentrism in South
Africa during and after the end of Apartheid. Utilizing the panel data, difference-in-differences
analyses suggest that exposure to radio with anti-Apartheid messages increased the vote shares for
a political party which opposed to Apartheid, and decreased the vote shares for a political party
which aimed to promote stricter racial segregation. In addition, using cross-sectional variation in
radio listenability which are determined by plausibly exogenous topographical variations, I find
that areas with the coverage of nationally-owned radio shows lower the vote share for an ethnic
party after the end of Apartheid for Black population. This result suggests that messages after
the end of Apartheid to unify different ethnic and racial groups dominated the messages during
73
Apartheid to promote ethnic division.
There are still multiple concerns in my analysis, which is going to be future tasks. First,
since the unit of analyses are the political-ward level and the number of observations is limited,
clustering by regional level is infeasible. Second, for the analysis during Apartheid, it would be
more interesting if I could include the data on the coverage of SABC radio, which was under the
governmental censorship and functioned as a governmental propaganda. This will become possi-
ble once I gain an access to SABC?s annual reports, which are stored in the headquarter of SABC.
Current travel restriction hampers the author?s travel. Finally, to estimate effects of SABC?s ra-
dio on ethnic preferences (and other outcomes) for Black population, outcome variables during
or just after the end of Apartheid play key roles. However, due to the data limitation, the election
data from 2000 is the best data I could access so far. Future disclosure of the election results
in 1994 or other potential variables (such as education) may enable us to evaluate the impact of
SABC radio on Black population during Apartheid.
74
Chapter 3: Labor Market Evolution in an Emerging Industry: Japanese Cotton
Spinning Industry before 1900
3.1 Introduction
The wages of labour are the encouragement of industry, which, like every other
human quality, improves in proportion to the encouragement it receives. Where
wages are high, accordingly, we shall always find the workmen more active, diligent,
and expeditious, than where they are low. (Smith 1776, pp. 99)
[H]ighly paid labour is generally efficient and therefore not dear labour. (Mar-
shall 1920, pp.294)
The views above, expressed 250 and 100 years ago, respectively, by two of the founders
of modern economic science have received renewed attention in recent studies that examine the
role of economic growth in raising income levels, alongside the reverse relationship whereby low
growth results in high poverty and high poverty in turn leads to low growth (see, e.g., Perry 2006).
While the possibilities stemming from virtuous circles between growth and poverty reduction
have thus been recognized for a long time, we still have only limited micro-level understanding of
the mechanisms that generate such virtuous circles, in particular, the roles played by competition
and firm heterogeneity through the process of industry evolution.
75
At the same time, a rich strand in the literature has examined the structure of labor mar-
kets, wage heterogeneity across firms and their wage-setting power. Most papers analyze wage
setting powers in developed countries: some studies labor markets requiring job-specific skills
( e.g.,school teachers by Ransom and Sims (2010) and Falch (2010) and nurses by Sullivan
(1989), Staiger, Spetz and Phibbs (2010), and Matsudaira (2014)) In addition, recent papers
consider labor market concentration in general (e.g. Lamadon, Mogstad and Setzler 2022 and
Rinz 2022). However, analyses focusing on developing and emerging contexts are still limited
(e.g.Brummund 2012, Naidu, Nyarko and Wang 2016). Moreover, the applications of this anal-
ysis to long-term issues of industry growth and development are still hard to come by in part
because of the lack of suitable data.
In this chapter, we utilize an unusually rich historical data set on firms, establishments,
workers and wages in the Japanese cotton spinning industry circa late 19th to explicitly link the
evolution of an industry to labor market outcomes. The industry represents an ideal case study for
this purpose because it was the first modern mechanized industry to emerge in the country that
had been isolated from the rest of the world for over two centuries. The industry also exhibited
spectacular growth, employing hundreds, then thousands and eventually hundreds of thousands
of factory workers. The growth was accompanied by increased productivity and workers wages
both of which, however, exhibited big variations across time and across firms and establishments
at any given point in time, as the industry went through the classical industry evolution stages (see
Gort and Klepper 1982)the initial stage with a few firms operating under high technological and
market uncertainty; the large-scale entry stage based on the industry coalescing around a winning
technological paradigm; the shake-out and market consolidation stage; and the maturation stage.
Throughout this whole process, at the end of which the industry emerged as a leading industry
76
in Asia and one of the leading industries in the world and which took about four decades, we are
able to bring together detailed data on the census of firms and establishments that populated the
industry.
The success story of the Japanese cotton spinning industry, including at the micro firm-
and establishment-levels has been documented in previous studies, including work by one of the
present authors (see, e.g., Saxonhouse 1974, Braguinsky, Ohyama, Okazaki and Syverson 2015),
but the aspects dealing with the market for factory workers and their wages are still waiting to be
explored. One common perception articulated in the literature is that the Japanese achievement
has not caught the attention of celebratory historians: the pain and labor that made it possible.
(Landes 1998, p. 381).
The only attempt we are aware of to utilize Japanese firm-level microdata to examine la-
bor outcomes in Japans cotton spinning industry is Saxonhouse (1976) and Saxonhouse and
Kiyokawa (1985). Notably, Saxonhouse and Kiyokawa (1985) conclude that [c]hanges in the
quality and organization of the labor force in Japan?s early industrialization in textiles appear to
be responsible for the bulk of the improvement in this industry?s total factor productivity at the
end of the 19th century and during the first three decades of the 20th. A better-educated, more
experienced labor force, working shorter hours, transformed an industry that initially could not
compete with Indian yarns even in its home market into the major force in world textile mar-
kets. (p. 285) They attribute most of the changes to exogenous factors, however (government-
sponsored primary education, international pressure that led to a reduction in working hours and
the natural process of increased worker experience as the industry matured). While these factors
were no doubt important, we focus on analyzing the interaction between increasing maturization
of the industry and formation of labor market measured by wages.
77
Our main findings can be summarized as follows. First, at its onset, in the early- to mid-
1880s, the industry, comprised mostly of small-scale government-sponsored mills, employed
labor practices such that factory workers, both male and female, were recruited from local social
networks and treated like family labor. While such practices can help mitigate information asym-
metry and the moral hazard problem, they are also known to be inefficient (e.g., Montgomery
1991; Heath 2018). In line with that, early Japanese cotton spinning mills were struggling to
make ends meet with no prospects of growing or putting a dent in the market dominated by
imports.
Second, the start of the industry transformation after the government abandoned its promo-
tion measures in 1886 was marked by entry by larger-scale, independent private entrants whose
initial labor practices in the late 1880s-early 1890s resembled the Lewis (1954) and Harris and
Todaro (1970)?s dual economy model. The new profit-motivated firms abandoned the master-
vassal notion of smaller mills and turned to recruiting cheap labor from impoverished urban and
suburban slums (Tsurumi 1990) and later from more remote rural areas. As we describe in more
detail in Section 3.2, firms generally started recruiting workers from areas close to their plants,
and then opened recruitment offices in rural areas further away once they could not find any ad-
ditional workers from their original locations. This recruitment process allowed firms to employ
a large number of unskilled workers with low wages.
Third, as criticisms of the Harris-Todaro model points out, the situation with the reserve
labor army was not one of a long-run equilibrium. It took only about five years for the nature of
the equilibrium to start changing in the Japanese cotton spinning industry. As the early profit-
seeking firms overcame industry-wide technological and market bottlenecks (see, e.g., Braguin-
sky and Hounshell 2016), a wave of new entrants followed, draining the reserve army and forcing
78
firms into fierce competition for workers, especially skilled workers.1 Due to the evolution of the
output market, firms had an incentive to poach skilled workers from other firms for efficient
production(Okamoto 1993). The poaching behaviors are generally observed between firms with
neighboring locations, and existence of neighboring firms determines cost of acquiring skilled
workers. Using geographically highly granulated data, we show that increased demand for skilled
workers dramatically changed the nature of the labor market. We investigate effects of increased
competition, measured by expansion of neighboring firms, on wages. Evidence indicates that the
mechanism behind these changes was intra-industry worker raiding and that a handful of firms
(many of which were the early independent private entrants) eventually came on top in this com-
petition for higher-quality labor. We find that expansions in peer firms? capacities by one standard
deviation from the mean increases the wages of focal firms for male workers by 7.1 percentage
points and for female workers by percentage points.
Cheap, low-quality labor and overall inefficiency are considered to be key features of de-
veloping nations. As mentioned, the first profit-seeking firms in Japans cotton spinning industry
fit this mode. However, the situation changed dramatically over time. The in-depth examination
of the process of transition to a large-scale, multi-plant production accompanied by better-quality
and higher-paid workforce we conduct in this chapter should foster our understanding of the in-
ner mechanisms that can generate virtuous circles in economic development.
In Section 3.2, we summarize contextual background of the cotton spinning industry before
1900. Section 3.3 introduces the main data source, as well as newly constructed data, such as
1Clark (1987) maintains that most textile jobs did not require great skill. This is only partially true as expensive
imported spinning machinery was complex to operate and required mastering of brand-new technologies (Tsurumi
1990), making experienced, skilled workers a crucial factor of improved productivity, as we demonstrate below.
79
geographical locations of firms. Section 3.4 examine formation of the labor market through the
evolution of the industry using historical documents and descriptive data. Section 3.5 provides
the regression analyses to examine the effect of local competition on wages, and Section 3.6
concludes the chapter.
3.2 Background
3.2.1 Japan?s Economy before 1900 and the Cotton Spinning Industry
Until 1883, Japan had literally no large-scale modern industry. In Tokugawa period which
ended in 1868, the government highly restricted communication and trade with foreign countries
(which was called Sakoku, the self-isolation policy). After the government decided to introduce
Western culture and industries, the government firstly opened government-run factories. The
cotton spinning industry was one of those.2 Starting from Kagoshima Boseki in 1867, the gov-
ernment opened or helped opening 19 mechanized small-scale cotton spinning factories across
the country with the capacity of 2,000 spindles each with machines imported from the United
Kingdom. Yet, they failed to deliver a competitive product due to the lack of competition and
skills. An early difference-maker did emerge, however, in the form of a firm set up without gov-
ernment support. The mill, called Osaka Cotton Spinning, grew to account for almost half of
total domestic output in 1886 and was also highly profitable. In that year, the government (which
also faced the need for budget austerity) wisely decided to get out of the way. Just a couple of
years later, the industry was firmly established on a robust growth path (see, e.g., Braguinsky et
al. 2015; Braguinsky and Hounshell 2016, for details). Locations of firms were highly dependent
2The others were silk mills started operating 1872 (Tomioka Silk Mill).
80
on local elites. Most firms were founded by local elites and local investors who became aware of
the new opportunities from modern industries.
Figure 3.1 summarizes the competition and growth of the cotton spinning industry from
1883 to 1914. Figure 3.1-(a) shows the number of firms by year, indicating there were many new
entries since the inception of the industry until around 1900, up to approximately 80 firms. As the
market competition became more severe, the number of firms decreased after 1900 due to both
exits and mergers. Figure 3.1-(b) plots the log of total output for the entire industry by year. The
output dramatically increased 20 times in the first 15 years. The growth of the output after 1900
was much slower, increasing 2.8 times in 14 years. Yet, since the dominant sector in Japan at this
period was agriculture, the cotton spinning industry was small relative to Japanese GNP when the
industry emerged. According to Long-Term Economic Statistics (LTES; Ohkawa, Shinohara and
Umemura 1967), the long-term estimation of economic activities in Japan from 1885 collected by
researchers in Hitotsubashi University, the cotton spinning industry comprised 0.72% of Japan?s
GNP in 1885. By 1899, following the period of rapid growth, the share of the industry in GNP
had increased to 4.06%. As part of the broader textile industry that overall comprised 25% of
GNP in 1899, the cotton spinning industry became one of the leading industries in Japan at that
time. Figure C.1 shows the historical trend of the fractions of the cotton spinning industry and
the textile industry in Japan?s GNP from 1885 to 1900.3
3The textile industry includes cotton spinning industry, as well as silk industry, small-scale weavers and other
industries producing final products related to textiles.
81
(a) Number of Firms by Year (b) Output by Year
Figure 3.1: Competition and Growth of the Cotton Spinning Industry
3.3 Data
The main data source for our analyses is Geppo Bulletins, which are the firm-level monthly
reports on inputs and outputs for all cotton spinning firms. This data is documented by published
by the All-Japan Cotton Spinners Association (hereafter, Boren using its Japanese acronym;
membership included basically all firms operating in the industry). As the cotton spinning in-
dustry was the very first and only modern nationwide industry in Japan in late 1800s, and there
was a strong industrial association set up for firms to help each other.4 The data record the number
of spindles in operation, number of days and average number of hours per day the firm operated,
output of the finished product (cotton yarn) in physical units, the average count (measure of fine-
ness) of produced yarn as well as a more detailed breakdown by different yarn varieties, the
number of factory operatives (separately for male and female workers), and average daily wages
separately for male and female workers. The data start from the very beginning of the industry
4The association was called the Cotton Spinners Association (Bouseki Rengokai), which was initiated by one
of the executive (from Aichi Boseki) in 1882. They renamed and reorganized the association in 1888 (to All Japan
Cotton Spinners Association) after experiencing new entries.
82
in 1883 and have been digitalized until the outbreak of the WWI in 1914. We focus our analyses
before 1900, when the industry experiences many new entrances. After 1900, several firms start
owning multiple plans due to new constructions and mergers, but the Geppo does not contain
information by plant-level. 5
3.3.1 Geographical distance
Figure 3.2: Example of a company report
Identifying geographical locations of each firm is the other essential component to inves-
tigate competition in local labor markets. In addition to Kinugawa (1937), semi-annual business
5We also have the plant-level data, which is annual statistics collected and published by government agencies.
The content of the data is the same as in firm-level data but split at the plant (establishment) level so that we can
observe within-firm differences, including but not limited to differences across geographic locations and trace all
ownership changes of which there were many during our sample (Braguinsky et al. 2015). These plant-level data are
available systematically only starting from 1899 but before that almost all firms, with just a few exceptions, were
single-plant firms, so that firm- and plant-level data coincide. We plan to expand the data to this periods by analyzing
importance of managerial practices in our future works.
83
reports for each company (called Kokajo in Japanese) stored in the Osaka University library are
used to identify locations. Figure 3.2 is a typical example of the first page of company reports.
It shows the location of the firm at the village-level.6 We use Google search to identify locations
of those villages: there are multiple webpages which store historical records of old addresses.
Because the geographical addresses in company reports are at village-level, we could not iden-
tify exact latitudes and longitudes for those firms except for a few of them. Yet, those villages
were very small (most villages are around 1km2) so that maximum of measurement errors should
be around 1km. Indeed, we will show in our analyses that results are very similar for different
definitions of peer firms. For those firms that we could not identify exact locations, we determine
latitudes and longitudes as most likely places in each village. As a ?most likely location?, we
look for places that are close to a river and currently occupied by a single agent such as schools
and industrial plants.
We use three different definitions for ?peers? or ?local labor market.? First is a previous
feudal domain (Kuni), which was the subset of current prefecture.7 Second, using guessed lati-
tudes and longitudes, we calculate distance from one firm to another. We define firms in the same
local labor market as those who are in a circle with 20km radius and 50 km radius from a focal
firm.
In Figure 3.3, we show location of firms by years. Borders in the maps represent the
previous feudal domains that were used for our analyses. Firms? locations were spread across
the country, and the number of firms increased over time. There were more firms concentrated in
some local area, such as Osaka, Okayama and Aichi prefectures, while other areas had relatively
6In this example, the picture shows the location of Amagasaki Boseki as Amagasaki-town.
7Typically, current prefectures are composed of one or two feudal domains.
84
smaller number of firms. Some of the government sponsored firms were located in the area
experiencing new entries in the end, while some others were in remote areas, and this is one of
the reasons of geographical heterogeneity.
(a) January, 1888 (b) December, 1890
(c) December, 1895 (d) December, 1900
Figure 3.3: Location of firms by years 8
8We used the map of old feudal domain digitized by Berman (2017).
85
3.4 Labor market institutions and descriptive analyses
3.4.1 The Beginning Stage
The Japanese labor market around 1900 was dominated by males. There existed a social
norm that kept females at home, and there were scarce working opportunities for females.9 The
majority of the population worked in farming, and there was an excess supply of workers in rural
area.
At its inception in the early-mid 1880s, the industry was comprised mostly of small-scale
government-owned and government-promoted mills, scattered across the country and employ-
ing their factory workers, both male and female, from local social networks, often from samurai
(former nobility) families, and treated like family.10 While such practices can help mitigate in-
formation asymmetry and the moral hazard problem, they are also known to be inefficient (e.g.,
Montgomery 1991; Heath 2018). Indeed, early Japanese cotton spinning mills were struggling
to make ends meet with no prospects of growing or putting a dent in the market dominated by
imports.11
One early exception was Osaka Spinning Company, a privately financed startup founded in
1882 whose first mill in Osaka, the commercial heart of the country, started operating in 1883.
The firm was the brain child of Eiichi Shibusawa, the father of Japanese capitalism. Right from
the beginning, it had spindle capacity more than five times larger than other mills, and it also
9One of the main exceptions was small scale manufactures, such as weavers.
10 The predominance of ex-samurai was not accidental, since governmental policy supported employment of
impoverished samurai first in some places samurai background became a qualification for employment. In the
Nagoya Cotton-Spinning Mill all the workers were samurai, and in the Okayama Cotton-Spinning Mill, supported
by the former feudal lord of the area, samurai wives and daughters, wearing kimonos with their family crests upon
them, went in and out of the mill gate. (Tsurumi 1990, pp. 36-37, citing Kinugawa 1937, Vol. 3, p. 307).
