University of Southern California Dissertations and Theses

Three essays on the microeconometric analysis of the labor market

(USC Thesis Other)

Three essays on the microeconometric analysis of the labor market

PDF

Download a page range

Download transcript

Copy asset link

Request this asset

Transcript (if available)

Content THREE ESSAYS ON THE MICROECONOMETRIC ANALYSIS OF THE LABOR MARKET
by
Jung Hyuk Lee
A Dissertation Presented to the
FACULTY OF THE USC GRADUATE SCHOOL
UNIVERSITY OF SOUTHERN CALIFORNIA
In Partial Fulfillment of the
Requirements for the Degree
DOCTOR OF PHILOSOPHY
(ECONOMICS)
May 2023
Copyright 2023 Jung Hyuk Lee
Dedication
I dedicate this thesis to the giants of human knowledge who generously allowed me to climb up their
shoulders. The honorable inventor of the dishwasher is one of them.
ii
Acknowledgements
Director Bong Joon-ho, a film director best known for his movie parasite, said, “The most personal is the
most creative” at the Academy Award 2020. The three research papers presented in this dissertation came
from the most personal experiences of mine.
I started my career in the Korean Ministry of Economy and Finance in 2011 after passing one of the most
competitive exams in Korea. However, I did not even know how much Korean public servants make every
year before I received my first paycheck. Like my past self, most college students are too busy to gather
comprehensive information about their potential job positions before making too much of irreversible
investments in preparation for a specific job or a sector. In this environment, prior beliefs usually have
the upper hand over new information that the job seekers gather about the job or sector characteristics.
My first chapter, “Information frictions in job search and occupational segregation by gender,” started
from this observation and tries to explain that one of the most persistent and pervasive phenomena in the
modern labor market – occupational segregation by gender – is largely driven by information frictions. It
challenges the conventional view that the gaps in preferences and skills entirely explain the differential
labor supply behaviors by gender.
During my Ph.D. program, Korean society has experienced an unprecedented social change. Femi-
nism, which had primarily been reserved for the educated few in the conservative Korean society before,
suddenly came to the center of the social discourse and began to change people’s views. It not only ex-
erted substantial influences on the political or public arena but also made people fundamentally rethink
iii
what should be the right role of both genders in their households. To me, the most challenging part of
the Ph.D. life was always taking proper care of endless housework and securing a reasonable amount of
time for research. Since exposure to feminist ideas through the media had such a substantial impact on the
bargaining process between my wife and me, I began to think about the impact on a more general scale,
ending up in the second chapter, “Changing gender norms and occupational segregation.”
Studying under the guidance of a prominent development economist, I faced one of the most funda-
mental problems in low-income countries’ industries: Why firms are still small and not scalable? Is it all
because of external barriers, such as lack of capital, social infrastructure, or human resources, as studied
by other researchers? Having witnessed the most exceptionally successful example of economic develop-
ment of Korea as a public servant at the heart of the government, I had a personal conviction that the real
culprit was inside. My co-authors and I zoomed in on understudied internal barriers and showed the link
between lack of specialization and scalability. And this became my third chapter, “Self-employment within
the firm.”
On the other hand, sparkling ideas from personal experiences were never a sufficient condition for
successful academic research. To me, what was particularly intriguing about the Ph.D. course, and what
really differentiated it from lower-level education, was that the knowledge was transmitted not by a set of
systematic and unilateral lectures but by some loose form of a collective apprenticeship from a network of
scholars. The system was operated as: (i) each student starts to present his/her rudimentary ideas to the
faculty→ (ii) scholars in diverse fields give the student different pieces of advice → (iii) an unexpected idea
emerges from the chaos that consists of the student’s prior knowledge and experiences, a fractured mixture
of economic theories, and the combinations of the faculty’s research fields → (iv) the student develops the
idea to a thesis by overcoming incessant criticisms. Notably, this process necessitates voluntary and well-
intended participations of numerous researchers, who usually do not have any type of personal interest in
iv
the student’s ultimate outcomes. I believe this is the true beauty of modern Ph.D. education (and more in
general, academia) that I have experienced in the United States.
Likewise, my dissertation is anything but the materialization of ideas that came exclusively from my
brain. During the journey, I have received countless pieces of advice and feedback from my teachers. And
I believe that what they only want from the investments of their valuable resources is that I also contribute
to society by giving back. My most memorable teachers (but not exclusively) are Vittorio Bassi, Jefferey
Weaver, Nicolas J. Duquette, Jeff Nugent, Daniel Bennett, Simon Quach, Augustin Bergeron, Tommaso
Porzio, Alessandra Peter, Ritwika Sen, Laura Schechter, TJ McCarthy, Michael Magill, Hayun Song, Hye-
won Kim, and Eunjee Kwon. In addition, I thank the scholars who wrote more than 100 papers I cited in
this dissertation for letting me climb up their shoulders.
Finally, I acknowledge that I have been the luckiest Ph.D. student who received endless support and
love from my family for five years. I send my love to my parents, wife, and adorable two daughters, Luna
and Hannah.
v
TableofContents
Dedication . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ii
Acknowledgements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iii
List of Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ix
List of Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xii
Abstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xiv
Chapter 1: Information Frictions in Job Search and Occupational Segregation by Gender . . . . . . 1
1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Institutional Context . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.3 Experimental Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.3.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.3.1.1 Superiority of MDS in Wages and Nonwage Amenities . . . . . . . . . . 7
1.3.1.2 Conceptual Framework . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.3.1.3 Key Challenges in Identification . . . . . . . . . . . . . . . . . . . . . . . 11
1.3.2 Sampling and Group Assignment at Baseline . . . . . . . . . . . . . . . . . . . . . 12
1.3.3 Information Treatment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
1.3.4 Follow-up Surveys and Outcome Variables . . . . . . . . . . . . . . . . . . . . . . . 16
1.3.5 Checks on Balance and Attrition . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
1.4 Empirical Specifications and Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
1.4.1 Empirical Specifications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
1.4.2 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
1.4.2.1 Impacts on Beliefs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
1.4.2.2 Impacts on Aspirations . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
1.4.2.3 Impacts on Sectoral Allocation of Job Search and Applications . . . . . . 25
1.4.2.4 Impacts on Labor Market Engagement and Employment Outcomes . . . 27
1.4.2.5 Checks on Other Components of the Conceptual Framework . . . . . . . 29
1.5 A Discussion on the Mechanism . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
1.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
Chapter 2: Changing Gender Norms and Household Resource Allocation . . . . . . . . . . . . . . . 34
2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
2.2 Background and Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
2.2.1 Korean Newspapers’ Coverage of Feminism . . . . . . . . . . . . . . . . . . . . . . 38
vi
2.2.2 Women’s Perceptions of Gender Norms . . . . . . . . . . . . . . . . . . . . . . . . 40
2.2.3 Household Resource Allocation and Other Outcomes variables . . . . . . . . . . . 42
2.2.4 Sample and Attrition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
2.3 The Impact of Feminism-related Articles on Perceptions . . . . . . . . . . . . . . . . . . . . 43
2.3.1 Exposure to Feminism-related Articles . . . . . . . . . . . . . . . . . . . . . . . . . 44
2.3.2 Empirical Strategy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
2.3.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
2.4 The Impact of the Changes in Perceptions on Resource Allocation . . . . . . . . . . . . . . 50
2.4.1 Previous Literature . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
2.4.2 Descriptive Statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
2.4.3 Empirical Strategy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
2.4.4 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
2.5 Robustness Checks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
2.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
Chapter 3: Self-employment within the Firm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
3.2 The Survey . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
3.2.1 Sampling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
3.2.2 Main Survey . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
3.2.3 Follow-up Survey . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