11For a more comprehensive analysis of the struggles of government-promoted mills see Braguinsky (2015).
86
introduced a whole bunch of technological and managerial innovations that made it the example
followed by the industry in years to come (Saxonhouse 1974; Braguinsky and Hounshell 2016).
By 1886, the mill had already tripled its capacity and alone accounted for almost half of total
industry output. Osaka Spinnings launch was financed by investors who subscribed to its shares.
The shares were actively traded. Already in the first half year, there were 24 separate transactions
in which 165 shares, or 5.9 percent of the total changed hands (Osaka Spinning, 1883-1914,
No. 2, 1884, p. 8-10). This imposed a tight financial discipline, and even more so after it was
listed on the Osaka Stock Exchange in March 1886 (Toyo Boseki, 1986, p. 55). In order to
prove to investors that it was worth investing, the company engaged in a quest for ever-larger
profits, hiring its workforce from impoverished first-generation urban dwellers who flocked to
Osaka hungry, empty-handed, and exhausted from struggles for livelihood and land that were rife
all over Japan. Predictably, the wages offered were lower than those paid at most other cotton
companies. (Tsurumi 1990, pp. 41)
While the above sounds like a perfect illustration of the Harris-Todaro type models of
reserve armies, the situation was actually more nuanced. Even though Osaka Spinning indeed
tapped into an entirely different labor pool than other mills, the inference that its wages were
lower than those paid at most other companies turns out to be misleading.12 In Figure C.2 in the
appendix we use all the available data for 1883-84 which show that Osaka Spinning indeed paid
less than other mills to its male workers but significantly more to its female workers. After 1884,
the wage data on other mills become available only starting with the second half of 1888, and, as
shown in the same Figure C.2, in 1888-1889 Osaka Spinning already paid significantly more to
12Tsurumi cites an authoritative study by the Japanese historian, Naosuke Takamura. The statistics on which his
assertion is based, however, come from just one month of observations (Takamura 1971, p. 79).
87
both its male and female workers, with female wages higher than in any other mill operating in the
industry in those years, including new privately financed startups. In fact, Osaka Spinning, which
was the only consistently profitable firm during this period, was also the only firm that increased
wages to its male and female workforce: its average male wages increased by 13 percent and
its average female wages increased by 9 percent between 1883-84 and 1888-89 (both increases
statistically significant at the 5 percent level), while among 13 government-promoted firms, male
and female wages declined on average by 22 percent and 14 percent, respectively, over the same
period.13 Crucially, the upward trend in wages paid by the Osaka Spinning mill can be traced to
improved workers skills.14
Prior to starting operations, Osaka spinning management decided to set the starting daily
wage for male workers at 12 sen and for female workers at seven sen.15 The male wage was
about half of the prevailing daily wages of masons and carpenters and can be considered to be
at or just above the subsistence level; as female wages were below that level, those had to be
supplementary family income.
Starting wages were set low in part because the cotton-spinning hands were all complete
novices. The plan was to raise their wages as they acquired skill with the training they received.
(Kinugawa 1937 Vol. 2, p. 420) To implement this plan, Osaka Spinning developed a wage-grade
classification system, with nine different grades for male workers, ranging from 10 to 24 sen per
13Part of the decline in wages paid by other mills could be due to the aftermath of the Matsukata deflation which hit
the agricultural sector especially hard (see, e.g., Ericson 2014). Female daily wages in agriculture declined by more
than 40 percent in the Osaka prefecture during this period with similar trends across the country. Osaka Spinning
thus bucked not just the trend in cotton spinning but in agricultural wages as well.
14Even though Clark (1987) maintains that most textile jobs did not require great skill, expensive imported spin-
ning machinery was complex to operate and required mastering of brand-new technologies (Tsurumi 1990), making
experienced, skilled workers a crucial factor of improved productivity, as we demonstrate below.
15According to Kinugawa (1937) (Vol. 2, p. 420). One sen = 1/100 yen. Twelve sen reportedly bought about
three kilograms of rice at the time.
88
day and six different grades for female workers, ranging from 7 to 17 sen per day (three sen less
for women in the same grade as men throughout). (Several more apprenticeship wage sub-grades
were established below those schedules for both male and female workers.) According to Osaka
Spinning first company report (Osaka Spinning, 1883-1914, No. 1, 1883, p. 18), there were 14
female workers (8.8 percent of the total) assigned to the higher grades that earned from 13-17 sen
a day. Just a year later, there were already 22 female workers, or 13.4 percent of the total in those
grades (Osaka Spinning, 1883-1914, No. 3, 1884, p. 8). Since then, average wages and the peak
of the wage distribution increased over years.
Why would firms such as Osaka Spinning that could hire from the reserve army at basically
the subsistence wage nevertheless pay more to workers as they acquired more skill? Mills spent
time and money on training workers, in particular by sending them as understudies to already
operating mills, incurring considerable costs (see, e.g., Saxonhouse 1974, p. 162). The mills
were rewarded for these expenses by higher productivity. But if the opportunity cost for workers,
regardless of their training, was still the subsistence wage, why would workers need to be paid
more? An important part of the answer, examined in detail in the next subsection, is that as
the industry experienced large-scale new entry, new mills raided skilled workers of incumbent
firms. However, when Osaka and Mie Spinning came up with their wage-grade schedules back in
1882, they did not have to worry about their workers being raided (worker raiding only became a
problem after 1887-88Kinugawa (1937), Vol. 5, Ch. 18; Tanaka (1957), pp. 49-50).
Workers needed to be paid more for acquiring skill even in the absence of competition,
because skill acquisition also required upfront effort on the part of the worker her(him)self, and
(s)he had to be given incentives to exercise that effort. One potential explanation is as follows:
workers are initially unskilled with low productivity, and both firms and workers can make costly
89
investment in order for workers to acquire skill in the next period which raises workers produc-
tivity. If a workers own effort investment is an essential part of the process, the firm needs to
offer a wage equal to the opportunity cost (subsistence wage), plus the cost of workers effort
even absent of intra-industry competition. Once intra-industry competition for the scarce pool of
skilled emerges and intensifies, skilled workers wages increase further.
From the workers? perspective, female workers supplied from agricultural areas worked
under 3-5 years contracts with cotton spinning firms, stayed at the dormitories provided by their
firms, and went back to their home regions at the end of their the contracts. Historical anecdotes
in Okamoto (1993) suggest that workers in general could save some amount of cash to bring back
their savings. Yet, according to the cross sectional survey in Shokko-jijo, significant fraction of
workers quit jobs within six months.16 A lot of those who worked more than six months also did
not work until expiration of contracts.17 The survey points out multiple reasons why workers quit
their jobs before the expiration of their contracts. Some workers quit jobs because they did not
get used to workplaces (since it was their first experiences working, or they found it difficult to
learn skills), and some others quit jobs for personal reasons (such as being home-sick, or family
issues). Firms paid transportation costs when they hired workers from remote areas. If workers
quit within a certain period (which, depending on firms, could be from a few months to a few
years), workers were asked to pay back these transportation cost as a penalty. Therefore, to be
incentive compatible to train workers, firms incorporate such costs of quitting when designing
16The fraction of workers who quit their workplace highly depends on firms. For instance, the fraction of workers
who quit the workplace within six months varies from 3% (Tokyo Boseki) to 80% (Noda Boseki) with the median of
40%.
17Again, the fraction of workers who worked until expiration of contract varied by firms. The fraction varies from
6% (Fukushima Boseki) to 90% (Totoumi Boseki) with the median of 29%. We should be careful when interpreting
this data, because some firms were established just before the survey time (1896), and thus there were mechanically
not so many workers who worked long enough.
90
wage schedule.
As mentioned, one thing that differentiated Osaka Spinning from other mills was its larger
mill size. The literature has consistently found that larger firms pay higher wages to observa-
tionally similar workers, both in developed (Brown and Medoff 1989; Abowd, Kramarz and
Margolis 1999) and developing nations contexts (Strobl, Thornton et al. 2004; So?derbom, Teal
and Wambugu 2005). This is also true in our data. However, since capacity expansion was
bulky, mills needed to hire a lot of new hands at the time of expanding capacity, and especially in
early years when skilled workers were scarce, these new hands were predominantly low-skilled,
temporarily reducing the average wage and increasing the skewness of the wage distribution. In
FigureC.3 we illustrate this on the example of Osaka Spinnings first capacity expansion. The
changes in the female wage distribution clearly reflect the impact of mass hiring of new hands
that, once again, had to be trained before they would acquire the necessary skills (see Appendix
for details).
3.4.2 Large-scale entry and industry expansion
We now turn to examining changes in the labor market, especially for female workers,
as the industry experienced rapid expansion starting from the late 1880s. As the government
abandoned its support to the industry (and intervention into its matters), a host of new privately
funded startups emerged, and, as can be seen in Figure 3.1 above, the number of mills increased
steadily until it reached a peak of 74 mills in 1898.
New entry, coupled with the expansions of some existing mills, increased the demand for
mill hands. By itself, however, it does not have to lead to any increase in wages. Wages only
91
start increasing if labor supply becomes less than perfectly elastic, and this is not the case in the
presence of a large reserve labor army. Indeed, industry-average wages were largely flat until the
mid-1890s but then started increasing sharply, especially for female hands.
After mid-1890s, workers were poached from one firm to another, and became one of the
main sources of conflicts between firms. If a firm figured out that another firm poached their
workers, they complained to Boren, the industrial association, to arbitrate the conflict. Geppo
records many such conflict cases, and most of them were during the mid-1890s.18 Except one
case, all the poaching and arbitration cases were between firms in the same local districts. This
implies differential costs of poaching workers by locations, and indicates that the more firms
located in same local areas, the more chance that workers were poached from others.
Consistent with recent literature in labor market concentration (e.g. Rinz 2022), wage
level depends on concentration of firms. As shown in Figure 3.3, new firms were established
year-by-year. This generates variation in the number of firms by locations. Figure 3.4 average
wages by number of peer firms. In Figure 3.4-(a) we present the breakdown of wage levels of
unskilled workers in the industry by the level of concentration of firms from 1880 to 1915 using
monthly data form Geppo. ?Isolated? refers to the average wages of the firms in areas with
only one or two firms, ?high? refers to the average wages in Osaka and surrounding areas which
accommodated up to 30 firms, and ?middle? refers to the average wages in the areas between
isolated and highly concentrated areas. Especially for female workers, wages were systematically
lower in the isolated areas, and higher in the highly concentrated areas. After 1900, annual
18Between 1892 to 1898, 3,757 inquiries were sent to the centralized industrial association regarding workers.
Among those, there were 73 inquiries for holding mediations for the poaching issues. Note that those are only the
observable cases which were detected by each firm. It is hard to count actual number of cases because firms might
not report some cases. Therefore, it is natural to imagine that there could be much more worker poaching actually
happening.
92
plant-level data is additionally available. Because of mergers and acquisitions of firms as well
as expansions of the industry, multiple firms owned multiple plants in different locations after
1900. In Figure 3.4-(b), we use those firms and classify plants into those in relatively high-
concentration areas and relatively low-concentration areas within the same firms, and plot those
wages by gender. Again, there is a systematic difference in wages: wages were higher in the areas
with more peer firms for both genders. Note that even though there were differences in product
varieties, production technologies were very similar among firms across the country.
(a) Monthly data for all firms (b) Same firm with different locations
Figure 3.4: Wage variation by location
This picture can be explained by changes in the nature of intraindustry competition, where
firms in different locations, depending on the number of peers, confront different costs to ac-
quire skilled workers. Figure 3.4-(a) uses data from Geppo, containing the monthly wages for
the universe of firms. In addition to this data source, an industry survey conducted by Boren in
1896 collected distribution of wages on almost all mills, so we have a large cross-section of wage
distributions at that point (Dainihon Menshi Boseki Dogyo Rengokai 1898). Table 3.1 shows
features of wage distribution by concentration of firms. First three columns show the average of
minimum and maximum wages by different level of concentration, and last three columns com-
93
pares those and conducted t-test. The table shows that wage levels for unskilled workers were
similar for all firms regardless of the concentration. In contrast, the highest wages were different
by the level of concentration. This is a consistent result with the historical evidence discussed
above, that firms could hire workers with low costs, but had different wage setting powers by
concentration due to the different chances of their skilled workers being poached.
Table 3.1: The lowest and the highest wages in 1896 by concentration
Wages by concentration Difference
Low Mid. High Low vs Mid. Mid. vs High Low vs. HIgh
The lowest 6.52 7.07 8.20 0.55 1.13 1.67
female wages (0.42) (0.36) (0.59) (0.97) (1.72) (2.32)?
The lowest 9.76 9.62 9.15 -0.14 -0.47 -0.61
male wages (1.00) (0.84) (0.62) (-0.11) (-0.51) (-0.41)
The highest 22.29 28.28 31.10 5.99 2.82 8.81
female wages (1.39) (1.09) (2.39) (3.46)?? (1.19) (3.24)??
The highest 47.14 60 60.5 12.86 0.5 13.35
male wages (2.35) (0.84) (1.98) (4.18)??? (0.15) (3.67)???
N 21 29 20
Standard errors are in parentheses for the first three columns, and t-statistics are in the parentheses
for the last three columns. For t-statistics, ? p < 0.05, ?? p < 0.01, ??? p < 0.001
Finally, Table 3.2 reports the summary statistics for selected years to see the trend of the
industry. The first three columns show means, standard errors, minima and maxima for firms
with not more than 4 peers,19 used in the second part of the analysis. The next three columns are
firms with any number of peers (i.e. all available firms in my data source). Using the monthly
data, we calculated annual averages for the year of 1888 (the starting year of my analyses), 1890,
1895 and 1900 (the ending year of my analyses). ?N? represents the number of firms at the end
19We describe multiple definitions of ?peers? in the next subsection. Here, we present the results with the defini-
tion of peers as firms within a circle with 20km radius from a focal firm.
94
Table 3.2: Summary statistics
? 4 peers All firms
Year Variable Mean (s.d.) Min Max Mean (s.d.) Min Max
Total output 97.02 (82.39) 17.02 489.37 155.44 (234.25) 17.02 1151.02
1888 # of male workers 22.41 (14.99) 6 79 37.63 (56.13) 6 305.5
# of female workers 51.71 (38.60) 19 203 66.70 (71.24) 19 430
Male wage (Yen) 15.39 (2.61) 9.9 20.8 15.79 (2.84) 9.9 22
Female wage (Yen) 6.91 (1.38) 4.9 10 7.21 (1.61) 4.9 11.7
N 27 35
Total output 269.86 (283.16) 14.19 1317.96 361.65 (420.52) 14.19 2192
1890 # of male workers 50.67 (49.52) 9 321 75.82 (101.61) 9 572
# of female workers 136.79 (146.29) 15 759 187.34 (210.35) 15 1015
Male wage (Yen) 16.89 (2.51) 10.2 26.1 17.19 (2.58) 10.2 26.1
Female wage (Yen) 7.93 (1.38) 5 12.9 8.18 (1.64) 5 14.8
N 30 40
Total output 587.94 (599.53) 10 3075.81 730.27 (716.99) 10 3075.81
1895 # of male workers 85.32 (101.08) 9 581 109.18 (113.89) 8 581
# of female workers 287.79 (327.51) 14 1983 353.32 (333.22) 24 1983
Male wage (Yen) 17.84 (3.23) 10.8 29 17.96 (3.12) 10.8 29
Female wage (Yen) 9.473 (2.08) 5 18.5 9.97 (1.97) 5 18.5
N 48 71
Total output 890.58 (1592.13) 17.35 11833 905.91 (1359.27) 17.35 11833
1900 # of male workers 121.59 (182.14) 6 1248 129.12 (165.33) 6 1248
# of female workers 449.72 (687.20) 23 4787 449.19 (575.58) 23 4787
Male wage (Yen) 28.89 (5.47) 17 51 28.63 (4.69) 17 51
Female wage (Yen) 16.90 (2.72) 9 23 17.91 (2.86) 9 26
N 52 93
of each year (December). The table shows consistent growth of the industry, both in terms of the
number of firms and the scale of each firm. The number of firms almost doubled in the region
with less than or equal to 4 firms, and more than doubled by counting all firms. The firm size
also expanded as time went on: rapid increase in the total output is observed, so as the number
of male and female workers. Wages for both male and female workers increased as the industry
grew, and our analyses focus on the relationships here: the industry evolution and the evolution
of the labor market. It is also notable that the average firm size in the area with less than or equal
to 4 peers was smaller than the average of all firms. This is because Osaka region is excluded
from ?? 4 peers? specification: there were a lot of firms located in Osaka (at most 25 in 1900)
95
and thus the region can be assumed to be in a competitive environment. Some of the leading
firms (e.g. Osaka boseki, Amagasaki boseki, etc) were in this region, and those firms push up the
mean for all firm averages.
3.5 Regression analyses
Motivated by descriptive evidence regarding the formation of local labor market through
the evolution of the industry, we now examine how competition affects wages. More precisely,
we focus on increase in peer?s labor demand through increased capacity. In the standard theory of
production of perfect competition, firms are small enough that one firm cannot affect equilibrium
wages: with a completely elastic labor supply curve, changes in labor demand in one peer firm
should not affect wages of the focal firm in an equilibrium. However, if workers accumulate
general skills within industry, where the number of those who have the skills are sufficiently
small and firms increase demand for those workers (or alternatively, firms are large enough in a
local economy and each firm faces an imperfectly elastic labor supply curve), a change in demand
on labor from one peer firm should affect wage levels in surrounding firms. Our basic empirical
strategy here is to use lagged variables of peer?s production variables at time t ? k, instead of
using current variables at time t to avoid potential problems discussed later.