3.3 The Organization of Production Inside the Firm . . . . . . . . . . . . . . . . . . . . . . . . 76
3.3.1 Firm characteristics, and markets for output and labor . . . . . . . . . . . . . . . . 77
3.3.2 Task Composition: What Do Firms Do? . . . . . . . . . . . . . . . . . . . . . . . . 80
3.3.3 Task Allocation: Who Does What Within the Firm? . . . . . . . . . . . . . . . . . . 84
3.3.4 Sectoral Heterogeneity and Barriers to Specialization . . . . . . . . . . . . . . . . . 91
3.4 Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
3.4.1 Environment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
3.4.2 Choices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
3.4.3 Equilibrium . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
3.4.4 Characterization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
3.5 Bringing the Model to the Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110
3.5.1 Empirical Validation of the Theoretical Predictions . . . . . . . . . . . . . . . . . . 110
3.5.2 Model’s Extension and Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . 112
3.5.3 Targeted Moments and Identification . . . . . . . . . . . . . . . . . . . . . . . . . . 113
3.5.4 Estimation Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
3.6 Quantifying within Firm Organization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114
3.6.1 Close to Self-Employment Within the Firm . . . . . . . . . . . . . . . . . . . . . . 114
3.6.2 Returns from Other Interventions Are Muted . . . . . . . . . . . . . . . . . . . . . 116
3.7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120
Appendices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130
A Appendix to Chapter 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130
A.1 Appendix Tables and Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130
A.2 An Experiment of “Tailored Job Information Newsletter” . . . . . . . . . . . . . . . 138
B Appendix to Chapter 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140
vii
B.1 Appendix Tables and Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140
B.2 The Differences in the Newspapers’ Framing of Feminism . . . . . . . . . . . . . . 150
C Appendix to Chapter 3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 154
C.1 Appendix Tables and Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 154
C.2 Attrition from the Follow-up Surveys . . . . . . . . . . . . . . . . . . . . . . . . . . 165
viii
ListofTables
1.1 Construction of Sector Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
1.2 Balance at Baseline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
1.3 Treatment Effects on Aspirations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
1.4 Treatment Effects on Sector Choices ( 2
nd
follow-up) . . . . . . . . . . . . . . . . . . . . . . 26
1.5 Treatment effects on Labor Market Engagement/Employment Outcomes ( 2
nd
follow-up) . 28
1.6 Treatment Effects on Other Components ( 1
st
and2
nd
follow-up) . . . . . . . . . . . . . . . 30
2.1 Components of the GNP Score . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
2.2 Construction of Outcome Variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
2.3 The First-stage Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
2.4 Descriptive Statistics (Averages) at Baseline (Wave 5) . . . . . . . . . . . . . . . . . . . . . 52
2.5 2SLS Estimates on Time Use in Household Labor . . . . . . . . . . . . . . . . . . . . . . . . 55
2.6 The 2SLS Estimates on Other Outcome Variables . . . . . . . . . . . . . . . . . . . . . . . . 58
3.1 Measuring Time Use . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
3.2 Descriptives on Firm Characteristics and Output Market . . . . . . . . . . . . . . . . . . . 80
3.3 Heterogeneity in Labor Specialization by Sector . . . . . . . . . . . . . . . . . . . . . . . . 91
3.4 Descriptives on Customization and Customer Interactions by Sector . . . . . . . . . . . . . 92
3.5 Cross-Sectoral Hetoreogeneity to Test Proposition 1 . . . . . . . . . . . . . . . . . . . . . . 111
3.6 Returns to Managerial Ability in Locations with Different Firm Size . . . . . . . . . . . . . 112
ix
A.1 Employment Shares and Gender Composition of the Top 20 Sectors in GOMS . . . . . . . 130
A.2 Construction of Main Outcome Variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
A.3 Attrition and Balance in the Follow-up Surveys . . . . . . . . . . . . . . . . . . . . . . . . 132
A.4 Parameters vs. Students’ Beliefs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133
A.5 Analyses of Heterogeneity in Treatment Effects on the Main Outcome variables . . . . . . 134
A.6 Reasons for Not Searching in Specific Sectors ( 2
nd
follow-up, aggregated) . . . . . . . . . . 135
A.7 Treatment Effects in the Newsletter Experiment ( 1
st
follow-up) . . . . . . . . . . . . . . . 139
B.1 Market Shares of the Big Six Newspapers (%) . . . . . . . . . . . . . . . . . . . . . . . . . . 140
B.2 Survey Questions Used to Measure Perceptions of Gender Norms in the Literature . . . . . 141
B.3 Pairwise Correlation Coefficients between the Components of the GNP Score . . . . . . . 142
B.4 Sample and Attrition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143
B.5 2SLS Estimates on Time Use in Housework/Childcare . . . . . . . . . . . . . . . . . . . . . 144
B.7 Robustness – The Results with and without Region Fixed Effects (First Stage) . . . . . . . . 145
B.8 Robustness – The results with and without region fixed effects (second stage) . . . . . . . 145
B.6 The OLS Estimates on Other Outcome Variables . . . . . . . . . . . . . . . . . . . . . . . . 146
B.9 Robustness - Replications of the Main Results without Gangwon and Jeju Regions . . . . . 147
B.10 Robustness - Relationship between Bartik exposure and Observable Correlates . . . . . . . 148
B.11 Robustness - Likelihood of Divorce/Separation (2SLS estimates) . . . . . . . . . . . . . . . 149
B.12 Subjects of Feminism-related Articles in 2018 . . . . . . . . . . . . . . . . . . . . . . . . . . 151
C.1 Measuring Time Use: Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 154
C.2 Comparison with IKEA Prices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155
C.3 Heterogeneity in Skill Intensity of Tasks . . . . . . . . . . . . . . . . . . . . . . . . . . . . 156
C.4 Heterogeneity in Skill Distribution within the Firm . . . . . . . . . . . . . . . . . . . . . . 157
C.5 Attrition Table – Owners . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 165
x
C.6 Attrition Table – Employees . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 166
xi
ListofFigures
1.1 Proportion of Women vs. Average Sector Characteristics for Female Workers . . . . . . . . 7
1.2 Conceptual Framework of the Experiment . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
1.3 Table Provided in the Information Treatment (translated to English) . . . . . . . . . . . . . 15
1.4 Belief Updating . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
2.1 Newspaper coverage of feminism and women’s perceptions of gender norms . . . . . . . . 39
2.2 The Relationship between the Changes in Exposure and the GNP Scores . . . . . . . . . . . 45
2.3 Estimated Effects of Bartik exposure on Components of the GNP Score . . . . . . . . . . . . 49
2.4 Trends in the (Standardized) Average GNP Scores across Regions . . . . . . . . . . . . . . 60
3.1 Firm Size Distribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
3.2 Task Composition across the Size Distribution . . . . . . . . . . . . . . . . . . . . . . . . . 81
3.3 Task allocation between Production and Non-production Tasks . . . . . . . . . . . . . . . 85
3.4 Task Allocation between Production and Non-production by Firm Size . . . . . . . . . . . 86
3.5 Task Allocation within Production across the Size Distribution . . . . . . . . . . . . . . . . 89
3.6 Within-Firm Organization and Unbundling Costκ 0
. . . . . . . . . . . . . . . . . . . . . . 109
3.7 Labor Specialization, Firm Size, and Unbundling Costκ 0
. . . . . . . . . . . . . . . . . . . 109
3.8 Model Fit for Within-firm Allocation of Time to Tasks . . . . . . . . . . . . . . . . . . . . . 114
3.9 Aggregate Effect of Changing the Unbundling Cost κ 0
. . . . . . . . . . . . . . . . . . . . 115
3.10 Returns from a Reduction in the Revenue Wedgeτ . . . . . . . . . . . . . . . . . . . . . . . 117
xii
A.1 Occupational Segregation by Gender among College Graduates (2-digit sectors) . . . . . . 136
A.2 Histogram of Female Graduates’ Sector Distributions by Stated Preferences . . . . . . . . . 137
B.1 Sections Where Feminism-focusing Articles Were Located . . . . . . . . . . . . . . . . . . 152
C.1 Matching in the Labor Market . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 158
C.2 Time Allocation within Production across Production Steps . . . . . . . . . . . . . . . . . . 159
C.3 Time Allocation within Idle Time . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 160
C.4 Task Allocation between Production and Non-production Tasks by Employee Skills . . . . 161
C.5 Employee Contribution to Demand Generation by Firm Size . . . . . . . . . . . . . . . . . 162
C.6 Heterogeneity in Share of Firms Performing a Production Step by Sector . . . . . . . . . . 163
C.7 Time Allocation to Non-production tasks (Heterogeneity by Sector) . . . . . . . . . . . . . 163
C.8 Heterogeneity in Allocation within Production between Steps . . . . . . . . . . . . . . . . 164
xiii
Abstract
In this dissertation, I adopt advanced microeconometric tools to analyze three important topics in modern
labor economics. In the first chapter, I discuss the role of incomplete information during a job search in
occupational segregation by gender. I administered a randomized experiment with female Korean under-
graduates and provided half of them with detailed statistics for representative male- and female-dominated
industrial sectors. The treatment group reported substantially less biased beliefs on sector characteristics
and showed relatively higher aspirations to enter jobs in male-dominated sectors than the control group.
In the second chapter, I examine what happened to household resource allocation when gender norms
began to change. By collecting newspaper articles about feminism in the past ten years in Korea and ex-
ploiting their region-year variations, I provided evidence that increased newspaper coverage of feminism
contributed to egalitarian attitudes among women. Then, I constructed a shift-share instrumental variable
with the growth of the feminism-related articles to show that the change in women’s perceptions of gender
norms induced by those articles affected both spouses’ time use in household labor and women’s welfare.
Finally, in the third chapter, I zoom in on the internal organization of manufacturing firms in Uganda to
analyze the relationship between limited specialization and scalability. Based on detailed time-use data,
I estimated the structural model of labor specialization within firms and demonstrated that the internal
barrier to specialization – delegation cost – was the key to understanding low returns to scale, resulting
in small firm size in developing countries.
xiv
Chapter1
InformationFrictionsinJobSearchandOccupationalSegregationby
Gender
1.1 Introduction
Occupational segregation has maintained a central role in explaining gender wage gaps. According to Blau
and Kahn (2017), gender differences in industry and occupation account for about half of the gender wage
gaps found in the U.S. Evidence on the prominence of occupational segregation in explaining gender wage
gaps has accumulated with respect to diverse cultural contexts, including Europe (Dolado, Felgueroso,
and Jimeno, 2002), Asia (Meng and Miller, 1995; Ismail, Farhadi, and Wye, 2017), and other developing
countries (World Bank, 2011; Borrowman and Klasen, 2020).
There is a thick literature on what causes this pervasive gender segregation. As summarized in an
extensive literature review by Cortes and Pan (2018), many analyses from the labor supply side focus on
how men and women differ in their preferences over job characteristics. For instance, female job seekers
tend to impose greater weights on certain nonwage aspects of jobs, including job security (Petrongolo,
2004), flexibility (Mas and Pallais, 2017), or commuting distance (Le Barbanchon, Rathelot, and Roulet,
2021). However, this classical approach, which implicitly assumes an occupational choice under complete
information, does not fully explain the fact that male-dominated sectors typically provide not only higher
1
wages but also better nonwage amenities to their female employees. We begin our analysis by reporting
evidence on this superiority of male-dominated sectors in Korea, which suggests that female job seekers
might be imperfectly informed about male-dominated sectors.
Drawing on this observation, this study examines the implications of information frictions for women’s
occupational choice, focusing on the transition from college to the labor force. By administering a sur-
vey experiment with female college students in Korea, we measured the impact of randomly providing
wage and nonwage information about representative male-dominated industrial sectors (MDS) and female-
dominated sectors (FDS) to the students in the treatment group when they were beginning their job search
in school. We then traced the trajectories of their beliefs, aspirations, and job search/application behaviors
for four months to investigate how differently they formulated their labor supply decisions on MDS or
FDS in comparison to the control group without intervention.
More specifically, 1,000 female senior college students with humanities and social sciences majors
were recruited through a professional survey firm. We administered our baseline survey to and received
responses from all of them in October 2021, which corresponded to the beginning of the last semester
in college for most of the senior students. The key feature of the baseline survey was that it included
an information treatment for 500 students in the treatment group (Group T) but not for the other 500
students in the control group (Group C). Using a nationally representative survey of college graduates in
Korea (Graduates Occupational Mobility Survey), we computed four average statistics on characteristics
– wages, work hours, welfare institutions index, and job security index – that female college graduates
reported about their first jobs. Then, we provided these numbers to Group T students in table format and
focused their attention on the seven sectors that are the largest employers – three male dominated, three
female dominated, and one neutral in terms of gender composition.
We administered two follow-up surveys to trace the treatment effects. The first one was administered
one month after the baseline (November 2021), with a specific focus on the students’ beliefs on sector
2
characteristics and the job search process. The second follow-up survey was implemented in February
2022, when many of the sample students had graduated from college and were actively engaging in the
job application process. In this round, students’ retrospective responses about the sectors that they had
searched and applied for, as well as details on their current labor market status, were collected.
We found substantial treatment effects on beliefs. Group C students showed a large and systematic
bias in their beliefs about job characteristics such that they underestimated the gaps between MDS and
FDS. In other words, while MDS jobs where recent female graduates were employed were much more
favorable in terms of wages, welfare institutions, and job security, control students did not have beliefs
corresponding to these realities and perceived MDS and FDS to be more similar than they actually are. We
found substantial treatment effects on Group T students, with their estimation of the magnitudes of the
gaps between MDS and FDS jobs being larger and closer to the actual magnitudes on all four characteristics.
We also found statistically significant effects on the stated aspirations to work in the seven experimen-
tal sectors. We computed (standardized) relative aspirations to enter the three MDS over the three FDS, and
the students in Group T showed higher relative aspirations to enter MDS by 0.2 standard deviations. The
magnitude of this treatment effect remained stable throughout the three rounds of surveys. In addition,
supporting our information story, the heterogeneity analysis showed that Group T students whose usual
job search costs were estimated to be higher at baseline showed larger treatment effects.
In the second follow-up survey, we found a significant difference among treated students in the com-
position of sectors in which they searched and applied for jobs. Compared to the corresponding figure
for the students in Group C, Group T’s proportion of searches covering the three MDS among the seven
experimental sectors was higher by 10%, and the proportion of MDS jobs applied for was higher by 13%.
We also evaluated general labor market engagement outside of the experimental sectors. While we did not
find statistically significant evidence that Group T students either increased their intensity of search (in
terms of hours spent) or expanded the span of search scope (in terms of number of sectors searched), there
3
was a change in the composition of sectors searched in favor of MDS. Furthermore, for those employed
at the time of the survey, we found suggestive evidence that the jobs of those in Group T had relatively
lower female representation and better work arrangements.
These findings contribute to the literature on the gender-specific implications of incomplete informa-
tion in job search. Since the pioneering work by Hanson and Pratt (1991), researchers have suggested that
men and women job seekers have different information sets – due, for example, to differences in their
personal networks (Lindenlaub and Prummer, 2016; Jain et al., 2021), role models (Breda et al., 2020), job
advertisements (Kuhn and Shen, 2013), or the professional advice (Gallen and Wasserman, 2021) that they
receive – leading to segregation in the candidate jobs that they aspire to work in. This paper provides
experimental evidence on the specific characteristics of the information set that female college students
develop through their job search process by leveraging detailed data focused on their beliefs and search be-
haviors. In particular, it shows that female students systematically underestimate the relative superiority
of jobs in male-dominated sectors and stick to these prior beliefs, resulting in women’s excessive concen-
tration in a few female-dominated service sectors despite relatively moderate differences in preferences
by gender.