In this analysis, we fully utilize two unique features of the data. First, the data contains
firms? production characteristics at monthly level. Our analysis uses 13 years (156 months) of
time periods, and that enables us to directly estimate firm fixed effects. Second, instead of the
random sampling of firms in most studies on firms, the universe of the cotton spinning firms that
existed in the study period are observed in the data. This allows us to capture peers? capacities
96
for all firms. Combining these two features with using lagged variables, we show that we are
able to estimate the effect of the increase in neighboring firms? capacity expansion on wages in a
fairly simple way.
3.5.1 Model and Identification
We estimate the following model:
wageit = Xi,t?k?1 +Xj(?i),t?k?2 +Xi,t?3 +ResWagei,t? + ?t + ?i + ?it. (3.1)
The dependent variable in this estimation is average monthly wages for manual workers, sepa-
rately by gender. Consider firms set wages considering three sets of information: Xi,t?k, a vector
of lagged variables which determines the production of firm i at time t?k, such as the number of
male and female workers, the number of total capital and the quantity of total output; Xj(?i),t?k,
a vector of the sum of lagged variables for peer firms; and Xi,t, the current production variable
for firm i. Xi,t?kand Xj(?i),t?k can be characterized as variables determining firm specific labor
supply function. As a main assumption, we conside that firms could not observe others? actual
labor demands at time t. Instead, they utilize information from previous period (t? k) to make a
best guess for the labor supply curve (See more discussion with a theoretical model in Appendix
C.1).20 Xi,t, the number of workers and total capital at time t, is also included in the model to
capture firms? decisions at time t. As in the previous section, we define peer firms in multiple
20This is just an assumption, but we believe this is plausible, if we consider the background of how the data was
constructed. Due to the difficulties on communication in the historical context, acquiring information on other firms
were likely to be costly. As explained in section 3.3, firms gave reports to centralized association every month, and
received the summarized report a few months later. The reports provided information regarding other firms, and this
should affect firms? behavior to formulate their ?best response?.
97
ways, such as those in the same feudal domain and those located within a circle with 20km and
50km radius from the sample firm. To mitigate potential bias as described below, we also in-
clude reservation wages in the agricultural sector as ResWage 21i,t. ?t and ?i are time and firm
fixed effects. We separately put year and month fixed effects. ?2 is our parameters of interest,
since changes in Xj(?i),t?k capture increased competition in the local labor market. For the main
specification, we directly estimate fixed effects (which is so-called LSDV: Least square dummy
variables estimator) to avoid some identification concerns described below.
There are multiple concerns on consistently identify ?2. The first is the simultaneous deci-
sion of the number of workers (in Xi,t) and wages. If we assume that firms only decide wages in
period t, and not the number of workers, then this is not a problem. In the standard literature on
imperfect labor markets, wages and the number of workers are simultaneously determined, and
thus estimating equation (3.1) by LSDV causes biases on ?3. ?2 is also likely to be biased as long
asXj(?i),t?k andXi,t?k are correlated (See Appendix C.1: we show a simple theoretical model on
why there is going to be a bias due to simultaneity). Though it may not seem intuitive to assume
that firms only decide wages but not the number of workers, it may actually be true if we think
about the nature of production in the industry. The production in the cotton spinning industry
requires a certain number of workers to work with a given number of machines. That is, even
though there is some substitutability between capital and labor, the number of spindles requires
a certain number of workers for smooth operation.22 In other words, if firms could not secure
21Data is annual level, digitized from various prefectural statistic books.
22More precisely, each frame must be serviced by workers who feed input, monitor the process, and pick up
output etc. The relationship is not strict enough to preclude estimating a production function with some degree of
substitutability between machines and labor, but it is strong enough to be a constraint, especially when considering
reducing the number of workers in response to higher wages.
98
workers per unit capital, they would need to pay the opportunity cost of idling machines. If this
opportunity cost was high enough and firms earned profits by enjoying monopsony power over
labor market, the problem that firms faced was to choose wages, but not the number of workers.
If the opportunity cost of idling machines was not high enough, we need to use instru-
mental variables to address the bias resulting from simultaneity. The ideal sets of instruments is
correlated with the number of workers and affect wages only through changes in the number of
workers. That is, instruments should only shift the labor supply curve, but not the labor demand
curve. We use institutional features of the labor market explained in Section 3.4.2 to construct
a set of instruments. Historical evidence suggests that workers were initially hired for subsis-
tence wages and accumulated skills through their work. Due to the shortage of workers, firms
hired workers from multiple regions, frequently far a way from the locations of plants. Daini-
hon Menshi Boseki Dogyo Rengokai (1898) documents the breakdown of regions from which
firms hired workers in 1896. Since establishing necessary ties in the regions (including setting
up recruitment officers, contracting with the recruiters, complying with local regulations) was a
time-consuming process, it makes sense to assume that there was a large degree of ?stickiness? in
regions from which firms hired workers. Assuming the regions listed for each firm for 1896 ap-
ply also to other years, we use ?average characteristics of firms that hired workers from the same
regions but were located in different (i.e. non-neighboring) regions? (hereafter ?characteristics
of IV firms?) as a set of instruments. Any sort of local labor supply shock in a certain district
should affect the number of workers for firms hiring workers from such regions, but should not
affect wages for firms hiring workers from that area because workers supplied from rural area
were hired for subsistence wages. We include firms in different locations, because if those firms
were located in the same district their wages are influencing each other because of the poaching
99
behaviors.23 That is to say, if firms A and B hire workers from the same region Z, but are located
in different regions (districts) X and Y, respectively, we can use the impact of the supply shock in
region Z on worker hiring by firms B as an instrument for the decision on the number of workers
to be hired by firm A. In addition, as discussed in Arellano and Bover (1995) and other papers
in dynamic panel mode, if Xi,t is endogenous, the fixed effect estimator becomes mechanically
biased. In Appendix C.2, we show that it does not cause a problem where T is relatively large in
our environment.
The second issue on the identification is one of the most common problems econometri-
cians need to address when they want to estimate peer effects in general: the reflection problem
(Manski 1993). Roughly speaking, if neighboring firms are influencing each other such that
Xi,t and wagei,t are influencing Xj,t and wagej,t, then ?1 and ?2 are not consistently estimated
because each firm simultaneously affects each other (e.g. firm i?s wage affects firm j?s wage
through ?1, and then firm j?s wage affects firm i?s wage, and so on). Most common approach
in dealing with this is to use spatial lag models such as Kapoor, Kelejian and Prucha (2007).
We overcome this issue by utilizing institutional feature that firms need some time to acquire
information: instead of considering simultaneous influences, we assume firms need some lags to
respond to other firms? capacity expansion, because of the friction to timely convey We overcome
23There is a potential problem with these instruments. If subsistence wages also changed due to the labor supply
shock in the source region, the instruments do not satisfy the exclusion restriction. However, historical records of
the firms? internal rules suggest that firms set certain wage policies by workers? rank and tenure. Since firms hired
workers from multiple regions, labor supply shock in a single district far away from the plant should not likely be
affecting wage policies. That is, we assume that the lowest wages in the internal rule should be similar to subsistence
wages in the firms? location, which might not be as same as the subsistence wages in the source region. (should be
verified by data, which is under construction). This means even if subsistence wages in source regions change under
labor supply shock, it does not matter for the wage setting policy. This holds when subsistence wage is lower in
source region than firm?s location, which is likely to be true because workers actually decide to take the job.
100
this issue by utilizing institutional feature that firms need some time to acquire information re-
garding the labor market. There is high correlation between the number of workers, total capital
and total output (i.e. it seems that production technologies are highly complementing each other),
and firms needed multiple months to expand their factories because machines had to be imported
from the United Kingdom. Hence, firms needed multiple months to prepare for capacity expan-
sions, and could not invoke a ?surprise? increase in labor demand. As a consequence, peer firms
should not have to consider best responses to those surprises. Rather, it is better for them to adjust
wage and other production technologies after observing capacity changes of peer firms.
Third, while the reverse causality is not a problem due to the lagged feature, estimated
parameters are going to be biased if unobservable characteristics influence both Xj(?i),t?k and
wagei,t. While labor market is formed in a local level, output market is fairly integrated (Kinu-
gawa 1937). Therefore, any shocks on output market should be absorbed by time fixed effects.
Firms? choices on expanding their capacity is likely to be made well before the actual expansions.
Therefore, unobserved, temporal and local economic shocks should only affect either Xj(?i),t?k
or wagei,t, not the both of them due to the differential timings of t ? k and t. The only unob-
servable that we cannot incorporate is the persistent local economic shocks. In such cases, the
error term andXj(?i),t?k are likely to be positively correlated, because a positive economic shock
would boost both capacity expansion of local firms and increase reservation wages for workers.
If this is the case, ?2 is overestimated. To incorporate this issue as much as possible, we include
wages for the agricultural sector (ResWagei,t) as reservation wages in the regression.
We cluster the error term by feudal domain level, as local economic shocks are likely to
affect firms in similar areas. Since the number of clusters is small (27 feudal domains), we
incorporate Cameron, Gelbach and Miller (2008)?s method to obtain t-statistics, utilizing wild-t
101
bootstrapping.
3.5.2 Results
Table 3.3 shows the results estimating equation (3.1) by LSDV. In this table, we define two
months lag (k = 2) for the regressors, and peer firms to be those in a circle with 20km radius.
All the numbers presented in the table are representing peer firms? characteristics (Xj(?i),t?k).
To make interpretations of the results easier, we divide Xi,t?k by 1,000. All specifications also
include own lag variables (Xi,t?k), firm, year and month fixed effects, and reservation wages in
agricultural sector in firms? locations. Dependent variables are female wages in columns (1) to
(5), male wages in columns (6) to (10), and male to female wage ratio in column (11). Since our
main focus is to analyze the imperfect labor market, we restrict the sample to firms which had
less than or equal to 4 peer firms in their local region. As one may notice, we use ?20km circle?
as a definition of local labor market in our analyses, while ?previous feudal domains? are used for
clustering. This is because firms categorized as peers differ by firm for ?20km circle? definition,
and we cannot achieve well-defined clusters.
Because we include both time and firm fixed effects, identification variations we are using
in these estimations are changes in peer firms? capacity, which are essentially changes in labor
demand in firms? locations. Under the assumptions on peer effects that we discussed in the pre-
vious subsection, one can interpret these estimated parameters as a peer effect of firms? capacity
expansion on the wage level.
Due to the high correlations between production variables, columns (1) to (4) and (6) to
(10) only include one out of four production variables. All of those results indicate that peers?
102
capacity expansions cause increase in wages. In column (5) and (10), we include all of the pro-
duction variables: it seems that high correlations between production variables leave only two
variable (the number of female workers and firm capacity) which are highly associated with
wages for sample firms. With those full specifications in column (5) and (10), we can interpret
the estimated coefficients as indicating that an increase in peers? capital by 1 standard deviation
from the mean increases the wages for male workers by 7.1% and for female workers by 9.5%.
Under infinite supply of labor, waged should not change under equilibrium with peer?s capacity
expansion. Instead, the estimation results suggest that expansion of peer cotton spinning firms
influences wages paid by focal firms.
Table 3.4 shows the results estimating equation (3.1) by instrumenting the number of work-
ers by ?characteristics of IV firms?. The tricky thing here is that, since we are instrumenting the
number of own workers, which is not the parameter of interest, and it is not clear the relation-
ship between the number of workers and peer?s capacity expansion both in terms of the sign and
magnitude, we could not analyze the direction of bias on the parameter of interest ex-ante. Yet,
the results presented in Table 3.4 are fairly similar to what we obtained with simple LSDV esti-
mation. As suggested in Table 3.3, we concentrate on the results with peers? ?firm capacity? as a
parameter of interest. we also present the results with different time lags (l = 2, 6 in columns (1)
to (4) ) and different definition of peers (20km and 50km from focal firms as peers), but results
are very similar to the results just employing LSDV (see Table C.2to C.6 in the Appendix). The
first stage results for this 2SLS specification are presented in Table C.1. Though the average
total output of IV firms are in general correlated with the number of workers at the focal firms
(especially when l = 6; t-statistics are close to 5% level). However, F-statistics testing joint sig-
103
nificance of the instrument is not that high, especially in the specification (1), (5) and (6). This is
because of high R2 resulting from two-way fixed effect specification, which absorb large fraction
of the variation.24
3.5.3 Discussion
Results presented above are robust to using different lag variables and different definition
of peers, as well as different criteria for the sample. To check the robustness of the results, we
tried different lagged variables (k = 6, 12), which are shown in the Appendix. We also show
different definition of peer firms. Again, results are very similar (see Table C.2 to C.6 in the
appendix section). Table C.8 shows the result with different selections of samples by different
threshold of the maximum number of firms in the peer regions. As explained in the previous
section, we constrain the number of peers to be less than or equal to 4 firms in the main analysis.
In Table C.8. we estimate same specifications for the model (5) and (10) in Table 3.3 with the
different number of peers, and show that the results are robust across specifications. As predicted
from the standard theory of labor economics, increasing the number of peers decrease the point
estimates of the effect of peer firms, both male and female wages.
24Therefore, those 2SLS results are just suggestive at this point showing that the results do not change that much
with simple LSDV. We will consider other potential instruments as our future work.
104
105
Table 3.3: Peer effects from other firms in 20 km: two months lag regressors
Female Wage Male Wage Wage gap
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11)
# of Male Workers 4.279??? 0.139 9.211??? -3.759
(0.00) (0.83) (0.00) (0.198)
# of Female Workers 1.173??? 1.246??? 2.554??? 2.282???
(0.00) (0.006) (0.00) (0.00)
Firm?s capacity 0.0376??? 0.0199?? 0.0858??? 0.0615??? 0.000286
(by # of spindles) (0.00) (0.014) (0.00) (0.00) (0.928)
Total output 0.411??? -0.276 0.946??? -0.168
(0.00) (0.332) (0.00) (0.382)
Observations 2368 2368 2368 2368 2368 2367 2367 2367 2367 2367 2366
R2 0.911 0.912 0.912 0.910 0.913 0.884 0.885 0.886 0.882 0.887 0.685
P-values are in parentheses, which are calculated by the wild-t bootstrap method suggested by Cameron et al. (2008). Each observation is clustered by previous
feudal domain level (27 regions). All regressions are estimated by LSDV(Least Square Dummy Variables), including firm fixed effects, year fixed effects, month
fixed effects, own lag variables (# of male and female workers, # of total capacity and total output), current number of workers and capitals, and reservation wages
for agriculture and weavers. ? p < 0.05, ?? p < 0.01, ??? p < 0.001
106
Table 3.4: Instrumental variable estimations on the effect of peers? capacity expansion
Firms in 20km as ?neighbors? Firms in 50km as ?neighbors?
l = 2 l = 6 l = 2 l = 6
(1) (2) (3) (4) (5) (6) (7) (8)
Female Male Female Male Female Male Female Male
Firm?s capacity (by # of spindles) 0.0352??? 0.0823??? 0.0333??? 0.0757??? 0.0349??? 0.0796??? 0.0359??? 0.0884???
of neighboring firms, lagged (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000)
Number of workers, 0328? 0.0482 0.0112 0.0051 0.038 0.141 -.0144 0.0296
current (0.0426) (0.483) (0.078) (0.752) (0.195) (0.217) (0.625) (0.423)
Observations 2368 2367 2268 2267 2152 2150 2010 2008
P-values are in parentheses, which are calculated by the wild-t bootstrap method suggested by Cameron et al. (2008). Each observation is
clustered by previous feudal domain level (25 clusters for the regression (1)-(4), and 20 clusters in (5)-(8)). All regressions are estimated by 2SLS
instrumenting the number of workers at current period by average firm characteristics of those located in different regions but hiring workers from
same regions. All regressions include firm fixed effects, year fixed effects, month fixed effects, own lag variables(# of male and female workers, #
of total capacity and total output), number of capitals and the number of workers at the current period, and reservation wages for agriculture and
weavers. ? p < 0.05, ?? p < 0.01, ??? p < 0.001
Also, one may concern there is better specification than the equation (3.1) we estimate.
Some variables may be not relevant for wage setting policies, and even worse, these variables
may be bad controls. In Table C.7, we re-estimate the model with alternative specifications.
As a benchmark, the same results as column (4) and (9) of Table 3.3 are presented in columns
(1) and (2). Specifications (3) and (4) exclude own current variables (# of workers and # of
capitals). Specifications (5) and (6) excludes own lagged variables(# of male and female workers,
# of total capacity and total output). Results in column (3) to (6) are very similar to our main
findings: peer firms capacity expansion affects own wages for both male and female. For all the
analyses above, we assume that the remaining error term after controlling time and firm fixed
effect is idiosyncratic. One may concern that there is unobservable shock which affects both
Xi,t?k and wi,t, even after controlling for agricultural wages. To address this concern, we also
add wi,t?k in specifications (7) and (8) as a control variable. Now, our assumption is that the
idiosyncratic shock in t ? k should affect wi,t (where there is k periods to adjust). As own
wage 2 month ago is highly correlated with current wage level, point estimates of the peers?
capacity expansion drop by approximately half. Yet, the effect of peers? capacity expansion is
still statistically significant.25
There are some missing observations due to either temporary shutdowns or missing reports.