In addition, our work relates to the literature on the role of information frictions in creating inefficient
labor market matching. Researchers emphasize the role of information frictions, which can be understood
as separate from that of classical search frictions. For instance, evidence shows that job seekers have
strongly biased beliefs (or expectations) about potential job offers and that their belief updating is far from
the Bayesian benchmark under complete information (Jäger et al., 2022; Bandiera et al., 2020; Abebe et al.,
2020; Conlon et al., 2018). Experimenting with Korean senior college students who are well-educated and
who have high accessibility to job information on the internet, our study verifies the existence of informa-
tion frictions and their substantial implications for job matching even in the most advanced information
settings.
4
Finally, we add to the literature on experimental evaluation of information treatments on labor supply.
Random provision of information has been proven to have a salient positive impact on labor market par-
ticipation (Jensen, 2012), job search (Belot, Kircher, and Muller, 2019; Belot, Kircher, and Muller, 2018), and
major choice (Conlon, 2019; Wiswall and Zafar, 2015). Our study finds significant effects on occupational
choice in a setting closely mirroring the actual job market faced by college graduates around graduation,
and the effects are not restricted to members of certain occupations such as military cadets (Kofoed et al.,
2019) or teachers (Coffman, Featherstone, and Kessler, 2017). Within this context, our results provide use-
ful policy implications – namely, that informing college job seekers of relevant job characteristics across
various sectors can substantially affect the allocation of job search efforts and result in more efficient labor
market outcomes.
1.2 InstitutionalContext
Before introducing our experiment, we briefly describe its institutional context – that is, key characteristics
of the Korean young labor market and occupational segregation by gender.
The Korean young labor market is known for its remarkable proportion of college-educated, high-
skilled workers. About 70% of the total adult population aged 25–34 holds a tertiary degree, which is the
highest share among OECD countries and 25 percentage points higher than the OECD average of 45%
(OECD, 2020). As is typical in OECD countries, this proportion is higher for women (76.5%) than for men
(63.8%). On the other hand, Korea’s gender wage gaps are severe (OECD, 2021). The wage gap calculated
as the difference in median earnings of men and women relative the median earnings of men is 32.5%, 2.5
times larger than the OECD average of 12.8%.
To explain this coexistence of extremely high educational attainment among women and severe gender
wage gaps, researchers point to the importance of occupational segregation and between-occupation wage
5
differences (Tromp, 2019). Evidence shows that in Korea, as in many other countries (Anker, Melkas, Ko-
rten, et al., 2003; Blau, Brummund, and Liu, 2013), there is a clear distinction between jobs predominantly
occupied by men and those held by women and that the former pay more than the latter. Similar to the
trajectories reported for other advanced countries, the time-series development of the segregation shows
a decreasing but flattening slope. Although occupational segregation among workers overall has been
slightly alleviated throughout the past 20 years, that of college-educated workers has remained stable (Sa,
2015). Appendix Figure A.1 provides the distribution of the sectors that college graduates of both genders
entered in their first jobs over the last five years, computed from the Graduates Occupational Mobility
Survey (GOMS).
1
While male graduates are relatively more dispersed across sectors, female graduates are
concentrated in a few service sectors, such as retail trade, professional services, education, human health
activities, and social work activities. On the other hand, female graduates’ proportions in manufacturing
sectors are sparse in comparison to those of men, with only a few exceptions. This indicates that occu-
pational segregation comes primarily from female workers’ exclusion from sectors that are not typically
female dominated.
1.3 ExperimentalDesign
In this section, we introduce our experiment designed to empirically evaluate the role of information
frictions in female college students’ skewed labor supply in FDS. We motivate our experiment by showing
that MDS dominate FDS jobs not only in wages but also in nonwage amenities and thus that there is
no trade-off between those two dimensions, in contrast to what is implied by voluntary sorting under
complete information settings.
1
The GOMS is a nationally representative annual survey administered from 2006 by the Korean Employment Information
Service, which interviews 5% of all Korean college graduates in that year. One important feature is that the survey is conducted
approximately 18 months after graduation for each year’s cohort, which enables researchers to compare across years. Throughout
this article, we use these data not only to provide descriptive statistics on college graduates’ labor market outcomes but also to
generate our information treatment.
6
1.3.1 Motivation
1.3.1.1 SuperiorityofMDSinWagesandNonwageAmenities
One important feature of college graduates’ labor market with respect to gender gaps is that MDS provide
not only higher average wages but also better nonwage amenities to female workers.
Figure 1.1: Proportion of Women vs. Average Sector Characteristics for Female Workers
(a) Wages (b) Work hours
(c) Welfare institutions index (d) Job security index
Notes: The relationship between average sector characteristics and the female proportion of 73 two-digit sectors, which we
compute using the first jobs of 48,277 college graduates in the past five waves (2015 ∼ 2019) of the Graduates Occupational Mobility
Survey (GOMS). The definitions of the four characteristics are presented in Table 1.1. The red line is a quadratic fit.
Figure 1.1 displays the relationship between average sector characteristics for female workers and the
proportion of women, computed from the GOMS data.
2
Strong negative correlations with the proportion
of women are observed for average wages, a welfare institutions index and a job security index. Although
2
Our definitions of sector characteristics are detailed in Section 1.3.3, where we introduce our information treatment mate-
rials.
7
work hours are generally lower in FDS, the negative correlation comes from a few FDS with very low
average work hours.
3,4
Intuitively, the relationship weakens arguments that underscore the importance
of gender gaps in preferences, which assume that women sacrifice their wages for better welfare and
security.
5
Furthermore, Appendix Figure A.2 shows that the labor market outcomes do not align with the stated
preferences of female graduates. While women who prioritize wages and work hours do tend to enter
sectors that provide higher wages and lower work hours, selection on preferences is weaker for welfare
institutions and job security.
Although this evidence does not rule out gender discrimination from the labor demand side, it does
show that differences in sector characteristics and women’s preferences over them do not completely
explain the observed sectoral distributions of female workers in equilibrium. The gap between their stated
preferences and labor market outcomes may be attributable to lack of complete information – in the sense
that they do not know enough about sector characteristics – during their job search.
1.3.1.2 ConceptualFramework
Based on the findings in the previous section, we locate our central identification problem by incorporating
the role of information frictions into the classical model of occupational choice to motivate our experiment
with information treatment. Taber and Vejlin (2020) develop a dynamic structural model that encompasses
the roles of the four most important theoretical frameworks in labor economics – the Roy model, the
search model, the compensating differentials model, and the human capital model – in the wage (and
3
If the two sectors with the lowest average work hours – creative, arts and recreation related services (34.4 hours) and
education (36.5 hours) – are excluded, the relationship with the proportion of women becomes flat.
4
Flat relationships are also found with subjective measures – work difficulty and relevance to major – in the GOMS data. We
provide the figures in the Appendix.
5
The superiority of male-dominated sectors has been primarily studied within the context of institutional segmented labor
market theory. The theory argues that labor markets are divided into “primary” and “secondary” markets and that the former,
which are dominated by men, provide higher wages and better working conditions (Anker, 1998). However, we cannot find any
literature that systematically compares the wage and nonwage arrangements of female workers over the gender composition
spectrum of occupations.
8
occupational) dispersion across observably similar workers. The key feature of their model is that it is
based on thereservationutility framework considering utilities from both nonwage amenities and wages.
6
We simplify their model for identification by restricting our focus to its static block, based on empirical
evidence over a short time horizon on college students’ first job market match.
7
A job seekeri’s expected utility function from jobj can be characterized as
8
:
logEU
ij
= logp
ij
| {z }
subjectiveprob.
+(α i1
µ p
j
+
X
c=2
α ic
µ u
cj
)
| {z }
pref− weightedjobchars.
+ Λ i
|{z}
studentf.e.
+ η ij
|{z}
match− specifichetero.
(1.1)
wherep
ij
isi’s subjective probability of receiving an offer from job j,α ic
is her personal weight on wages
and other job characteristicsc (c = 1 represents wages), µ p
j
is the log wage of jobj, µ u
cj
is the common
worker utility from nonwage characteristicc of jobj,Λ i
isi’s individual characteristics common across
jobs, andη ij
is a match-specific idiosyncrasy.
Note that this classical equation assumes complete information on job characteristics, in that there are
no discrepancies between what job j’s characteristics actually are (µ p
j
, µ u
cj
) and what i believes them to
be. However, ample evidence has been reported that there are substantial gaps between college students’
beliefs about parameters and the actual parameter values.
9
By formally incorporating bias in beliefs on
population parameters into the equation, we obtain:
logEU
ij
=logp
ij
+(α i1
ˆ µ p
ij
+
X
c=2
α ic
ˆ µ u
icj
)+Λ i
+η ij
6
The literature on compensating differentials focuses on nonwage characteristics of jobs and solves the job search problem
within the context of reservation utility that explicitly considers utilities from those characteristics rather than just utility from
wages (Blau, 1991). For example, Groh et al. (2015) experimentally show that voluntary unemployment above the reservation
wage level in Jordan arose from young job seekers’ preferences over nonwage attributes.
7
In our baseline survey, about 90% of the students reported that aligning graduation and employment is important, and 60%
of the students were planning to apply for jobs in one year. More details are provided in the Appendix.
8
The detailed derivation of the equation is provided in the Appendix.
9
For example, Conlon’s (2019) survey of freshmen at the Ohio State University shows that the mean absolute errors about
the average salary of graduates range from 22% to 42% of the true values. Wiswall and Zafar’s (2015) examination on New York
University students and Betts’s (1996) study on University of California San Diego students show no less bias. In the next section,
we demonstrate the gaps observed in our own survey.
9
=logp
ij
+(α i1
µ p
j
+
X
c=2
α ic
µ u
cj
)− (α i1
bias
p
ij
+
X
c=2
α ic
bias
u
icj
)
| {z }
info.frictionscomponent
+Λ i
+η ij
(1.2)
wherebias
ij
=µ j
− ˆ µ ij
and ˆ µ ij
stands fori’s beliefs about jobj’s characteristics.
This equation adds the information frictions component to Taber and Vejlin’s (2020) model. If we as-
sume that the match-specific heterogeneity ( η ) is distributed randomly, the variations in students’ expected
utility can come from (i) subjective probabilities (p), (ii) preferences (α ), (iii) job characteristics (µ ), (iv) bi-
ases (bias), and (v) individual motivation (Λ ). Note that preferences (α ) are represented as weights in both
the second term and the third term. If the preference-weighted bias (the third term) is large enough to
offset the preference-weighted job characteristics (the second term), the role of preferences over job char-
acteristics in occupational choice may be limited, while that of subjective probabilities becomes relatively
more prominent. For example, if female students systematically underestimate the relative advantages of
MDS over FDS for some reason, their expected utility from an MDS job would not be substantially higher
than that from an FDS job.
Our experiment boils down to comparing female college students’ behaviors (aspirations, job search,
and applications) that correspond to the left hand-side of equation (2) with and without a randomized in-
tervention onbias
ij
. More specifically, we provide students in the treatment group with direct information
on the population parameters of diverse jobs without affecting other components – subjective probabilities
and preference weights – to evaluate the importance of information frictions in occupational segregation
from the labor supply side. Since we randomly assign students to the treatment and control groups, the
distributions ofΛ i
andη ij
are orthogonal to the treatment assignment. Figure 1.2 summarizes the struc-
ture of our experimental provision of information and the mechanism through which it affects beliefs,
aspirations, job search and applications, and the resulting equilibrium.
10
Figure 1.2: Conceptual Framework of the Experiment
1.3.1.3 KeyChallengesinIdentification
We face two key problems in identifying the theory-based expected utility function suggested in the pre-
vious section in existing employee–firm match datasets. First, the data reflect the equilibrium determined
not only by labor supply factors but also by demand factors, particularly gender discrimination, whether
taste based or statistical. Second, it is difficult to quantitatively measure diverse compensating differen-
tials and properly weight them with preferences. The first problem can be addressed by using only data on
supply-side behaviors instead of those on labor market outcomes.
10
We implement a survey that directly
asks students about their aspirations from the labor supply side at the beginning of their job search process
and trace their job search and application behaviors around graduation.
The second problem is more subtle and has been less addressed by the literature because unlike wages,
nonwage attributes are multidimensional and difficult to measure.
11
Ideally, an econometrician should be
able to observe characteristics of all sectors and each student’s preference weights and (biased) beliefs about
these characteristics to perfectly decompose the variations on the right-hand side. Since this is infeasible
10
Good recent examples are studies that analyze application behaviors based on unemployment insurance data (Fluchtmann
et al., 2020; Eriksson and Lagerström, 2012), data from general job search platforms (Marinescu and Skandalis, 2021), or responses
to retrospective surveys conducted a few years after graduation (Cortés et al., 2021).
11
Mas and Pallais (2017) address the difficulty in measuring the (subjective) value of nonpecuniary attributes by implementing
a discrete choice experiment that makes participants choose between jobs with diverse attributes, particularly worker flexibility
and discretion, which are given as perfectly known information.
11
in the real world, how to reduce the dimensionality becomes important to the empirical strategy.
12
This
study’s novel solution is to implement a framed experiment with four sector characteristics and seven
prespecified sectors computed from a nationally representative survey of college graduates and carefully
chosen to represent the sample’s first job choice problem. The experiment compares female participants’
aspirations and behaviors related to the seven sectors with and without provision of information on four
characteristics – wage, work hours, welfare, and job security. Details of this strategy are explained in the
following sections.
1.3.2 SamplingandGroupAssignmentatBaseline
Our experiment targeted female senior college students in Korea. At baseline, we invited 1,000 female
senior college students with humanity and social sciences majors from a preconstructed panel from Em-
brain.com, a representative online survey company in Korea. The voluntary participants were screened by
the company’s panel management program and compensated by the contract between the two parties.
13
Randomization was administered at the individual level, with 500 respondents assigned to the treat-
ment group (Group T) and the other 500 to the control group (Group C). The baseline survey was admin-
istered throughout October 2021, which was the beginning of the last semester in college for most senior
students.
14
1.3.3 InformationTreatment
The baseline survey questionnaires consisted of four sections – demographics, job search behaviors, job in-
formation sources, and an experiment with information treatment. The questionnaires for the two groups
12
This is the primary reason why experimental evaluations of the occupational choice problem are scarce. Kofoed et al. (2019)
recognize this problem and cite it as the motivation for their framed experiment on cadets’ occupational choice in the U.S. Army,
an approach that dramatically reduced dimensionality at the cost of generalizability.
13
We also invited 100 male students to compare gender differences in job search behaviors. They received the same baseline
questionnaire as female students in the control group. The analysis using their responses is provided in our Appendix.
14
In Korea, the academic year starts in the spring every year. The fall semester usually starts in September.
12
differed only in the last section. While Group T was provided with the information treatment, Group C
was not.
In the fourth section, Group T students were shown a newspaper article about college graduates’ labor
market outcomes, which included a table (Figure 1.3)
15
demonstrating the populationaverages of the four
focal characteristics – wages, work hours, welfare institutions index, and job security index – for first-year
female employees’ jobs
16
in 20 industrial sectors. The average numbers were computed with the data from
the most recent five waves (Wave 2015 to Wave 2019) of the GOMS.
17
The 20 sectors are where the largest
proportion of female college graduates were employed, and the four characteristics are those that they
reported as their primary considerations (preferences) in searching for jobs.
18
As explained in Table 1.1,
each measure of the four characteristics was computed for each graduate’s first job. Then, the averages
(weighted by sampling weights) were taken across all female graduates in each of the sectors. In the article,
we explained to the students in detail exactly how all these numbers were computed.
When showing the numbers to the students, we adopted a few strategies to maximize the treatment
effects while minimizing experimenter demand effects.