This may cause measurement errors on peers? characteristics, which are our key variables. Some
of the temporary shutdowns can be identified from the original data source, while others may be
not. In such a case, peer firms? characteristics are systematically smaller than actual values, and
generate a bias to estimate ?2. Though it is infeasible to fully address the problem of measurement
25As pointed out in Galiani and Gonzalez-Rozada (2005), bias for LSDV estimate of the covariants should be
small. To reduce the bias further, we also conducted LSDVC (Corrected LSDV) suggested by Kiviet (1995) using
IV estimator of Anderson and Hsiao (1982). The results are very similar to what we present in specification (7) and
(8). The results are up on request.
107
errors by finding instruments, we present additional results in Table C.9 and Table C.10 with
imputed missing variables in different ways. In Table C.9, we impute missing observations if we
could not find the reason for missing data (i.e. we could not identify if it was due to a temporary
shutdown or a missing report). The imputation takes following steps. First, if we can find some
observations before and after the missing observation, we take the average of the most recent
values from the same firm before and after the missing observation. Second, if we cannot find
any observations either before (or after) the missing observation, we impute the most recent value
of the same firm after (before) the missing observation. By imputing these missing observations
and missing values, the number of observations increases by approximately 1,000. The results
remain very similar after this procedure. In Table C.10, we also impute missing observations for
those identified as temporary shutdowns. Since firm capacity (i.e. the number of spindles) is a
stock variable, rather than a flow, it may be reasonable to assume that firms still hold production
capacity as they restart production. The imputation procedure is exactly as same as above. The
number of observations in Table C.10 now becomes 5,132, but the results remain similar. For both
additional estimations, point estimates for key variables become approximately 20-25% smaller
than the main results in Table 3.3 for both genders. One of the reasons is the measurement errors
due to imputation process (though the original analyses also involve measurement errors, and it
is hard to quantify which errors are larger). In addition, it may be also because firms anticipated
future shutdowns, and had less incentives to increase wages in these cases.
108
3.6 Conclusion and further agenda
In this chapter, we utilize an unusually rich historical data set on firms, establishments,
workers and wages in the Japanese cotton spinning industry in late 19th to explicitly link the
evolution of an industry to labor market outcomes.
In the early- to mid-1880s, Japan did not have any modern industry hiring hundreds of
thousands of workers. The cotton spinning industry was one of the first modern industries in
Japan, which had enormous impacts on economic growth and formation of the labor market.
We firstly summarize historical documents and descriptively show how wages were set in the
very early stage of development. Initially, the industry was comprised mostly of small-scale
government-sponsored mills, employed typical labor practices in agricultural economyfactory
workers, both male and female, were recruited from local social networks and treated like family.
In line with that, early Japanese cotton spinning mills were struggling to make ends meet with no
prospects of growing or putting a dent in the market dominated by imports. We then document
how the industry departed from such a situation, and became one of the largest industries in
Japan. Though there were independent private entrants that amplify the scale of the industry,
initial labor practices in the late 1880s-early 1890s still resembled the Lewis (1954) and Harris
and Todaro (1970)?s dual economy model. The new profit-motivated firms abandoned the master-
vassal notion of smaller mills and turned to recruiting cheap labor from impoverished urban and
suburban slums and later from more remote rural areas.
We then show that the situation with the reserve labor army was not one of a long-run equi-
librium, and the industry departed from such a situation in mid-1890s. Due to the evolution of
the output market, firms had an incentive to poach skilled workers from other firms for efficient
109
production, and wages for workers increased in areas with high concentration of firms that had
relatively low costs to find skilled workers in other firms. Using geographically highly granu-
lated data, we show that increased demand for skilled workers dramatically changed the nature
of the labor market. In our regression analyses, we utilize monthly data covering the universe of
the industry, we examine how the increased competition influences wages. Evidence indicates
that the mechanism behind these changes was intra-industry worker raiding and that a handful of
firms (many of which were the early independent private entrants) eventually came on top in this
competition for higher-quality labor. We find expansions in peer firms? capacities by one standard
deviation from the mean increases the wages of focal firms for male workers by 7.1 percentage
points and for female workers by percentage points.
There is additional agenda not examined in this chapter, but important to consider in future
research. First, it is still not clear what precise mechanisms were that led to wage increases.
Previous literature shows that an imperfect labor market may be present for multiple reasons.
As analyzed, competition may directly affect wages by changing the market structure from the
monopsony to oligopsony or by increasing labor demand . In addition to such direct effects, there
may be some additional indirect effects of improved labor market structure. That is, competition
improves other labor market features, and these features improve wages. In that case, estimated
results in this chapter are the combination of direct and indirect effects. For instance, compe-
tition may reduce search frictions for workers to switch jobs or look for new jobs. There are
multiple studies considering the search friction as one of the important channels that can gen-
erate imperfect competition in the labor market (Burdett and Mortensen 1998, Manning 2003).
These empirical literature estimating search friction applies the model in Manning (2003), uti-
110
lizes matched employer-employee data (Hirsch, Schank and Schnabel 2010; Ransom and Oaxaca
2010). Firms may also improve internal rules and practices to treat workers. Okamoto (1993)
documents that some firms, such as Kurashiki Boseki, introduced a school within a factory to
educate their workers. Due to the limitation of the data on employees, we are not able to dig in
detail on such indirect mechanisms.
Second, although we assumed that more competition in labor market should occur due to
new entries and capacity expansion, historical anecdotes summarized in Okamoto (1993) suggest
that the labor market situation was not close to perfect competition. Because of the lack of any
agency that monitored competition, cotton spinning firms formulated a strong industrial associ-
ation to cooperate together to lower wages in the 1890s. In addition, there were a lot of cases
where new firms sent their workers to incumbent firms to educate them. Those stories suggest
that firms were not just pure rivals, but also likely to formulate cooperative relationships. As-
suming that the labor market was not perfectly competitive, wages might be much higher under
perfectly competitive situation.
We also have a plan to write a follow-up paper. After 1900, the industry became matured
and moved to the new stage. Instead of experiencing many new entries, merges, acquisitions,
construction of brunch plants and shut down started to occur in this period. This is characterized
as the stage where innovation and new knowledge no longer come from the outside but instead
shift to be developed inside the firms. This raises barriers to entry and leads to progressive
dominance of the industry by a handful of large firms. However, further disruptions can cause
repeated entry and shakeout cycles. As a future agenda, we plan to investigate how profitability
and productivity of firms relate to workers? compensation in this period.
111
Appendix A: Appendices for Chapter 1
A.1 Subjective attitudes on national unity and ethnicity
Attitudes toward cultural diversity and national unity is hard to measure. In my main anal-
ysis, I investigate outcome variables that are associated with cultural diversity and national unity.
In this section, I discuss effects of television exposure on subjective attitudes measured by large-
scale survey, Afrobarometer.
I use repeated cross-sectional data on the attitudes toward national and ethnic identities
from Afrobarometer (2000). Afrobarometer is a research institution conducting surveys on public
attitudes in more than 30 African countries. I use data from five waves of survey (wave 1-5) in
South Africa, conducted in 2000, 2002, 2006, 2008, and 2010. There are approximately 2,400
observations for each wave.
Table A.1 shows effect of exposure to television on subjective attitudes toward various
political opinions, using Afrobarometer. Since Afrobarometer takes multiple subjective attitudes
in Likert scales, I employ multiple specifications to test the effect of television. Column (1), (3),
(5), and (7) are the results with OLS, using linear probability model as follows:
yit = ?1Singalit + ?2FreeSpaceit + ExposeAreai +Xit? + ?it + ?st, (A.1)
112
where yit is a various outcomes for individual i at year t, Singalit is the measure of expo-
sure to television (either SABC1 or SABC2), FreeSpaceit is a measure of free-space signaling,
ExposeAreai is an indicator variable taking 1 if the area is ever exposed to television (i.e. in-
dicator for treated areas), Xit is a vector of controls (indicator for race and ethnicity, gender and
age), and ?it is a municipality and year fixed effects.1 ?1 is the parameter of interest, which is
the generalized difference-in-differences estimator of television exposure on subjective attitudes.
Table A.1 shows the estimated results. In column (1), (3), (5), and (7), outcome variables
are dummy variables, taking either 0 or 1. Column (2), (4), (6), and (8) are the results with fixed-
effect ordered logit model, suggested by Baetschmann, Staub and Winkelmann (2015).2 Panel A
reports the result using the dummy variable of television signal (taking 1 if the signal is stronger
than 60 dB?V/m, as same as the previous specifications), and Panel B reports the result using the
continuous measure of signal strength in dB?V/m as the main explanatory variable. As same as
the previous analyses, I restrict the sample to individuals who were exposed to television within
survey periods, or those without television exposure throughout the survey periods.
As the dependent variables, column (1) and (2) measure whether respondents are proud to
be South Africans. The original question asks whether a respondent agrees or disagrees with the
statement ?It makes you proud to be called a South African? with 5 scales (1: strongly disagree,
2: disagree, 3: neither agree nor disagree, 4: agree, 5: strongly agree). The binary variable in
the column (1) takes 1 if a respondent answers either 4 (agree) or 5 (strongly agree), while the
column (2) uses the Likert measure as it is. Column (3) and (4) measure whether respondents are
1Since the data structure of Afrobarometer is repeated cross-section, instead of panel, I could not include indi-
vidual fixed effects.
2Including fixed effects for logit and ordered logit model is challenging due to the well-known incidental param-
eters problem. Baetschmann et al. (2015) propose blow-up and cluster (BUC) estimator, which is almost as efficient
as more complex models such as GMM. See Baetschmann et al. (2015) for more detail. I use the stata module of
feologit by Baetschmann, Ballantyne, Staub and Winkelmann (2020).
113
desired to create a united South Africa. The original question asks whether a respondent agrees
or disagrees with the statement ?desirable to create a united South Africa out of all the groups?
with the same scale as the column (1) and (2). I generate the binary outcome variable in column
(3) as same way as the column (1). Column (5) and (6) measure whether respondents consider
their own ethnic group is treated unfairly by the national government (0: never, 1: sometimes, 2:
often, and 3: always). The binary dependent variable in column (5) takes 1 if a respondent thinks
his/her ethnic group is treated unfairly (answering 2 or 3). Column (7) and (8) measure whether
respondents are attached to ethnic identity or racial identity in five scales (1: ethnic identity only,
2: ethnic identity more than national identity, 3: both are equal, 4: national identity more than
ethnic identity, and 5: national identity only). The binary variable in column (7) takes 1 if a
respondent thinks that he/she is attached to ethnic identities (i.e. answering 1 or 2). The last two
questions only exist for the wave 3, 4 and 5, so that the number of observations are smaller than
the first two questions. The first two dependent variables indicate importance and pride toward
national identity, while the later two indicate importance toward ethnic identity.
114
115
Table A.1: Effects of television on subjective attitudes
(1) (2) (3) (4) (5) (6) (7) (8)
Proud SA Desire unite SA ethnic treated fairly ethnic vs national
OLS ordered logit OLS ordered logit OLS ordered logit OLS ordered logit
Panel A: Results with dummy variable (Signal stronger than 60 dbu)
Signal stronger than 60 dbu 0.0259 -0.139 0.0558 0.188 -0.0971 0.0922 0.129 0.0841
(0.0520) (0.397) (0.0502) (0.330) (0.0959) (0.483) (0.0845) (0.340)
Municipality FE Yes Yes Yes Yes Yes Yes Yes Yes
Year FE Yes Yes Yes Yes Yes Yes Yes Yes
Control Yes Yes Yes Yes Yes Yes Yes Yes
R2 0.114 0.133 0.336 0.132
N 1916 1842 1862 1834 1107 1042 1184 1126
Panel B: Results with continuous signal strength
Signal (dBu) of SABC -0.000564 -0.00890 0.000273 -0.00107 -0.00148 0.00619 0.00123 -0.00663
(0.000548) (0.00539) (0.000737) (0.00511) (0.00166) (0.00983) (0.000792) (0.00646)
Municipality FE Yes Yes Yes Yes Yes Yes Yes Yes
Year FE Yes Yes Yes Yes Yes Yes Yes Yes
Control Yes Yes Yes Yes Yes Yes Yes Yes
R2 0.114 0.132 0.337 0.133
N 1916 1842 1862 1834 1107 1042 1184 1126
? p < 0.05, ?? p < 0.01, ??? p < 0.001. Standard errors are clustered by municipality level. Column (1), (3), (5), and (7) use the linear probability model,
where the outcome variables are the dummy variables. Column (2), (4), (6), and (8) use the fixed-effect ordered logit model suggested by Baetschmann et
al. (2020). All the reported coefficients are marginal effects. For all regresions, I include indicator for race and ethnic groups, gender and age as controls.
Definitions of the dependent variables are explained in the main article.
Results indicate that there is no clear evidence that exposure to television change attitudes
toward ethnicity and nationality. None of the specification is statistically significant, and the sizes
of coefficients are small. This result may be driven by multiple reasons. First, since I restrict the
sample to those who were exposed to television during survey periods and those who were never
exposed to television during the period, the number of observations becomes small. Among
11,799 entire observations, I could use small portion of them. Second, since all answers in
Afrobarometer are not incentivized and respondents answer questions with their own subjective
measures, data may be too noisy to detect any effect in my analyses.
116
A.2 Additional figures and tables
A.2.1 Figures: the construction of transmitters for each year
(a) new transmitters between 2002-2003 (b) new transmitters between 2002-2005
(c) new transmitters between 2002-2007 (d) new transmitters between 2002-2009
Figure A.1: Construction of new transmitters by year
117
A.2.2 Figures: Different assumptions on clustering standard errors
As mentioned in Section 1.4 I cluster standard errors by municipality level. Here, Fol-
lowing figure reports the result with clustering school level, instead of municipality-level cluster.
Figure A.2: Event study estimate with school-level cluster
(a) Vote share for ANC (b) Vote share for IFP
Figure A.3: Effects of television exposure on vote shares for political parties, with municipality
fixed effects with voting ward level clustering
118
A.2.3 Figures: Effect of TV access on language choices: standard two-way
fixed effect
Figure A.4: Event study estimate with year and school fixed effects, standard TWFE
A.2.4 Figure: Effect of TV access on language choices: adding covariant from
Census
119
Figure A.5: Event study estimate with year and school fixed effects, additional covariant from
Census 2001, interacting with year
120
A.2.5 Figures: Effect of TV access on language choices with different cutoffs
Figure A.6: Event study estimate with year and school fixed effects: cutoff 55dB?V/m
Figure A.7: Event study estimate with year and school fixed effects: cutoff 65dB?V/m
121
Figure A.8: Event study estimate with year and school fixed effects: cutoff 70dB?V/m
122
A.2.6 Figures: Effect of TV access on voting outcomes
Figure A.9: Event study estimate with year and ward FEs: votes for ANC: cutoff 55dB?V/m
Figure A.10: Event study estimate with year and ward FEs: votes for IFP: cutoff 55dB?V/m
A.2.7 Tables: difference-in-differences
123
Figure A.11: Event study estimate with year and ward FEs: votes for ANC: cutoff 65dB?V/m
Figure A.12: Event study estimate with year and ward FEs: votes for IFP: cutoff 65dB?V/m
124
Figure A.13: Event study estimate with year and ward FEs: votes for ANC: cutoff 70dB?V/m
Figure A.14: Event study estimate with year and ward FEs: votes for IFP: cutoff 70dB?V/m
125
126
Table A.2: Difference-in-differences: Effect of SABC exposure on language choice
(1) (2) (3) (4) (5) (6) (7) (8) (9)
Panel A: Clustering by schools
Signal stronger than 60dbu 0.00628 0.00681 0.00549
(0.00444) (0.00444) (0.00455)
Signal (dBu) of SABC 0.000252?? 0.000264?? 0.000205?
(0.0000874) (0.0000875) (0.0000907)
Signal dummy ? linear time trend 0.00545??? 0.00523??? 0.00346??
(0.00110) (0.00110) (0.00111)
school FE Yes Yes Yes Yes Yes Yes Yes Yes Yes
Year FE Yes Yes Yes Yes Yes Yes Yes Yes Yes
Free-Space No No No Yes Yes Yes Yes Yes Yes
Census control No No No No No No Yes Yes Yes
Dep.Var.Mean 0.369 0.369 0.369 0.369 0.369 0.369 0.369 0.369 0.369
R2 0.599 0.599 0.600 0.599 0.600 0.600 0.603 0.603 0.603
N 52078 52078 52078 52078 52078 52078 52015 52015 52015
Panel B: Clustering by municipalities
Signal stronger than 60dbu 0.00631 0.00683 0.00549
(0.00611) (0.00601) (0.00638)
Signal (dBu) of SABC 0.000256 0.000268 0.000205
(0.000151) (0.000151) (0.000166)
Signal dummy ? linear time trend 0.00542 0.00521 0.00346
(0.00292) (0.00289) (0.00179)
school FE Yes Yes Yes Yes Yes Yes Yes Yes Yes
Year FE Yes Yes Yes Yes Yes Yes Yes Yes Yes
Free-Space No No No Yes Yes Yes Yes Yes Yes
Census control No No No No No No No No No
Dep.Var.Mean 0.369 0.369 0.369 0.369 0.369 0.369 0.369 0.369 0.369
R2 0.599 0.599 0.600 0.599 0.599 0.600 0.603 0.603 0.603
N 52023 52023 52023 52023 52023 52023 52015 52015 52015
? p < 0.05, ?? p < 0.01, ??? p < 0.001. Standard errors are clustered by school level in Panel A and municipality level in Panel B. All Specifications. exclude schools
which have access to television throughout the sample periods. Specifications (4) to (9) control for free-space signaling. Specification (6) to (9) include control variables
from census in 2001, interacted with year dummies. Controls include average age, education level, income level, sex and employment rate. Specifications (2), (5), and (8)
use the continuous number of signal strength, and specifications (1), (4), and (7) use the dummy variable of signal strength, where the dummy takes 1 if the signal is stronger
than 60 dBu at school locations. Specifications (3), (6) and (9) use the dummy variably of signal strength, interacted with linear trend.