19
First, we did not provide the gender composi-
tions of the sectors or explicitly classify those sectors as MDS or FDS in the table to detach the role of
job characteristics information from other confounding considerations on sectors dominated by certain
genders. Instead, we colored the top three sectors on each characteristic in blue and the bottom three in
red to let students identify by themselves that MDS are more likely to be included in the top three sectors
while many FDS are in the bottom three sectors. Second, we showed the numbers in absolute terms in odd
15
This is an English-translated version with a simplified design. The full treatment materials used (in Korean) are provided in
the Appendix.
16
Providing job seekers with statistics on employees who recently entered their jobs instead of average statistics for all em-
ployees in the sectors has a clear advantage in that the former data are more easily interpretable by job seekers with less dispersion
and are not distorted by differential increases in benefits by tenure across sectors.
17
Compared to the respondents in our survey sample, GOMS respondents are more dispersed with respect to region and major,
since the GOMS is conducted with stratified randomization.
18
For the full list of considerations (preferences) reported in the survey, see the Appendix.
19
Haaland, Roth, and Wohlfart (2020) provide excellent suggestions on implementing an information treatment, including
on issues related to experimenter demand effects and anchoring. We acknowledge that many features of our experiment were
designed based on their suggestions.
13
Table 1.1: Construction of Sector Parameters
Parameters Methods of construction
Wages The (weighted) average of annual salaries (including bonuses, if any) by sector,
inflation adjusted by the consumer price index (2020=100)
Work hours The (weighted) average of reported weekly work hours (regular and overtime) by
sector
Welfare institutions
index
(Step 1) Count the number of applicable welfare institutions among the following:
- severance pay, paid leave, overtime pay, bonus, weekly allowance, pension,
medical insurance, employment insurance, industrial accident insurance
(Step 2) Take the (weighted) average by sector and rescale to 0–5
Job security index (Step 1) Count the applicable job security status between the two:
- whether the worker is a regular employee
- whether the worker’s employment contract is not fixed term
(Step 2) Take the (weighted) average by sector and rescale to 0–5
columns and additionally provided the numbers relative to our gender-neutral benchmark sector, retail
trade, in even columns to allow easier cognitive comparison across sectors and alleviate concerns over
unconscious numerical anchoring.
Among the top 20 sectors provided in the table, we selected seven experimental sectors that in to-
tal account for 30% of the first jobs of college graduates with humanities and social sciences majors for
our evaluation on beliefs and behaviors throughout the surveys. To compare female students’ job search
behaviors across MDS and FDS, we chose three MDS (wholesale trade; manufacture of electronic com-
ponents, computer, visual, sounding and communication equipment; and financial service activities) and
three FDS (business support activities, education, and social work activities), as well as the gender-neutral
retail trade sector.
20
The numbers for the seven sectors were provided in bold in the table.
20
The shares of employment and the female proportion of the top 20 private sectors are reported in Appendix Table A.1.
Considering female graduates’ concentration in a few sectors, we selected three FDS with a proportion of women over 60% while
picking three MDS with a proportion under 45%.
14
Figure 1.3: Table Provided in the Information Treatment (translated to English)
After showing the participants the article, we administered two exercises to ensure that they read and
understood the numbers properly. First, adopting Conlon’s (2019) strategy, we let students type in all 28
numbers (4 characteristics× 7 sectors) to two decimal places. Most of the students diligently undertook
this tedious job and filled out the correct numbers.
21
Then, they were asked Likert-scale questions com-
paring those numbers to their prior beliefs – whether the numbers were larger or smaller than they had
previously thought. Our intention with this set of exercises was to allow students in the treatment group
21
A total of 54.8% of the students correctly typed in all 28 numbers to two decimal places, and 84.6% did so for more than 25
numbers.
15
to develop a sense of the absolute levels of the four characteristics and the gaps across the experimental
sectors without directly mentioning that “MDS are better than FDS”.
Students in Group C were not provided the table. Instead of providing information, we asked about
their beliefs about the average values of the four characteristics for the seven experimental sectors. Since
it is difficult for students to assess nonwage amenities in numbers, we provided the exact same definition
of each variable as what we showed in the article for Group T.
All in all, it should be noted that we carefully designed the questionnaires to provide information on
average sector characteristics without affecting any other factors that influence expected utility – especially
subjective probabilities and preferences – as defined in equation (2). This design allows us to separate out
the role of information frictions with limited concerns over the confounding effects that are prominent in
interventions related to role models or mentoring.
1.3.4 Follow-upSurveysandOutcomeVariables
After the baseline survey with treatment, we administered two rounds of follow-up surveys in November
2021 and February 2022 to keep track of students’ beliefs and behaviors. Since our research is about occu-
pational segregation, we primarily focus on the composition of the sectors that female students aspired to
enter and in which they searched and applied for jobs across the three rounds of surveys.
At baseline, the students’ aspirations to enter each of the seven experimental sectors on a 1∼ 10 Likert
scale were elicited right after the information treatment was given (Group T), or the students were simply
asked about their beliefs (Group C).
22
In the first follow-up survey, the students’ beliefs and aspirations were elicited again to check their
trajectories one month after the intervention. The differences in beliefs and aspirations across treatment
arms were evaluated to verify the first-stage validity of the information treatment in the short run. In
22
The survey question for eliciting the aspirations was: “Based on the information you have, please state your aspirations
(willingness to get a job) to enter each of the seven sectors in 1∼ 10 Likert scale.”
16
addition, rich questions on the modes of information acquisition and intensity and sectoral allocation of
job search were included to shed light on how their information set was being developed during their job
search process. In addition, we embedded a “tailored job information newsletter” experiment in this round,
where students were asked to choose among the names of 49 actual hiring firms about which they wanted
to learn more information.
23
The second follow-up survey was administered in February 2022, when most students had graduated
and had engaged in job search. Our main outcome variables in this round were constructed with the
students’ actual history of job search and applicationsinside andoutside of our experimental sectors. First,
we measured the students’ beliefs and aspirations again to evaluate whether the treatment effect on bias
has strengthened or dissipated over the four months. Then, the students were asked about whether they
had searched/applied for jobs in our seven experimental sectors and why they did not searched for jobs
in each of those sectors if they had not. Finally, we asked about their general labor market engagement,
including their job search, application, and employment history. Those who were employed at the time of
the second follow-up survey were asked to provide their detailed job characteristics.
The construction of the main outcome variables is explained in Appendix Table A.2 in detail. Since the
most data points were gathered at the individual–sector level (e.g., student i’s aspiration to enter sector
s), aggregated variables at the individual level were used for estimation. For example, to measure the
treatment effects on aspirations, we computed the (standardized) sum of aspirations to enter the three
MDS (wholesale trade, manufacture of electronic components, and financial service activities) divided by
those to the three FDS (business support activities, education, and social work activities). Likewise, the
treatment effects on the sectoral allocation of job search/applications were measured as the proportion of
MDS jobs searched/applied for among our seven experimental sectors.
23
We report the details of this experiment in Appendix A.2. Students’ choice of firms in this experiment turned out to be
consistent with their self-reported sectoral allocation of search efforts.
17
From the three rounds of surveys, we could construct panel data that capture the four months of college
graduates’ transition into the labor market. The unique feature of our dataset is that we look into the black
box of the understudied relationships between information (bias), aspirations, job search behaviors, and
the resulting labor supply decisions.
1.3.5 ChecksonBalanceandAttrition
Table 1.2 shows the balance of observable characteristics between the two groups at baseline. The students
were about 23 years old and had spent seven semesters out of eight in college. Most of them had started
their job search efforts at the time of the survey, but only less than 30% had ever applied for any job
openings, making them our ideal sample for scrutinizing the process between initial job search and the
resulting labor market outcomes.
Students in both groups were well balanced in terms of demographics, individual traits (risk preferences
and patience), and job search history at baseline. We were unable to reject the null hypothesis that the
observable variables were jointly orthogonal to treatment status (p = 0.537). Following De Mel, McKenzie,
and Woodruff (2019), we report the normalized differences in the fourth column.