Table A.3: Effect of SABC exposure on voting outcomes by political ward
Full observations Only IFP exist
Dep. var: Vote Share of ANC Dep. var: Vote Share of IFP
(1) (2) (3) (4) (5) (6)
Panel A: Results with municipality FEs
Signal (dBu) 0.000262? 0.000228 -0.0000810
(0.000131) (0.000185) (0.000181)
Signal stronger than 60dbu 0.0277??? 0.0202 -0.00672
(0.00799) (0.0137) (0.00884)
Year FE Yes Yes Yes Yes Yes Yes
Municipality FE Yes Yes Yes Yes Yes Yes
Dep.Var.Mean 0.667 0.667 0.492 0.492 0.397 0.397
R2 0.697 0.697 0.791 0.792 0.894 0.894
N 5075 5075 1597 1597 1597 1597
Panel B: Results with political wards FEs
Signal (dBu) 0.000370? -0.000146 0.000459
(0.000187) (0.000350) (0.000288)
Signal stronger than 60dbu 0.0273? 0.0193 -0.0107
(0.0106) (0.0167) (0.0137)
Year FE Yes Yes Yes Yes Yes Yes
Municipality FE Yes Yes Yes Yes Yes Yes
Dep.Var.Mean 0.667 0.667 0.480 0.480 0.422 0.422
R2 0.886 0.886 0.938 0.939 0.958 0.957
N 4663 4663 1456 1456 1456 1456
? p < 0.05, ?? p < 0.01, ??? p < 0.001. Standard errors are clustered by local-municipality level. All specifications
control for free-space signaling and a dummy for treated regions. Specifications (1), (3), and (5) use the continuous
number of signal strength, and specifications (2), (4), and (6) use the dummy variable of signal strength, where the
dummy takes 1 if the signal is stronger than 60 dBu at school locations. Specifications (3) and (4) restrict voting district,
where IFP receive non-zero vote shares.
127
Appendix B: Appendices for Chapter 2
B.1 Additional background information
B.1.1 Detailed background of the Capital Radio 604
In this section, I summarize detailed information regarding the Capital Radio 604, which
are not covered in the main article. Again, most contents are based on the autobiographic docu-
mentary by Johnston (2019).
As mentioned in the main article, Capital Radio 604 was established by taking advantage of
the legal defectiveness in newly independent homelands. Founders got an idea to establish radio
station mimicking Capital Radio 194 in the United Kingdom, which was established in 1972.
Capital Radio 194 was initially illegal in the U.K., which broadcasted their programs from a ship
in North Sea to avoid governmental regulations, because the public broadcasting was dominated
by BBC.
Founders of the Capital Radio 604 communicated well with the independent government
of Transkei to start up their own station. Initially, the station was owned by Transkei Government
with 51% share, founders with 26%, and future use for Transkei citizens with 23%1. The initial
station was located in Port St. Johns in Transkei. Some programmers from UK also came to help
1Transkei government seized 100% of the share, after receiving financial support (3/4 of the budget) from South
Africa in 1980.
128
starting the program.
After some preparation in Port St. Johns, they moved their studio to Johannesburg, the
largest city in South Africa and far a way from Transkei, and installed a line to Transkei. They
bought the largest transmitter in the southern hemisphere at the time: ?megawatt transmitter?
66dBm signal in Johannesburg, and locate the transmitter at Herchel in 1979 after convincing
local chiefs for the construction. However, since the ground conductivity at Herchel was weaker
then a consulting firm predicted, it failed to have enough signal in Johannesburg. They, in the
end, gave up the large transmitter in Herchel and constructed smaller transmitters in Umtata and
Umzimkhulu with 100kW capacities targeting Eastern Cape and Natal regions respectively in
the same year. Umtata was the central part of Transkei which was in favor of the independent
Transkei government, and Umzimkhulu was the north edge of Transkei which could maximize
the coverage in Natal region.
Contrasting to SABC which broadcasted programs in presenters? languages, Capital Radio
broadcasted their programs in English regardless of presenters? identities. As a consequence,
it became the first South African radio broadcasting Black presented in English, not their home
languages. Since the program sometimes contain contents conflicting with governmental policies
of the central South Africa, members of the station were sometimes attached by guns. While there
was no medium wave radio in South Africa at the period and it requires additional equipment
(though it was not that expensive to obtain), more than 200,000 people are said to listen to Capital
radio2.
Listeners were active, but the station started losing their audience in 1990s: they initially
2This number was self-reported by the engineers at the station, and need to take the number with cautions.
Therefore, in my main article, I present the advertising rate for Capital Radio to make the existence of listeners more
compelling.
129
targeted the White population in Durban with medium waves (sounds worse than FM), and not
targeting people in Transkei. However, in the late 1980s, one former worker turned in a proposal
to a Minister of Finance in Transkei, and change the program as ?station with Afritude?, tar-
geting African population in Transkei. This sharply decreased both listeners and advertisement
revenues. The program was terminated by the new government after the general election in 1994.
B.1.2 Other radio in independent Bantustan
Main information sources for the resident of Bantustan are Radio Bantu programs operated
by SABC, as explained in Section 2.2.2.1. In addition, independent homelands sometimes estab-
lished small radio stations targeting homeland population. According to Hamm (1991), programs
were ?largely overseen by the (central) government, but gradually developed programming pat-
terns which ran counter to Radio Bantu?s mission of stressing ethnic and linguistic separation?.
Therefore, they also played European and American music, which were not played in Bantu
programs.
Radio Freedom was the other alternative option, which had the strongest political mes-
sages. Radio station organized by ANC, the main illegal counter group, and send propaganda on
anti-Apartheid. The program was broadcasted from other countries, such as Ethiopia, Mozam-
bique, Zambia and Madagascar using short waves. Short wave radio uses lower frequencies than
medium (AM) and very-high (FM) frequencies, and it is able to send signals to much longer dis-
tance than other radio. Exiled ANC members in other countries used Radio Freedom to contact
ANC members in South Africa: ?served as a source of inspiration for ANC supporters, partic-
ularly for its loyal, youthful and politically savvy audience? and ?most effective way through
130
which the banned ANC could project its own voice and impact onto internal struggles against
the apartheid state? ( Lekgoathi 2010). According to Hamm (1991), most Black South African
could not listen to the short wave radio, as FM-AM receivers were inexpensive but short wave
receivers were not. Short-wave radio requires additional equipment, and most Black population
did not own them (Hamm 1991). In addition, one of the listener, Murphy Morobe, the leader
of South African Student Movement (SASM) mentioned in his interview that short wave radios
were quite scarce (Lekgoathi 2010), as well as the wave were likely to be blocked by the central
government. It was illegal to listen to Radio Freedom, and listeners would be punished: hard to
measure listenership (Lekgoathi 2010).
While Radio Freedom may have significant impact especially for Black South African ac-
tivists who countered Apartheid regime, it was hard to predict listenership as well as the coverage.
My estimate do not take into account the effect of Radio Freedom, because (1) due to the feature
of short-wave radio, coverage of radio seemed to be nation-wide and there were not geographical
variations, and (2) according to historical literature, majority of the population did not listened to
it.
B.1.3 Potential radio exposure from other countries
There may be potential exposure of radio from other countries. I summarize the potential
exposures in this subsection.
LM radio from Mozambique was popular among White South Africans in 1960s, before
the independence of Mozambique. As the South African government regulated rock and pop
music, LM radio broadcasted such music from outside the country. LM radio played American
131
and British rock and pop music. 45% of white matriculates listened LM radio, while only 20%
listened radio Springbok (music program by SABC) (Hamm 1991). LM?s popularity were purely
musical, not political because Mozambique was still under Portuguese control and no criticism
of South African policies. Obtaining the coverage of LM radio and estimating its impact on
voting behaviors, which may be able to isolate demand to popular culture from anti-Apartheid
motivation, is my future agenda. Swaziland Commercial Radio was the other radio program
which played popular music targeting South African White population. Their programs were
nothing to do with Swaziland, but played 702 and other music programs
B.2 Engineering model of calculating AM radio propagation
I follow Russo (2021)?s application of the engineer model of predicting AM radio coverage
on economic research. In Russo (2021), he uses simplified version of Sommersfeld-Norton model
(Trainotti 1990). The original Sommerfeld-Norton model contains some complicated functional
form which makes it difficult to compute radio coverage, and Trainotti?s model simplifies and
approximates the model without losing predicting power. In the Trainotti?s model, electrical field
(E, mV/m) is calculated as:
E = (30WD)1/2
AA1 (B.1)
R
where W is transmitter?s power (watts), D is antenna directivity and R is distance from transmit-
ter (meters).
Here, if p0 < 4.5,
?
A = exp(?0.43p0 + 0.01p20)? ( p0/2sin(b)exp((?5/8)p0)) (B.2)
132
and if p0 > 4.5
1 ?
A = ? ( p0/2sin(b)exp((?5/8)p0)) (B.3)
2p0 ? 3.7
where
?(R/1000)f 2cos(b) (+ 1)f
p0 = , b = . (B.4)
54? 102? 18? 103?
? is ground conductivity (siemens by meters),  is permittivity (highly correlated with ?), f is fre-
quency (mHz). While we have data on ground conductivity (see Figure 2.2), we do not know ex-
act permittivity. I use table regarding the relationship between ground conductivity and permittiv-
ity from CCIR (1992) and use the linear approximation of permittivity, where  = 8.035+1.12?.
Finally,
?A2R
A1 = (B.5)
104 ? 8/f (1/3)
where A2 is an experimentally determined coefficient which typically takes numbers around 0.25
(Trainotti 1990). We can convert electrical field in mV/m to Db/m by
E(Db/m) = 20? log10(E(mV/m)) + 120. (B.6)
In general, radio has good listenability if E(Db/m) > 88 in urban areas, E(Db/m) > 74 in
residential areas and E(Db/m) > 54 in rural areas. The method proposed by Trainotti (1990), as
well as the original Sommerfeld-Norton model, assume that the grand conductivity is same across
location. That is, the model does not take into account the attrition of radio waves occurring
due to the change in ground conductivity through the travel of waves. I follow Russo (2021) to
calculate weighted average of ground conductivity through travels of waves. Figure B.1 illustrates
the way I gave weight to calculate approximated ground conductivity. I first take average of
133
ground conductivities between a transmitter and receiver?s location, and repeat this calculation
once again to obtain the approximated ground conductivity. Figure B.2 shows the approximated
ground conductivity taking into account the different level of ground conductivity through the
travel of waves. When I take weighted average, each cell represents 100m2 of the ground.
T 3 3 6 T
? : Location of a transmitter
T 3 3 4 ? 3 + 3 + 6
= 4
3
3 + 3
= 3
2
T 3 3 10/3 ?
3 + 3 + 4 10
=
3 3
3 + 3
= 3
2
Figure B.1: Methods to take weighted average of ground conductivity
Figure B.2: Approximation of ground conductivity, taking into account the travel of waves
134
B.3 Additional Tables
B.3.1 Tables: results with different signal strength as treatment definitions
Table B.1: Difference in differences: Effects of AM radio on voting behavior
Treated if the signal strength is higher than 50dBu
(1) (2) (3) (4)
Left-wing Right-wing Middle-Left All-left
Panel A: Results with proportion of votes
Areas >50dBu? 1981 0.0904??? -0.0831??? -0.100?? -0.00992
(0.0265) (0.0169) (0.0388) (0.0372)
Location FE Yes Yes Yes Yes
Year FE Yes Yes Yes Yes
R2 0.701 0.668 0.740 0.833
N 801 651 801 801
Panel B: Results with the actual numebr of voters
Areas >50dBu? 1981 840.0?? -705.3??? 313.4 1153.4?
(266.7) (155.9) (406.2) (446.5)
Location FE Yes Yes Yes Yes
Year FE Yes Yes Yes Yes
R2 0.726 0.674 0.656 0.767
N 801 651 801 801
Panel C: Results with the log of actual numebr of voters
Areas >50dBu? 1981 1.256? 0.423 2.003? 2.128??
(0.592) (0.649) (0.791) (0.803)
Location FE Yes Yes Yes Yes
Year FE Yes Yes Yes Yes
R2 0.667 0.643 0.589 0.586
N 801 651 801 801
? p < 0.05, ?? p < 0.01, ??? p < 0.001
135
Table B.2: Difference in differences: Effects of AM radio on voting behavior
Treated if the signal strength is higher than 55dBu
(1) (2) (3) (4)
Left-wing Right-wing Middle-Left All-left
Panel A: Results with proportion of votes
Areas >55dBu? 1981 0.0978??? -0.104??? -0.0856? 0.0121
(0.0283) (0.0179) (0.0415) (0.0397)
Location FE Yes Yes Yes Yes
Year FE Yes Yes Yes Yes
R2 0.702 0.675 0.739 0.833
N 801 651 801 801
Panel B: Results with the actual number of voters
Areas >55dBu? 1981 924.4?? -858.2??? 705.0 1629.5???
(284.3) (165.4) (432.6) (474.2)
Location FE Yes Yes Yes Yes
Year FE Yes Yes Yes Yes
R2 0.727 0.678 0.657 0.769
N 801 651 801 801
Panel C: Results with the log of actual number of voters
Areas >55dBu? 1981 1.490? -0.491 3.035??? 2.754??
(0.631) (0.692) (0.839) (0.854)
Location FE Yes Yes Yes Yes
Year FE Yes Yes Yes Yes
R2 0.667 0.643 0.593 0.588
N 801 651 801 801
? p < 0.05, ?? p < 0.01, ??? p < 0.001
136
Table B.3: Difference in differences: Effects of AM radio on voting behavior
Treated if the signal strength is higher than 65dBu
(1) (2) (3) (4)
Left-wing Right-wing Middle-Left All-left
Panel A: Results with proportion of votes
Areas >65dBu? 1981 0.0685 -0.104?? 0.0143 0.0828
(0.0500) (0.0318) (0.0730) (0.0695)
Location FE Yes Yes Yes Yes
Year FE Yes Yes Yes Yes
R2 0.696 0.658 0.737 0.833
N 801 651 801 801
Panel B: Results with the actual number of voters
Areas >65dBu? 1981 760.3 -865.1?? 668.6 1428.8
(502.0) (292.4) (759.8) (837.7)
Location FE Yes Yes Yes Yes
Year FE Yes Yes Yes Yes
R2 0.723 0.665 0.656 0.765
N 801 651 801 801
Panel C: Results with the log of actual number of voters
Areas >65dBu? 1981 1.830 -0.788 3.628? 2.211
(1.109) (1.201) (1.479) (1.508)
Location FE Yes Yes Yes Yes
Year FE Yes Yes Yes Yes
R2 0.666 0.643 0.588 0.583
N 801 651 801 801
? p < 0.05, ?? p < 0.01, ??? p < 0.001
137
B.3.2 Tables: results using data after 1970
Table B.4: Difference in differences: Effects of AM radio on voting behavior
Treated if the signal strength is higher than 50dBu
(1) (2) (3) (4)
Left-wing Right-wing Middle-Left All-left
Panel A: Results with proportion of votes
Areas >50dBu? 1981 0.0965??? -0.0831??? -0.0904?? 0.00614
(0.0288) (0.0169) (0.0341) (0.0332)
Location FE Yes Yes Yes Yes
Year FE Yes Yes Yes Yes
R2 0.721 0.668 0.769 0.866
N 651 651 651 651
Panel B: Results with the actual number of voters
Areas >50dBu? 1981 948.6??? -705.3??? 199.0 1147.6??
(286.3) (155.9) (392.6) (427.4)
Location FE Yes Yes Yes Yes
Year FE Yes Yes Yes Yes
R2 0.744 0.674 0.708 0.806
N 651 651 651 651
Panel C: Results with the log of actual number of voters
Areas >50dBu? 1981 1.456? 0.423 1.910? 2.050??
(0.627) (0.649) (0.777) (0.776)
Location FE Yes Yes Yes Yes
Year FE Yes Yes Yes Yes
R2 0.684 0.643 0.641 0.647
N 651 651 651 651
? p < 0.05, ?? p < 0.01, ??? p < 0.001
138
Table B.5: Difference in differences: Effects of AM radio on voting behavior
Treated if the signal strength is higher than 55dBu
(1) (2) (3) (4)
Left-wing Right-wing Middle-Left All-left
Panel A: Results with proportion of votes
Areas >55dBu? 1981 0.106??? -0.104??? -0.0778? 0.0279
(0.0308) (0.0179) (0.0365) (0.0355)
Location FE Yes Yes Yes Yes
Year FE Yes Yes Yes Yes
R2 0.721 0.675 0.768 0.866
N 651 651 651 651
Panel B: Results with the actual number of voters
Areas >55dBu? 1981 1061.1??? -858.2??? 522.5 1583.6???