24
This scale-invariant
measure also shows good balance, in that the largest differences are around 0.1 standard deviations.
The attrition and balance in the follow-up surveys are reported in Appendix Table B.4. In the first
follow-up survey, the attrition rate was 19.3% (193 out of 1,000) and was balanced across treatment groups
(p value for the differential attrition test = 0.379). On the other hand, there was selective attrition on two
observable characteristics, parents’ monthly income and the specific major distribution within humanities
and social sciences major categories, with the differences in both significant at 10%. In the second follow-
up survey, the attrition rate relative to the first follow-up was higher at 27.8% (225 out of 807) but was still
balanced across groups (p value for the differential attrition test = 0.249). There was selective attrition on
24
The normalized differences are computed as ( (XT − XC)/
p
(s
2
T
+s
2
C
)/2), where X and s
2
are the sample mean and
variance of each group.
18
two observable characteristics, mother’s education, and the specific major distribution, with the differences
significant at 10% and 5%, respectively. This statistically significant selective attrition indicates that these
variables should be controlled for when we evaluate the treatment effects. As explained in the next section,
we adopted the double LASSO procedure to automatically select relevant control variables among those
in Table 1.2, which effectively controlled for those with selective attrition.
Table 1.2: Balance at Baseline
Group C Group T p
value
Norm.
Diff.
(1) (2) (3) (4)
Panel A. Demographics & background
Age 22.72 22.56 0.103 -0.103
Num. of semesters in college 7.22 7.19 0.684 -0.026
GPA (/4.0) 3.59 3.59 0.791 -0.017
Num. of formal certificates 1.58 1.54 0.581 -0.035
College in metropolitan area (0/1) 0.60 0.58 0.607 -0.033
Parents’ monthly income (10,000 KRW) 549.7 523.7 0.119 -0.099
Mother’s yrs. of education 14.5 14.5 0.906 0.008
Panel B. Attitudes & personal traits
Risk preference in financial matters (1 ∼ 7) 3.50 3.49 0.941 -0.004
Risk preference in daily life (1∼ 7) 3.42 3.48 0.457 0.047
Patience (1∼ 7) 4.63 4.66 0.725 0.022
Panel C. Job search history at baseline
Started job search (0/1) 0.95 0.97 0.115 0.100
Ever applied for job openings (0/1) 0.28 0.26 0.476 -0.045
Planning to apply in one year (0/1) 0.58 0.58 0.848 0.012
Kolmogorov–Smirnovp value of majors dist. 0.111
Joint orthogonalityp value 0.537
Sample size 500 500
Notes: This table shows the balance of observable characteristics between the treatment and control groups at baseline. Column
(3) reports thet testp values, and Column (4) reports the normalized differences. Risk preference and patience were self-assessed
on a 1∼ 7 Likert scale, following the strategies of Cortés et al. (2021) and Schnitker (2012), respectively. Kolmogorov–Smirnov
p value is the result of the Kolmogorov–Smirnov test of identical distribution of majors within humanities and social sciences
categories. Joint orthogonalityp value is the result of anF test with the null hypothesis that the variables are jointly orthogonal
to treatment status. The attrition and balance in the follow-up surveys are reported in Appendix Table B.4.
We also checked for potential spillover of treatment effects across groups by asking students whether
they knew anyone else who had participated in our experiment, although the chance of spillover was
19
expected to be low because our experiment took the form of a survey experiment independently adminis-
tered at home. As expected, 91% of the respondents in the second follow-up survey reported that they did
not know any other participants. Among the remaining 9%, 4.2% had never shared any information with
other respondents.
1.4 EmpiricalSpecificationsandResults
1.4.1 EmpiricalSpecifications
Since the individual-level randomization process successfully created a balanced assignment between
Group T and Group C, the form of our main empirical specification is straightforward:
Y
i
=β 0
+β 1
Treat
i
+X
′
i
θ +ϵ i
where Y
i
is the outcome variables for student i, Treat
i
is an indicator equal to 1 if student i belongs
to Group T and 0 otherwise, and X
′
i
is a vector of the control variables represented in Table 1.2. Our
main specification adopts the double LASSO procedure, suggested by Belloni, Chernozhukov, and Hansen
(2014), to systematically select relevant control variables without overfitting the model.
25
We use robust
standard errors for the regressions since the randomization was administered at the individual level.
26
25
In the Appendix, we provide a list of the control variables selected by the double LASSO procedure in all regressions in
this paper, showing that the only selected variable is specific majors (within the humanities and social sciences category) in most
cases.
26
Our empirical specifications and key outcome variables were prespecified in the preanalysis plan registered to the AEA RCT
Registry (AEARCTR-0008243).
20
1.4.2 Results
1.4.2.1 ImpactsonBeliefs
Figure 1.4 compares the trajectories of the students’ beliefs in the two follow-up survey rounds and the
population parameters that were provided to Group T at baseline.
27
The beliefs of female students in both groups about the four sector characteristics show substantial
gaps with the actual population parameters (dashed gray lines). In the first follow-up survey, the students
in Group C significantly underestimated the wages, welfare institutions, and job security of MDS jobs. On
the other hand, they overestimated FDS jobs in terms of wages and job security. It is important to note
that the gaps in beliefs about MDS and FDS job characteristics were smaller than the real gaps for all four
characteristics. After three months of job search (between the first and the second follow-up surveys),
their beliefs came closer to reality for most sectors and characteristics, but substantial gaps still remained
in February 2022, their month of graduation.
28
Our treatment had substantial impacts, shifting the beliefs of Group T students away from those of
their Group C counterparts in the first follow-up survey. Large treatment effects were observed in the
beliefs on the welfare institutions and job security of MDS jobs, while there were smaller but significant
corrections to beliefs on the wages and job security of FDS jobs.
In the second follow-up survey, on the other hand, we could see some catching-up of Group C over time,
consistent with the students acquiring information through their additional three months of job search.
However, compared to Group C students, Group T students still evaluated MDS jobs as better on average
in all four characteristics and FDS jobs as worse in wages, welfare, and job security, with statistically
significant differences from the outcomes of Group C for the job security of MDS jobs ( p=0.001) and the
welfare of FDS jobs (p=0.099).
27
Note that we do not compare the beliefs at baseline because at that time, we provided Group T with the actual parameters
instead of asking about their beliefs. Therefore, the first comparable beliefs are those assessed in the first follow-up survey.
28
Appendix Table A.4 compares the parameters and the students’ beliefs in numbers (for wages and nonwage amenities,
respectively). The table also includes thep values for thet test comparing the gaps between the two groups.
21
Figure 1.4: Belief Updating
(a) Salary (annual) (b) Work hours (weekly)
(c) Welfare institutions index (0∼ 5) (d) Job security index (0∼ 5)
Notes: The figures summarize the students’ beliefs on the averages of the four characteristics of MDS (wholesale trade, manufacture of electronic components, financial service
activities), FDS (business support activities, education, social work activities), and a neutral sector (retail trade). Means and 95% confidence intervals reported. Dashed gray lines
represent the population parameters, computed from the GOMS. Units on y-axis: wages (10 thousand KRW/year), work hours (hours/week), welfare institutions index (0∼ 5 Likert),
job security index (0∼ 5 Likert).
22
All in all, this result indicates two important facts. First, there were substantial information frictions
such that the students underestimated the actual differences between MDS and FDS jobs – even those
students who were actively searching for job information in their last semester.
29
Second, the first stage
effect of our treatment on beliefs was valid throughout the experimental period.
1.4.2.2 ImpactsonAspirations
The first main outcome variable of our experiment is self-reported aspirations to enter the seven exper-
imental sectors, which were collected in all three rounds of surveys and reported in Table 1.3. The de-
pendent variable is the (standardized) relative aspirations to enter MDS over FDS jobs, as defined in Table
A.2.