(305.3) (165.4) (418.6) (453.9)
Location FE Yes Yes Yes Yes
Year FE Yes Yes Yes Yes
R2 0.744 0.678 0.709 0.808
N 651 651 651 651
Panel C: Results with the log of actual number of voters
Areas >55dBu? 1981 1.715? -0.491 2.856??? 2.591??
(0.668) (0.692) (0.824) (0.826)
Location FE Yes Yes Yes Yes
Year FE Yes Yes Yes Yes
R2 0.684 0.643 0.646 0.649
N 651 651 651 651
? p < 0.05, ?? p < 0.01, ??? p < 0.001
139
Table B.6: Difference in differences: Effects of AM radio on voting behavior
Treated if the signal strength is higher than 65dBu
(1) (2) (3) (4)
Left-wing Right-wing Middle-Left All-left
Panel A: Results with proportion of votes
Areas >65dBu? 1981 0.0871 -0.104?? -0.00403 0.0830
(0.0539) (0.0318) (0.0637) (0.0614)
Location FE Yes Yes Yes Yes
Year FE Yes Yes Yes Yes
R2 0.716 0.658 0.766 0.867
N 651 651 651 651
Panel B: Results with the actual number of voters
Areas >65dBu? 1981 990.3 -865.1?? 766.3 1756.6?
(534.6) (292.4) (726.6) (793.6)
Location FE Yes Yes Yes Yes
Year FE Yes Yes Yes Yes
R2 0.740 0.665 0.709 0.805
N 651 651 651 651
Panel C: Results with the log of actual number of voters
Areas >65dBu? 1981 2.313? -0.788 3.920?? 2.540
(1.163) (1.201) (1.437) (1.444)
Location FE Yes Yes Yes Yes
Year FE Yes Yes Yes Yes
R2 0.683 0.643 0.642 0.644
N 651 651 651 651
? p < 0.05, ?? p < 0.01, ??? p < 0.001
140
B.3.3 Tables: others
Table B.7: Difference in differences: replication of Wilse-Samson (2013)
(1) (2) (3) (4)
Left-wing Right-wing Middle-Left All-left
Mine recruitment area ? treated time 0.0660??? -0.0328?? -0.0194 0.0467?
(0.0157) (0.0108) (0.0232) (0.0220)
Location FE Yes Yes Yes Yes
Year FE Yes Yes Yes Yes
R2 0.704 0.657 0.737 0.834
N 801 651 801 801
? p < 0.05, ?? p < 0.01, ??? p < 0.001
B.4 Additional figures
141
(a) proportion of left-wing party (b) proportion of right-wing party
(c) proportion of middle-left party (d) proportion of left wing parties, combined
Figure B.3: Proportion of parties by location and year
142
Figure B.4: Turnout by year for different by exposure to AM radio
143
Appendix C: Appendices for Chapter 3
Figure C.1: Proportion of the cotton spinning industry and the textile industry to GNP
In Figure C.3 we illustrate this on the example of Osaka Spinnings first capacity expansion.
As the solid blue line shows, in the second half of 1885 the female wage distribution was already
spread out and rather symmetric. In 1886 the mill embarked on its first big expansion, tripling the
capacity from 10,500 to 31,320 spindles. The hands to work with new machines were expanded
too; the number of female workers on fixed daily wages (itself about half of the total number of
144
(a) Male wages (b) Female wages
Figure C.2: Male and Female Workers?s average wages, 1883-84 and 1888-89, comparison be-
tween Osaka and others
Figure C.3: Osaka Spinning mill female wage distribution around first capacity expansion
145
female workers employed) rose from 59 to 130 between the second half of 1885 and the first half
of 1886 and then to 317 in the second half of 1886. Figure C.3 clearly shows how this changed
the shape of the wage distribution: the two solid lines, orange and red, corresponding to wage
distributions in the first and second half of 1886, respectively, show an increasing spike at the
lowest daily wage of five sen and general fattening of the left tail especially in the second half of
1886. Clearly, most of almost 300 newly hired female workers were unskilled and by the second
half of 1886, their low wages dominate the whole distribution. The average wage fell 36 percent
from the second half of 1885 to the second half of 1886. After that, the number of female workers
stabilized and as the newly hired hands acquired more skill, the distribution started spreading out
again until it basically going back to what it was prior to expansion by the first half of 1888 (solid
green line). The average wage also almost recovered although it remained somewhat below the
level achieved at the end of 1885. The impact of the next round of expansion (doubling the plant
capacity and the number of female workers in 1889-1890, not shown) is very similar.
146
147
Table C.1: First stage of the 2SLS regression dependent variable: number of male and female workers
Firms in 20km as ?neighbors? Firms in 50km as ?neighbors?
l = 2 l = 6 l = 2 l = 6
(1) (2) (3) (4) (5) (6) (7) (8)
Female Male Female Male Female Male Female Male
# of Male Workers -0.184? -0.126? -0.304 ? -0.159 ? -0.0903 -0.0894 -0.215? -0.0667
(0.093) (0.084) (0.239) (0.0914) (0.051) (0.057) (0.098) (0.0410)
# of Female Workers -0.0203? -0.00777 -0.0500? -0.0106 -0.0227 -0.0129 -0.00964 -0.0101
(0.0092) (0.00554) (0.0281) (0.0195) (0.0125) (0.00954) (0.0074) (0.0085)
Firm?s capacity (by # of spindles) 0.000427 0.000245 -0.00129 -0.0000256 0.000937 0.000355 -0.000152 0.000172
(0.000776) (0.000235) (0.00125) (0.000372) (0.000788) (0.000225) (0.000935) (0.000249)
Total output 0.0175 0.00518 0.0610 0.0178 0.00605 0.00112 0.0301 0.00905
(0.0221) (0.00588) (0.0345) (0.00934) (0.0252) (0.00693) (0.0271) (0.00678)
Observations 3144 3146 2905 2906 1926 1928 1764 1765
R2 0.968 0.965 0.933 0.929 0.970 0.969 0.941 0.941
F(4,25) 2.54 3.64 4.45 2.45
F(4,20) 2.52 2.84 3.13 2.05
Standard errors are clustered by previous feudal domain level, which are shown in parentheses (25 clusters for the regression (1)-(4), and 20 clusters in
(5)-(8).) All regressions are estimated by LSDV(Least Square Dummy Variables), including firm fixed effects, year fixed effects, month fixed effects and own
lag variables (# of male and female workers, # of total capacity and total output). ? p < 0.05, ?? p < 0.01, ??? p < 0.001
148
Table C.2: Peer effects from other firms in 20 km: six months lag regressors
Female Wage Male Wage Wage gap
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11)
# of Male Workers 4.802??? 1.955 10.51??? -4.780
(0.00) (0.466) (0.00) (0.208)
# of Female Workers 1.273??? 0.812 2.932??? 1.654
(0.00) (0.208) (0.00) (0.334)
Firm?s capacity 0.0447?? 0.0245 0.0959??? 0.0290 0.000103
(by # of spindles) (0.006) (0.21) (0.00) (0.048) (0.992)
Total output 0.472??? -0.266 1.195??? 0.765
(0.00) (0.254) (0.00) (0.38)
Observations 2219 2219 2219 2219 2219 2218 2218 2218 2218 2218 2217
R2 0.914 0.914 0.914 0.912 0.915 0.891 0.893 0.890 0.894 0.894 0.686
P-values are in parentheses, which are calculated by the wild-t bootstrap method suggested by Cameron et al. (2008). Each observation is clustered by
previous feudal domain level (27 regions). All regressions are estimated by LSDV(Least Square Dummy Variables), including firm fixed effects, year fixed
effects, month fixed effects, own lag variables (# of male and female workers, # of total capacity and total output), current number of workers and capitals, and
reservation wages for agriculture and weavers. ? p < 0.05, ?? p < 0.01, ??? p < 0.001
149
Table C.3: Peer effects from other firms in 20 km: twelve months lag regressors
Female Wage Male Wage Wage gap
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11)
# of Male Workers 5.370??? 5.982? 10.77??? 2.729
(0.00) (0.032) (0.00) (0.618)
# of Female Workers 1.357??? -1.258 2.941??? -1.549
(0.00) (0.242) (0.00) (0.348)
Firm?s capacity 0.0544??? 0.0402??? 0.0903??? 0.0237? -0.00190
(by # of spindles) (0.00) (0.00) (0.00) (0.046) (0.388)
Total output 0.589??? 0.0842 1.383??? 1.495
(0.00) (0.698) (0.00) (0.114)
Observations 2037 2037 2037 2037 2037 2037 2037 2037 2037 2037 2036
R2 0.914 0.913 0.916 0.913 0.917 0.892 0.892 0.889 0.895 0.896 0.697
P-values are in parentheses, which are calculated by the wild-t bootstrap method suggested by Cameron et al. (2008). Each observation is clustered by previous
feudal domain level (27 regions). All regressions are estimated by LSDV(Least Square Dummy Variables), including firm fixed effects, year fixed effects, month
fixed effects, own lag variables (# of male and female workers, # of total capacity and total output), current number of workers and capitals, and reservation wages
for agriculture and weavers. ? p < 0.05, ?? p < 0.01, ??? p < 0.001
150
Table C.4: Peer effects from other firms in 50 km: two months lag regressors
Female Wage Male Wage Wage gap
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11)
# of Male Workers 4.961??? 1.376 9.473??? 0.247
(0.00) (0.426) (0.00) (0.92)
# of Female Workers 1.341??? 1.116??? 2.572??? 1.768???
(0.00) (0.00) (0.00) (0.00)
Firm?s capacity 0.0471??? 0.0322??? 0.0896??? 0.0596??? -0.00106
(by # of spindles) (0.00) (0.00) (0.00) (0.00) (0.724)
Total output 0.475??? -0.414?? 0.950??? -0.348??
(0.00) (.002) (0.00) (.002)
Observations 1387 1387 1387 1387 1387 1385 1385 1385 1385 1385 1385
R2 0.918 0.919 0.920 0.915 0.921 0.911 0.911 0.913 0.908 0.914 0.647
P-values are in parentheses, which are calculated by the wild-t bootstrap method suggested by Cameron et al. (2008). Each observation is clustered by previous
feudal domain level (27 regions). All regressions are estimated by LSDV(Least Square Dummy Variables), including firm fixed effects, year fixed effects, month
fixed effects, own lag variables (# of male and female workers, # of total capacity and total output), current number of workers and capitals, and reservation wages
for agriculture and weavers. ? p < 0.05, ?? p < 0.01, ??? p < 0.001
151
Table C.5: Peer effects from other firms in 50 km: six months lag regressors
Female Wage Male Wage Wage gap
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11)
# of Male Workers 5.291??? 4.403 11.21??? -1.007
(0.00) (0.052) (0.00) (0.758)
# of Female Workers 1.375??? 0.227 3.052??? 0.0477
(0.00) (0.814) (0.00) (1.00)
Firm?s capacity 0.0524??? 0.0346??? 0.101??? 0.0101 -0.00122
(by # of spindles) (0.00) (0.00) (0.00) (0.574) (0.636)
Total output 0.517??? -0.338??? 1.252? 1.241???
(0.00) (0.048) (0.00) (0.00)
Observations 1296 1296 1296 1296 1296 1294 1294 1294 1294 1294 1294
R2 0.922 0.921 0.922 0.919 0.924 0.921 0.922 0.916 0.924 0.924 0.648
P-values are in parentheses, which are calculated by the wild-t bootstrap method suggested by Cameron et al. (2008). Each observation is clustered by previous
feudal domain level (27 regions). All regressions are estimated by LSDV(Least Square Dummy Variables), including firm fixed effects, year fixed effects, month
fixed effects, own lag variables (# of male and female workers, # of total capacity and total output), current number of workers and capitals, and reservation
wages for agriculture and weavers. ? p < 0.05, ?? p < 0.01, ??? p < 0.001
152
Table C.6: Peer effects from other firms in 50 km: twelve months lag regressors
Female Wage Male Wage Wage gap
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11)
# of Male Workers 5.773??? 7.658??? 11.79??? 2.123
(0.00) (0.00) (0.00) (0.14)
# of Female Workers 1.457??? -1.865? 3.216??? -2.683??
(0.00) (0.02) (0.00) (0.002)
Firm?s capacity 0.0629??? 0.0485??? 0.102??? 0.0232 -0.00299
(by # of spindles) (0.00) (0.00) (0.00) (0.164) (0.266)
Total output 0.631??? 0.108 1.502??? 2.146???
(0.00) (0.554) (0.00) (0.00)
Observations 1195 1195 1195 1195 1195 1194 1194 1194 1194 1194 1194
R2 0.920 0.917 0.924 0.918 0.926 0.922 0.922 0.917 0.927 0.928 0.669
P-values are in parentheses, which are calculated by the wild-t bootstrap method suggested by Cameron et al. (2008). Each observation is clustered by previous
feudal domain level (27 regions). All regressions are estimated by LSDV(Least Square Dummy Variables), including firm fixed effects, year fixed effects, month
fixed effects, own lag variables (# of male and female workers, # of total capacity and total output), current number of workers and capitals, and reservation wages
for agriculture and weavers. ? p < 0.05, ?? p < 0.01, ??? p < 0.001
Table C.7: Peers? capacity expansion on own wages: different specifications
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10)
Female Male Female Male Female Male Female Male Female Male
[1em] Firm?s capacity 0.0376??? 0.0858??? 0.0330??? 0.0803??? 0.0328??? 0.0762??? 0.0138??? 0.0315??? 0.0304??? 0.0741???
(by # of spindles) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000)
Firm fixed effect X X X X X X X X X X
Time fixed effect X X X X X X X X X X
Own lagged variables X X X X X X X X
Own current variables X X X X X X X X
IV for current # of worker X X
Own wages 2 periods ago X X
Observations 2368 2368 2371 2371 2389 2389 2368 2368 3514 3512
P-values are in parentheses, which are calculated by the wild-t bootstrap method suggested by Cameron et al. (2008). Each observation is clustered by previous
feudal domain level (27 regions). All regressions are estimated by LSDV(Least Square Dummy Variables), including firm fixed effects, year fixed effects, month
fixed effects. Specification (1) and (2) are the benchmark, which are as same specification as the column (4) and (9) in Table 3.3. Specification (3) and (4) exclude
own current variables (# of workers and # of capitals). Specifcation (5) and (6) excludes own lagged variables(# of male and female workers, # of total capacity and
total output). Specification (7) and (8) in addition include lagged wages. ? p < 0.05, ?? p < 0.01, ??? p < 0.001
153
Table C.8: Peer effects from other firms in 20 km: two months lag regressors with different criteria of the sample
Less than 3 peers Less than 4 peers Less than 5 peers Less than 6 peers
(1) (2) (3) (4) (5) (6) (7) (8)
Female Male Female Male Female Male Female Male
Firm?s capacity (by # of spindles) 0.0388? 0.0891??? 0.0376?? 0.0858??? 0.0324 0.0750?? 0.0308 0.0742??
(0.02) (0.00) (0.01) (0.00) (0.10) (0.01 ) (0.15) (0.008)
Observations 2209 2208 2368 2367 2599 2598 2659 2658
R2 0.916 0.889 0.912 0.886 0.908 0.873 0.907 0.871
P-values are in parentheses, which are calculated by the wild-t bootstrap method suggested by Cameron et al. (2008). Each observation
is clustered by previous feudal domain level (27 regions). All regressions are estimated by LSDV(Least Square Dummy Variables),
including firm fixed effects, year fixed effects, month fixed effects, own lag variables (# of male and female workers, # of total capacity
and total output), current number of workers and capitals, and wages in agricultural sector. ? p < 0.05, ?? p < 0.01, ??? p < 0.001
154
Table C.9: Peer effects from other firms in 20 km: two months lag regressors with imputing missing variables
Female Wage Male Wage
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10)
# of Male Workers 3.136 -4.474 9.223??? -1.777
(1.740) (3.503) (1.159) (4.521)
# of Female Workers 1.024? 1.616 2.678??? 1.981
(0.438) (0.841) (0.293) (1.716)
Firm?s capacity 0.0376??? 0.0289? 0.0818??? 0.0362?
(by # of spindles) (0.00962) (0.0131) (0.0131) (0.0166)
Total output 0.373 -0.0300 1.008??? 0.165
(0.189) (0.190) (0.173) (0.483)
Observations 4939 4939 4939 4939 4939 4941 4941 4941 4941 4941
R2 0.889 0.890 0.891 0.890 0.892 0.865 0.866 0.864 0.865 0.867
P-values are in parentheses, which are calculated by the wild-t bootstrap method suggested by Cameron et al. (2008). Missing observations are
imputed if I could not find any reason of missing. Each observation is clustered by previous feudal domain level (27 regions). All regressions are
estimated by LSDV(Least Square Dummy Variables), including firm fixed effects, year fixed effects, month fixed effects, own lag variables (# of
male and female workers, # of total capacity and total output), current number of workers and capitals. ? p < 0.05, ?? p < 0.01, ??? p < 0.001
155
Table C.10: Peer effects from other firms in 20 km: two months lag regressors with imputing missing variables
Female Wage Male Wage
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10)
# of Male Workers 3.234 -4.316 9.491??? -1.466
(1.829) (3.567) (1.255) (4.628)
# of Female Workers 1.056? 1.532 2.751??? 1.945
(0.461) (0.839) (0.321) (1.723)
Firm?s capacity 0.0377??? 0.0288? 0.0822??? 0.0362?
(by # of spindles) (0.00953) (0.0134) (0.0132) (0.0175)
Total output 0.384 -0.00711 1.023??? 0.165
(0.193) (0.194) (0.180) (0.477)
Observations 5143 5143 5143 5143 5143 5143 5143 5143 5143 5143
R2 0.891 0.892 0.892 0.891 0.893 0.863 0.865 0.863 0.864 0.866
P-values are in parentheses, which are calculated by the wild-t bootstrap method suggested by Cameron et al. (2008). Missing observations
are imputed for all observations. Each observation is clustered by previous feudal domain level (27 regions). All regressions are estimated by
LSDV(Least Square Dummy Variables), including firm fixed effects, year fixed effects, month fixed effects, own lag variables (# of male and
female workers, # of total capacity and total output), current number of workers and capitals. ? p < 0.05, ?? p < 0.01, ??? p < 0.001
C.1 Simple model of monopsony and peer effects
In this section, we consider a simple model of monopsony and show why it is reasonable
to assume that the lagged variables of the focal firm and peer firms affect current wages and the
number of workers in our empirical model. In addition, we show why we need to take care of
potential simultaneity bias. Let?s start from the simplest static model. In a complete competition
in labor market, we have a profit function which is
? = R(L)?WL, (C.1)
where R(L) is a revenue function of the firm, L represents the number of workers, and W is the
wage. We consider the model with short run, so that firms cannot choose capital. Obviously,
FOC gives
R?(L) = W. (C.2)
On the other hand, under a pure monopsony, firms face an upward-sloping labor supply
function (W (L)), which implies wage is an increasing function of employment. In this case, we
have the profit function as
? = R(L)?W (L)L, (C.3)
and FOC as
R?(L) = W ?(L)L+W (L). (C.4)
Here, firms choose W and L simultaneously, so that if we regress W on L (or L on W ), then
estimated coefficients are not causal due to the well-known simultaneity bias. As a consequence,
156
estimated coefficients for any covariants (say X) are going to be biased as far as X and L are
correlated.
Now, let?s consider the case with multiple firms in a same local labor market, but firms still
hold market powers to determine wages (i.e. imperfect competition). Denote i as an identifier of
firms. Firms are going to maximize
?i = R(Li)?Wi(Li, Lj(?i),Wj(?i))Li. (C.5)
That is, the labor supply function faced by individual firms are going to be function of not only
Li but also Lj(?i) and Wj(?i), which are the labor and wages for peer firms. If the firm i exactly
knows Lj(?i) and Wj(?i), then they could just take the first order condition as same as in the
equation (C.4). However, if firms in a same labor market choose wages and labors at the same
time, the firm i cannot observe actual labor supply function, because she does not know actual
Lj(?i) andWj(?i). Instead, it is fair enough to assume that firms use previous information of other
firms to predict the labor supply function. Let?s add t subscripts for time, and denote Xi,t?1 and
Xj(?i),t?1 as vectors of L,W and other potential production variables such as capital and outputs.
Firms now choose Li by maximizing
?i = R(Li)? E[Wi(Li, Lj(?i),Wj(?i))|Xi,t?1, Xj(?i),t?1, Xi,t?2, Xj(?i),t?2, ...]Li. (C.6)
That is, firms make the best guess for the the labor supply function by using previous observations
for themselves and others (i.e. Xi,t?1, Xj(?i),t?1, Xi,t?2, Xj(?i),t?2, ...) to maximize their profits
157
(Let?s denote this as Case 1). Of course, there is no reason to include all the history of Xi,. and
Xj(?i),.. In that case, we either assume Xi,t?1 and Xj(?i),t?1 carry all the previous information
(Case 2), or there exist some costs to recall all the previous information (Case 3). It is important
to note that in Case 1, including Xi,t?1, Xj(?i),t?1, Xi,t?2, Xj(?i),t?2, ... to the model implies this
list of variables is going to affect both Wi and Li. In Case 2 only Xi,t?1 and Xj(?i),t?1 affect
Wi and Li, and Xi,t?2, Xj(?i),t?2, ... do not have any additional explanatory power to predict Wi
nor Li. Finally, in Case 3, firms could gain more profit by also using Xi,t?2, Xj(?i),t?2, ... if we
ignore the cost of recalling.
If we think about the FOC for the equation (C.6), it is obvious that optimal Li,t is a func-
tion of Wi,t and Xi,t?1, Xj(?i),t?1, Xi,t?2, Xj(?i),t?2, .... Li,t and Wi,t are still simultaneously
determined. As a consequence, regressing Wi,t on Li,t and Xi,t?1, Xj(?i),t?1 by OLS (or LSDV)
in the equation (3.1) becomes inconsistent.
One may think that additional lag variables, Xi,t?2, Xj(?i),t?2, ..., can be instruments to
take into account the simultaneity bias. Here, we show that these variables cannot be valid in-
struments in any case. In the Case 1, firms use Xi,t?2, Xj(?i),t?2, ... for maximizing their profit.
Therefore, Xi,t?2, Xj(?i),t?2, ... are going to affect both Wi,t and Li,t, and this is going to be a
violation of exclusion restriction. That is, there is no credible reason to believe that these vari-
ables only affect Li,t but not Wi,t. In the Case 2 and Case 3, firms do no use Xi,t?2, Xj(?i),t?2, ...
because these variables do not contain any additional information to improve profits. As a con-
sequence, Xi,t?2, Xj(?i),t?2, ... should not predict Li,t after controlling Xi,t?1 and Xj(?i),t?1, and
thus they are invalid to be instruments (weak instruments). If there is any correlation between
Xi,t?2, Xj(?i),t?2, ... and Li,t, this implies that firms use those previous information for maxi-
mization.
158
Yet, if we think the specific production technologies in the cotton spinning industry, firms
might only have freedom to choose wages, but not the number of workers because of the high op-
portunity cost of idling machines. In this case, firms just choose the level of wages by competing
against peer firms.
C.2 Discussions on estimating dynamic panel model
While the reverse causality is not a problem due to the lagged feature in equation 3.1,
estimated parameters are going to be mechanically biased if we estimate the model with stan-
dard fixed effect approach by subtracting within and between means (denote X?i,. and X?.,t?k) and
adding the overall mean (denote X? .,.). Applying the argument in Arellano and Bover (1995) and
other papers on estimating dynamic panel models, Xi,t that is included in X?i,. may mechanically
correlate with ?it, so that estimated parameters would be biased. In general, dynamic panel liter-
ature addresses this problem by running GMM instrumenting endogenous regressors by lagged
variables (i.e. (yi,t?k?1, yi,t?k?2..., Xi,t?k?1, Xi,t?k?2...)), but this is feasible only when T is rel-
atively small (see detail discussion in Baltagi (2013)). Because the data I am using involves
approximately 200 time periods, there are too many moment conditions and the results are going
to be biased and under-powered. Therefore, I do not use GMM to estimate the model, and instead
directly estimate fixed effects by including all the dummies for individual firms. It is possible to
do so since the data contains long time periods with relatively small number of firms. That is,
we can avoid the well-known incidental parameter problem. As shown analytically by Nickell
(1981), linear dynamic panel models with fixed effects are going to be biased if T is small, but
159
the bias term goes to 0 as T ? ?. Note that Kiviet (1995) points out that the result on Nickell
(1981) does not hold under a small number of observations (i.e. not assuming N ? ? asymp-
totic). Yet, Galiani and Gonzalez-Rozada (2005) showed Monte Carlo simulation that the bias
for the covariant (x) is small under the dynamic panel model withN = 50 and T = 50. Since our
analyses have similar N with much longer T (= 156), bias caused by estimating the model with
LSDV is expected to be small. For potentially nonlinear panel models, Hahn and Kuersteiner
(2002) propose the way to correct bias due to nonlinearity.
160
Bibliography
Abel, Martin, ?Long-run effects of forced resettlement: evidence from Apartheid South Africa,?
The Journal of Economic History, 2019, 79 (4), 915?953.
Abowd, John M, Francis Kramarz, and David N Margolis, ?High wage workers and high
wage firms,? Econometrica, 1999, 67 (2), 251?333.
Adena, Maja, Ruben Enikolopov, Maria Petrova, Veronica Santarosa, and Ekaterina Zhu-
ravskaya, ?Radio and the Rise of the Nazis in Prewar Germany,? The Quarterly Journal of
Economics, 2015, 130 (4), 1885?1939.
Advertising and Press Annual of Southern Africa, ?Advertising and Press Annual of Southern
Africa,? 1980.
Afrobarometer, ?South Africa, Round 1-5,? 2000. Restricted geocoded data https://
afrobarometer.org/.
Akerlof, George A and Rachel E Kranton, ?Economics and identity,? The quarterly journal of
economics, 2000, 115 (3), 715?753.
ANC, African National Congress, ?Nation-formation and nation building: The national ques-
tion in SouthAfrica,? 1997.
Anderson, T.W. and Cheng Hsiao, ?Formulation and estimation of dynamic models using panel
data,? Journal of Econometrics, jan 1982, 18 (1), 47?82.
Angrist, Joshua D and Victor Lavy, ?The Effect of a Change in Language of Instruction on
the Returns to Schooling in Morocco,? Journal of Labor Economics, 1997, 15 (1, Part 2),
S48?S76.
Arellano, Manuel and Olympia Bover, ?Another Look at the Instrumental Variable Estimation
of Error-components Models,? Journal of Econometrics, 1995, 68 (1), 29?51.
Armand, Alex, Paul Atwell, and Joseph Gomes, ?The reach of radio: Ending civil conflict
through rebel demobilization,? American Economic Review, 2020, 110 (5), 1395?1429.
Baetschmann, Gregori, Alexander Ballantyne, Kevin E Staub, and Rainer Winkelmann,
?feologit: A new command for fitting fixed-effects ordered logit models,? The Stata Journal,
2020, 20 (2), 253?275.
161
, Kevin E Staub, and Rainer Winkelmann, ?Consistent estimation of the fixed effects or-
dered logit model,? Journal of the Royal Statistical Society: Series A (Statistics in Society),
2015, 178 (3), 685?703.
Baines, Gary, ?The rainbow nation? Identity and nation building in post-apartheid South Africa,?
Mots pluriels, 1998, 7, 1?12.
Baltagi, Badi H., Econometric Analysis of Panel Data, 5th ed., West Sussex, U.K.: Willy, 2013.
Banerjee, Abhijit V, Esther Duflo, Clement Imbert, and Rohini Pande, ?Entry, exit and can-
didate selection: Evidence from India,? 2013.
Barnard, Ian, ?The language of multiculturalism in South African soaps and sitcoms,? Journal
of Multicultural Discourses, 2006, 1 (1), 39?59.
Barnett, Clive, ?Broadcasting the Rainbow Nation: Media, Democracy, and Nation-Building in
South Africa,? Antipode, 1999, 31 (3), 274?303.
Baumann, Gerd, The multicultural riddle: Rethinking national, ethnic and religious identities,
Routledge, 2002.
Bazzi, Samuel, Arya Gaduh, Alexander D Rothenberg, and Maisy Wong, ?Unity in diversity?
How intergroup contact can foster nation building,? American Economic Review, 2019, 109
(11), 3978?4025.
Beall, Jo, Sibongiseni Mkhize, and Shahid Vawda, ?Emergent democracy and resurgenttradi-
tion: Institutions, chieftaincy and transition in KwaZulu-Natal,? Journal of Southern African
Studies, 2005, 31 (4), 755?771.
Beaman, Lori, Raghabendra Chattopadhyay, Esther Duflo, Rohini Pande, and Petia
Topalova, ?Powerful women: does exposure reduce bias?,? The Quarterly journal of eco-
nomics, 2009, 124 (4), 1497?1540.
Berman, Lex, ?Japan Tokugawa GIS - Japan Historical GIS,? 2017.
Bertrand, Marianne and Esther Duflo, ?Field experiments on discrimination,? in ?Handbook
of economic field experiments,? Vol. 1, Elsevier, 2017, pp. 309?393.
Blouin, Arthur and Sharun W Mukand, ?Erasing ethnicity? Propaganda, nation building, and
identity in Rwanda,? Journal of Political Economy, 2019, 127 (3), 1008?1062.
Booysen, Susan, The African National Congress and the regeneration of political power, Wits
University Press, 2011.
Botha, Susan, ?South Africa?s party system,? Journal of theoretical politics, 1996, 8 (2), 209?
225.
Braguinsky, Serguey, ?Knowledge diffusion and industry growth: the case of Japans early cot-
ton spinning industry,? Industrial and Corporate Change, 2015, 24 (4), 769?790.
162
and David A Hounshell, ?History and nanoeconomics in strategy and industry evolution re-
search: Lessons from the Meiji-Era Japanese cotton spinning industry,? Strategic Management
Journal, 2016, 37 (1), 45?65.
, Atsushi Ohyama, Tetsuji Okazaki, and Chad Syverson, ?Acquisitions, Productivity, and
Profitability: Evidence from the Japanese Cotton Spinning Industry,? American Economic Re-
view, 2015, 105 (7), 2086?2119.
Brown, Charles and James Medoff, ?The employer size-wage effect,? Journal of political
Economy, 1989, 97 (5), 1027?1059.
Brummund, Peter, ?Variation in Monopsonistic Behavior Across Establishments: Evidence
From the Indonesian Labor Market,? Working Paper, 2012.
Burdett, Kenneth and Dale T. Mortensen, ?Wage Differentials, Employer Size, and Unem-
ployment,? International Economic Review, 1998, 39 (2), 257.
Callaway, Brantly and Pedro HC SantAnna, ?Difference-in-differences with multiple time
periods,? Journal of Econometrics, 2020.
Cameron, A. Colin, Jonah B. Gelbach, and Douglas L. Miller, ?Bootstrap-Based Improve-
ments for Inference with Clustered Errors,? Review of Economics and Statistics, 2008, 90 (3),
414?427.
CCIR, ?Recommendation 832: World Atlas of Ground Conductivity,? Technical Report 1992.
Charney, Craig R, ?Towards rupture or stasis? An analysis of the 1981 South African general
election,? African Affairs, 1982, 81 (325), 527?545.
Chisholm, Linda, ?Apartheid education legacies and new directions in post-apartheid South
Africa,? Apartheid Education Legacies and New Directions in Post-Apartheid South Africa,
2012, pp. 81?103.
Clark, Gregory, ?Why isn?t the whole world developed? Lessons from the cotton mills,? The
Journal of Economic History, 1987, 47 (1), 141?173.
Clots-Figueras, Irma and Paolo Masella, ?Education, language and identity,? The Economic
Journal, 2013, 123 (570), F332?F357.
Corno, Lucia, Eliana La Ferrara, and Justine Burns, ?Interaction, stereotypes and perfor-
mance: Evidence from South Africa,? Technical Report, IFS Working Papers 2018.
Crush, Jonathan, ??The long-term averted clash?: Labour competition in the South African
countryside,? Canadian Journal of African Studies, 1993, 27 (3), 404?423.
Dainihon Menshi Boseki Dogyo Rengokai, Menshi-Boseki Shokkou Jijo (Workers in the Cotton
Spinning Industry), Osaka Insatsu Inc., 1898.
163
Dasgupta, Nilanjana and Anthony G Greenwald, ?On the malleability of automatic attitudes:
combating automatic prejudice with images of admired and disliked individuals.,? Journal of
personality and social psychology, 2001, 81 (5), 800.
DBE, Department of Basic Education, Report on Dropout and Learner Retention Strategy to
Portfolio Committee on Education 2011.
de Chaisemartin, Cle?ment and Xavier d?Haultfoeuille, ?Two-way fixed effects estimators with
heterogeneous treatment effects,? Technical Report, National Bureau of Economic Research
2019.
DellaVigna, Stefano and Ethan Kaplan, ?The Fox News effect: Media bias and voting,? The
Quarterly Journal of Economics, 2007, 122 (3), 1187?1234.
, Ruben Enikolopov, Vera Mironova, Maria Petrova, and Ekaterina Zhuravskaya,
?Cross-border media and nationalism: Evidence from Serbian radio in Croatia,? American
Economic Journal: Applied Economics, 2014, 6 (3), 103?32.
Department of Communication, ?Broadcasting digital migration policy for South Africa,? Gov-
ernment Gazette, 2008, 31408.
Depetris-Chauvin, Emilio, Ruben Durante, and Filipe R Campante, ?Building nations
through shared experiences: Evidence from African football,? American Economic Review,
2020, 110 (5), 1572?1602.
der Berg, Servaas Van, ?Apartheid?s enduring legacy: Inequalities in education,? Journal of
African economies, 2007, 16 (5), 849?880.
Dinkelman, Taryn, ?Long-run Health Repercussions of Drought Shocks: Evidence from South
African Homelands,? The Economic Journal, 2017, 127 (604), 1906?1939.
DOE, Department of Education, Revised National Curriculum Statements Grades R-9., Gov-
ernment Publisher, 2002.
, Revised National Curriculum Statements Grades R-9. Teachers Guide for the Development
of Learning Programmes,Foundation Phase, Government Publisher, 2003.
Drewett, Michael, ?An analysis of the censorship of popular music within the context of cultural
struggle in South Africa during the 1980s.? PhD dissertation, Rhodes University 2004.
Eifert, Benn, Edward Miguel, and Daniel N Posner, ?Political competition and ethnic identi-
fication in Africa,? American Journal of Political Science, 2010, 54 (2), 494?510.
Ericson, Steven J, ?The? Matsukata Deflation? Reconsidered: Financial Stabilization and
Japanese Exports in a Global Depression, 1881?85,? The Journal of Japanese Studies, 2014,
pp. 1?28.
Falch, Torberg, ?The Elasticity of Labor Supply at the Establishment Level,? Journal of Labor
Economics, 2010, 28 (2), 237?266.
164
FCC, Federal Communications Commissions, ?FCC-0416: Technical Standards for Determin-
ing EligibilityFor Satellite-Delivered Network Signals PursuantTo the Satellite Home Viewer
Improvement Act,? Technical Report 2000.
Ferrara, Eliana La, Alberto Chong, and Suzanne Duryea, ?Soap operas and fertility: Evi-
dence from Brazil,? American Economic Journal: Applied Economics, 2012, 4 (4), 1?31.
Fouka, Vasiliki, ?Backlash: The unintended effects of language prohibition in US schools after
World War I,? The Review of Economic Studies, 2020, 87 (1), 204?239.
Gagliarducci, Stefano, Massimiliano Gaetano Onorato, Francesco Sobbrio, and Guido
Tabellini, ?War of the waves: radio and resistance during World War II,? 2018.
Galiani, Sebastian and Martin Gonzalez-Rozada, ?Inference and Estimation in Small Sample
Dynamic Panel Data Models,? Working Paper, 2005.
Goodman-Bacon, Andrew, ?Difference-in-differences with variation in treatment timing,?
Technical Report 2019.
Gort, Michael and Steven Klepper, ?Time paths in the diffusion of product innovations,? The
Economic Journal, 1982, 92 (367), 630?653.
Green, Donald P and Srinivasan Vasudevan, ?Diminishing the effectiveness of vote buying:
Experimental evidence from a persuasive radio campaign in India,? Unpublished Manuscript,
Columbia University, 2016.
Green, Elliott, ?Ethnicity, national identity and the state: evidence from sub-Saharan Africa,?
British Journal of Political Science, 2018, pp. 1?23.
Gunner, Liz, ?Resistant medium: The voices of Zulu radio drama in the 1970s,? Theatre Re-
search International, 2002, 27 (3), 259?274.
Hahn, Jinyong and Guido Kuersteiner, ?Asymptotically Unbiased Inference for a Dynamic
Panel Model with Fixed Effects When Both n and T Are Large,? Econometrica, 2002, 70 (4),
1639?1657.
Hamm, Charles, ?Rock?n?roll in a very strange society,? Popular music, 1985, 5, 159?174.
, ?The constant companion of man: Separate Development, Radio Bantu and music,? Popular
music, 1991, 10 (2), 147?173.
Harris, John R and Michael P Todaro, ?Migration, unemployment and development: a two-
sector analysis,? The American Economic Review, 1970, 60 (1), 126?142.
Hayman, Graham and Ruth Tomaselli, ?Ideology and technology in the growth of South
African broadcasting, 1924-1971,? Currents of power: State broadcasting in South Africa.
Bellville, South Africa: Anthropos, 1989.
Heath, Rachel, ?Why do firms hire using referrals? Evidence from Bangladeshi garment facto-
ries,? Journal of Political Economy, 2018, 126 (4), 1691?1746.
165
Hirsch, Boris, Thorsten Schank, and Claus Schnabel, ?Differences in Labor Supply to
Monopsonistic Firms and the Gender Pay Gap: An Empirical Analysis Using Linked
Employer-Employee Data from Germany,? Journal of Labor Economics, 2010, 28 (2), 291?
330.
ICASA, ?Nortice of Publication of Terrestrial Broadcast Frequency,? Government Gazette, 2002,
23695.
, ?Nortice of Publication of Terrestrial Broadcast Frequency,? Government Gazette, 2008,
32728.
, ?Notice in terms of Section 34 of the South African Electronic Act, No. 36 of 2005,? Gov-
ernment Gazzette, 2008.
ICCJ, Information and Communication Council of Japan, ?Hoso system Iinkai Hokoku (in
English: Reports on broadcasting system),? 2007.
IEC, Election Commission of South Africa, ?National and Provincial election re-
sults,? https://www.elections.org.za/electionresults/Downloads/
NPE-Results 1994?2019.
ITU, International Telecommunication Union, ?Minimum field strengths for which protection
may be soughtin planning an analogue terrestrial television service,? Rec. ITU-R, 2002, 417-5.
Ivlevs, Artjoms and Roswitha M King, ?2004 Minority Education Reform and pupil perfor-
mance in Latvia,? Economics of Education Review, 2014, 38, 151?166.
Jensen, Robert and Emily Oster, ?The power of TV: Cable television and women?s status in
India,? The Quarterly Journal of Economics, 2009, 124 (3), 1057?1094.
Johnston, Craig James, Music Power: The Story of Capital Radio 604, Independent documen-
tary, 2019.
Kadt, Daniel De and Horacio A Larreguy, ?Agents of the regime? Traditional leaders and
electoral politics in South Africa,? The Journal of Politics, 2018, 80 (2), 382?399.
Kaplan, Irving, James L. McLaughlin, Barbara J. Marvin, Harond D. Nelson, Ernestine E.
Rowland, and Donald P. Whitakar, Area Handbook for the Republic of South Africa, US
Government Printing Office, 1971.
Kapoor, Mudit, Harry H. Kelejian, and Ingmar R. Prucha, ?Panel Data Models with Spatially
Correlated Error Components,? Journal of Econometrics, 2007, 140 (1), 97?130.
Kearney, Melissa S and Phillip B Levine, ?Media influences on social outcomes: The impact
of MTV?s 16 and pregnant on teen childbearing,? American Economic Review, 2015, 105 (12),
3597?3632.
and , ?Early childhood education by television: Lessons from Sesame Street,? American
Economic Journal: Applied Economics, 2019, 11 (1), 318?50.
166
Kinugawa, Taichi, Honpo Menshi Bousekishi (History of Japanese Cotton Spinning Industry),
Nihon Mengyo Club, 1937.
Kiviet, Jan F., ?On bias, inconsistency, and efficiency of various estimators in dynamic panel
data models,? Journal of Econometrics, 1995, 68 (1), 53?78.
Kymlicka, Will, Multicultural citizenship: A liberal theory of minority rights, clarendon Press,
1995.
Lai, Calvin K, Maddalena Marini, Steven A Lehr, Carlo Cerruti, Jiyun-Elizabeth L Shin,
Jennifer A Joy-Gaba, Arnold K Ho, Bethany A Teachman, Sean P Wojcik, Spassena P
Koleva et al., ?Reducing implicit racial preferences: I. A comparative investigation of 17
interventions.,? Journal of Experimental Psychology: General, 2014, 143 (4), 1765.
Lamadon, Thibaut, Magne Mogstad, and Bradley Setzler, ?Imperfect Competition, Compen-
sating Differentials, and Rent Sharing in the US Labor Market,? American Economic Review,
2022, 112 (1), 169?212.
Landes, David S., The Wealth and Poverty of Nations: Why Some Are So Rich and Some So
Poor, W W Norton and Co Inc, 1998.
Lekgoathi, Sekibakiba Peter, ?You are Listening to Radio Lebowa of the South African Broad-
casting Corporation: Vernacular Radio, Bantustan Identity and Listenership, 1960?1994,?
Journal of Southern African Studies, 2009, 35 (3), 575?594.
, ?The African National Congress?s Radio Freedom and its audiences in apartheid South Africa,
19631991,? Journal of African Media Studies, 2010, 2 (2), 139?153.
, ?Ethnic separatism or cultural preservation? Ndebele radio under apartheid, 1983?1994,?
South African Historical Journal, 2012, 64 (1), 59?80.
Levinsohn, James, ?Globalization and the Returns to Speaking Englishin South Africa,? in
?Globalization and Poverty,? University of Chicago Press, 2007, pp. 629?646.
Lewis, W Arthur, ?Economic Development with Unlimited Supplies of Labour,? The Manch-
ester School, 1954, 22 (2), 139?191.
Manning, Alan, Monopsony in Motion : Imperfect Competition in Labor Markets, Princeton;
NJ: Princeton University Press, 2003.
Manski, Charles F., ?Identification of Endogenous Social Effects: The Reflection Problem,? The
Review of Economic Studies, 1993, 60 (3), 531.
Marshall, Alfred, Principles of Economics, 8th ed. ed., Macmillan, 1920.
Matsudaira, Jordan D., ?Monopsony in the Low-Wage Labor Market? Evidence from Mini-
mum Nurse Staffing Regulations,? Review of Economics and Statistics, 2014, 96 (1), 92?102.
Montgomery, James D, ?Social networks and labor-market outcomes: Toward an economic
analysis,? The American Economic Review, 1991, 81 (5), 1408?1418.
167
Mousa, Salma, ?Building Tolerance: Intergroup Contact and Soccer in Post-ISIS Iraq,? 2019.
Msila, Vuyisile, ?From apartheid education to the Revised National Curriculum Statement: Ped-
agogy for identity formation and nation building in South Africa,? Nordic Journal of African
Studies, 2007, 16 (2).
Naidu, Suresh, Yaw Nyarko, and Shing-Yi Wang, ?Monopsony power in migrant labor mar-
kets: evidence from the United Arab Emirates,? Journal of Political Economy, 2016, 124 (6),
1735?1792.
Ndlovu, Andile, ?The Wild nominated for seven awards,? 2012.
NEEDU, ?National Report 2012 The State of Literacy Teaching and Learning in the Founding
Phase,? 2013.
NEPI, National Education Policy INvestigation (Research Group), Language : Report of the
NEPI Language Research Group 1992.
News24, ?DStv ropes in 1m SA subscribers,? https://www.news24.com/fin24/
dstv-ropes-in-1m-sa-subscribers-20051222 2005.
Nickell, Stephen, ?Biases in Dynamic Models with Fixed Effects,? Econometrica, 1981, 49 (6),
1417?1426.
Ohkawa, Kazushi, Miyohei Shinohara, and Mataji Umemura, Long-Term Economic Statis-
tics (LTES), Tokyo: Toyo Keizai Shinposha, Inc., 1967.
Okamoto, Yukio, Meijiki Boseki Rodo Kankeishi : Nihon-teki Koyo, Roshi Kankei Keisei heno
Sekkin (History of Labor Relations in Cotton Spinning Firms in Meiji), Kyushu University
Press, 1993.
Okuyama, Yoko, ?Toward Better Informed Decision-Making: The Impacts of a Mass Media
Campaign on Womens Outcomes in Occupied Japan,? 2019.
Olken, Benjamin A, ?Do television and radio destroy social capital? Evidence from Indonesian
villages,? American Economic Journal: Applied Economics, 2009, 1 (4), 1?33.
O?meara, Dan, Forty lost years: The apartheid state and the politics of the National Party,
1948-1994, Ravan Pressof South Africa, 1996.
Paluck, Elizabeth Levy and Donald P Green, ?Prejudice reduction: What works? A review
and assessment of research and practice,? Annual review of psychology, 2009, 60, 339?367.
Perry, Guillermo, Poverty reduction and growth: virtuous and vicious circles, World Bank Pub-
lications, 2006.
Piper, Laurence, ?The Inkatha Freedom Party: between the impossible and the ineffective,? in
?Electoral politics in South Africa,? Springer, 2005, pp. 148?165.
168
Plaut, Martin, ?South Africa: how the ANC wins elections,? Review of African Political Econ-
omy, 2014, 41 (142), 634?644.
Plessis, Elize Du and Petro Marais, ?Reflections on the NCS to NCS (CAPS): Foundation
Phase teachers? experiences,? The Independent Journal of Teaching and Learning, 2015, 10
(1), 114?126.
Posel, Deborah, ?The meaning of apartheid before 1948: Conflicting interests and forces within
the Afrikaner Nationalist alliance,? Journal of Southern African Studies, 1987, 14 (1), 123?
139.
Ransom, Michael R and David P. Sims, ?Estimating the Firm?s Labor Supply Curve in a New
Monopsony Framework: Schoolteachers in Missouri,? Journal of Labor Economics, 2010, 28
(2), 331?355.
and Ronald L. Oaxaca, ?New Market Power Models and Sex Differences in Pay,? Journal
of Labor Economics, 2010, 28 (2), 267?289.
Reed, Holly E, ?Moving across boundaries: migration in South Africa, 1950?2000,? Demogra-
phy, 2013, 50 (1), 71?95.
Rinz, Kevin, ?Labor market concentration, earnings, and inequality,? Journal of Human Re-
sources, 2022, 57 (S), S251?S283.
Robinson, Amanda Lea, ?National versus ethnic identification in Africa: Modernization, colo-
nial legacy, and the origins of territorial nationalism,? World Politics, 2014, 66 (4), 709?746.
Rogers, Barbara, Divide & Rule: South Africa?s Bantustans, International defence & aid Fund,
1980.
Russo, Gianluca, ?Media and Assimilation: Evidence from the Golden Age of Radio,? Working
Paper, 2021.
Saxonhouse, Gary, ?A Tale of Japanese Technological Diffusion in the Meiji Period,? The Jour-
nal of Economic History, 1974, 34 (1), 149?165.
, ?Country Girls and Communication among Competitors in the Japanese Cotton-Spinning
Industry,? in H. Patrick, ed., Japanese Industrialization and Its Social Consequences, Berkeley:
University of California Press, 1976.
Saxonhouse, Gary R. and Yukihiko Kiyokawa, ?Supply and Demand for Quality Workers in
Cotton Spinning in Japan and India,? in Kazuchi Ohkawa and Gustav Ranis, eds., Japan and
the Developing Countries, Oxford: Basil Blackwell, 1985, pp. 177?211.
Setati, Mamokgethi, Jill Adler, Yvonne Reed, and Abdool Bapoo, ?Incomplete journeys:
Code-switching and other language practices in mathematics, science and English language
classrooms in South Africa,? Language and education, 2002, 16 (2), 128?149.
Smith, Adam, The Wealth of Nations, Londonm: W. Strahan and T. Cadell, 1776.
169
So?derbom, Ma?ns, Francis Teal, and Anthony Wambugu, ?Unobserved heterogeneity and the
relation between earnings and firm size: evidence from two developing countries,? Economics
letters, 2005, 87 (2), 153?159.
Staiger, Douglas O., Joanne Spetz, and Ciaran S. Phibbs, ?Is There Monopsony in the Labor
Market? Evidence from a Natural Experiment,? Journal of Labor Economics, 2010, 28 (2),
211?236.
Strobl, Eric, Robert Thornton et al., ?Do large employers pay more? The case of five develop-
ing African countries,? Journal of Economic Development, 2004, 29, 137?162.
Stro?mberg, David, ?Radio?s impact on public spending,? The Quarterly Journal of Economics,
2004, 119 (1), 189?221.
Sullivan, Daniel, ?Monopsony Power in the Market for Nurses,? The Journal of Law and Eco-
nomics, 1989, 32 (2, Part 2), 135?178.
Sun, Liyang and Sarah Abraham, ?Estimating dynamic treatment effects in event studies with
heterogeneous treatment effects,? Journal of Econometrics, 2020.
Takamura, Naosuke, Introduction to Japanese Cotton Spinning Industry (Nohon Boseki Gyoshi
Josetsu), Hanawa Shobo, 1971.
Tanaka, Hiroshi, Kanebo and Its Group: Japanese Cotton Spinning Industry (Kanebo to
Keiretsu: Nihon no Boseki), Nihon Konzern Kankokai, 1957.
Taylor, N and J Moyane, ?Khanyisa Education Support Programme: Baseline study part 1:
communities, schools and classrooms,? Johannesburg: JET Education Services, 2005.
Taylor, Nick and Penny Vinjevold, Getting Learning Right. Report of the President Education-
Initiative Research Project, The Joint Education Trust, 1999.
Taylor, Stephen and Marisa von Fintel, ?Estimating the impact of language of instruction in
South African primary schools: A fixed effects approach,? Economics of Education Review,
2016, 50, 75?89.
Tella, Rafael Di, Sebastian Galiani, and Ernesto Schargrodsky, ?Persuasive propaganda dur-
ing the 2015 Argentine ballotage,? Journal of Comparative Economics, 2021, forthcoming.
Todd, Andrew R, Galen V Bodenhausen, Jennifer A Richeson, and Adam D Galinsky, ?Per-
spective taking combats automatic expressions of racial bias.,? Journal of personality and so-
cial psychology, 2011, 100 (6), 1027.
Trainotti, Valentine, ?Simplified Calculation of Coverage Area for MF AM Broadcast Stations,?
IEEE Antenas and Propagation Society Magazine, 1990, pp. 41?44.
Tsurumi, Patricia E., Factory Girls: Women in the Thread Mills of Meiji Japan, Princeton:
Princeton University Press, 1990.
Wang, Tianyi, ?Media, Pulpit, and Populist Persuasion: Evidence from Father Coughlin,? 2020.
170
Webb, Vic, Michel Lafon, and Phillip Pare, ?Bantu languages in education in South Africa:
An overview. Ongekho akekho!?The absentee owner,? Language Learning Journal, 2010, 38
(3), 273?292.
Wilse-Samson, Laurence H, ?Structural change and democratization: Evidence from rural
apartheid,? 2013.
Wooldridge, Jeffrey M, Econometric analysis of cross section and panel data, MIT press, 2010.
World Bank, ?LOUD AND CLEAR: Effective Language of InstructionPolicies For Learning,?
A World Bank Policy Approach Paper, 2021.
Yanagizawa-Drott, David, ?Propaganda and conflict: Evidence from the Rwandan genocide,?
The Quarterly Journal of Economics, 2014, 129 (4), 1947?1994.
171