In Column (1), Group T respondents show higher aspirations to enter MDS jobs by 0.2 standard devia-
tions at baseline in comparison to Group C respondents. It means that the students who were provided with
accurate information on the seven experimental sectors expressed significantly higher (lower) aspirations
to enter MDS (FDS) jobs.
30
The magnitude of the treatment effect is remarkably persistent throughout the
three rounds of surveys, and the 0.2 standard deviation gap is sustained (Columns (2) and (3)).
To test whether the treatment effects came from the dissolution of information frictions, we conducted
a heterogeneity analysis by measuring each student’s job search costs and examining how the treatment
effects varied across those with different search costs. More specifically, we computed a student’s job search
cost by multiplying her self-reported opportunity cost for an hour of work
31
and her expected hours to
29
This finding contributes to the literature in that it indicates heterogeneity in information frictions, which vary across jobs and
personal characteristics such as gender. Previous direct assessments of college students’ beliefs have mostly referred to average
wages and considered graduates with certain majors (Wiswall and Zafar, 2015; Conlon (2019)). Zooming in on information
frictions across jobs and genders not only mitigates concerns over the variance of measured beliefs by restricting the scope of
jobs about which beliefs are elicited but also underscores the importance of understanding the implications of gender-job specific
information barriers for women’s occupational choice.
30
If we zoom in on the treatment effects by individual sector, the effects come primarily from higher aspirations to enter
wholesale trade and manufacture of electronic components and lower aspirations to enter education. The three sectors are those
with the largest gaps between Group C students’ measured beliefs and the population parameters at baseline. The results are
provided in the Appendix.
31
The question for measuring opportunity cost was as follows: “Assume that someone wants to hire you part time. What is
the minimum hourly wage you would accept to do this job over the next few weeks?”
23
Table 1.3: Treatment Effects on Aspirations
Std. Relative Aspirations MDS > FDS
(0/1)
(1) (2) (3) (4) (5) (6)
Treated (0/1) 0.213*** 0.203*** 0.240*** 0.170** 0.133* 0.107***
(0.060) (0.067) (0.080) (0.068) (0.073) (0.029)
Search Cost (Wage) -0.010***
(0.004)
Treated× Search Cost (Wage) 0.015
(0.010)
Search Cost (Nonwage) -0.006
(0.005)
Treated× Search Cost (Nonwage) 0.023*
(0.013)
Mean of Dep Var in C -0.100 -0.093 -0.118 -0.100 -0.100 0.362
Survey Round Base 1
st
follow-up
2
nd
follow-up
Base Base Base
Observations 1,000 807 582 1,000 1,000 1,000
Notes: This table shows the regression results of the treatment status on aspirations to enter MDS and FDS jobs. Robust standard errors in parentheses. The regressions are OLS
with control variables chosen by the double LASSO method among the variables in Table 1.2. Columns (1)∼ (3) show the treatment effects on the standardized relative aspirations
across the three rounds of surveys. For Column (3), a dummy for current employment status (employed = 1) is added in the regression. Columns (4) and (5) add interaction terms
between the treatment status and estimated search costs for wage and nonwage information. The dependent variable of Column (6) is a dummy taking 1 if a respondent’s average
aspiration to enter MDS jobs is higher than her average aspiration to enter FDS jobs.
24
obtain enough information (including wages and nonwage attributes) about one firm through her usual
information sources. The results in Columns (4) and (5) provide supportive evidence for our argument.
Students in Group C had lower relative aspirations to enter MDS jobs when their usual search costs were
higher. On the other hand, the treatment effects were stronger for the students in Group T with higher
search costs, especially costs for obtaining nonwage information. This indicates that our information
treatment complemented the students’ usual information sources to correct their beliefs and reduce the
costs of evaluating MDS and FDS jobs accurately.
Finally, Column (6) shows the result of a linear probability regression with a dummy for having higher
average aspirations to enter MDS than FDS jobs as dependent variable. Students in Group T were 11%
more likely to have higher aspirations to enter MDS than FDS, suggesting two facts. First, the treatment
effects were salient enough for some female students to prefer MDS to FDS jobs. Second, from the opposite
perspective, the difference in probability of going over the threshold to prefer MDS is 11%, indicating that
only a limited number of students’ actual labor supply decision would be swayed toward MDS jobs, despite
the substantial aspirational boost for MDS.
32
1.4.2.3 ImpactsonSectoralAllocationofJobSearchandApplications
The second set of outcome variables is the students’ job search and application history during the four-
month transition. Our primary interests lie in whether they searched or applied for jobs in each of the seven
experimental sectors. The job search/application history in the seven sectors were elicited retrospectively
in the second follow-up survey. Since those sectors are the major employers of Korean college graduates
with humanities and social sciences majors, students were highly likely to be exposed to some level of
information about at least one of them even without treatment.
32
Note that occupational choice is intrinsically a discrete choice problem. In our conceptual framework,i chooses jobj such
thatj = argmax
j∈J
logEUij , whereJ is a finite discrete choice set. Therefore, not every increase in continuous expected utility
causes a change in occupational choice.
25
Table 1.4: Treatment Effects on Sector Choices ( 2
nd
follow-up)
7 Sectors Num. MDS Prop. MDS 7 Sectors Num. MDS Prop. MDS
Ever Searched Searched Searched Ever Applied Applied Applied
(0/1) (0/1)
(1) (2) (3) (4) (5) (6)
Treated (0/1) 0.091** 0.095* 0.097** 0.044 0.049 0.126**
(0.038) (0.055) (0.041) (0.037) (0.040) (0.061)
Mean of Dep Var in C 0.653 0.487 0.409 0.307 0.203 0.447
Observations 582 582 407 582 582 188
Notes: This table shows the regression results of the treatment status on the sector choices during job search/application reported retrospectively at the second follow-up survey.
Robust standard errors in parentheses. Sample: 582 female students who responded the second follow-up survey. Dependent variables: (1) A dummy for the seven sectors ever
searched / (2) The number of MDS searched (among the three MDS) / (3) The proportion of MDS searched (among sectors searched) / (4) A dummy for the seven sectors ever
applied for / (5) The number of MDS applied for (among the three MDS) / (6) The proportion of the MDS applied for (among sectors applied for). The regressions are OLS with
control variables chosen by the double LASSO method among the variables in Table 1.2. The variables selected by the double LASSO method are provided in the Appendix.
26
Column (1) of Table 1.4 shows that 65% of the students in Group C searched for information about
jobs in at least one of the seven sectors during the four months. In addition, the students in Group T
were 9% more likely to search for jobs in those sectors with a statistical significance at 5%. According to
the following two columns, students in Group T searched more in MDS in terms of absolute number and
proportion among the sectors in which they searched for jobs. To summarize, the students provided with
the information treatment not only searched more for jobs in the seven sectors about which information
was provided but also significantly corrected their job search composition according to the changed beliefs.
On the other hand, the regression results with application history for the seven sectors and the three
MDS (Columns (4) and (5)) show smaller treatment effects in terms of both magnitude and statistical sig-
nificance. However, the proportion of MDS applications was substantially larger for Group T, indicating
that some of the students were discouraged from applying for FDS jobs.
33
All in all, this sector-level evidence shows that our information treatment had tangible impacts not only
on students’ stated aspirations but also on their actual job search and application behaviors. The result is
consistent with the students’ sectoral allocation measured at the firm level in the “tailored job information
newsletter” experiment that we embedded in the first follow-up survey, as explained in Appendix A.2.
1.4.2.4 ImpactsonLaborMarketEngagementandEmploymentOutcomes
In this section, we examine the actual labor market outcome variablesoutsideof our experimental sectors
(Table 1.5). In the second follow-up survey administered in February 2022, 43% out of the 582 students
who responded the survey had graduated from college, and 15% were employed.
34
There are no significant
33
We implemented the same heterogeneity analyses as we did in the previous subsection but with students’ job search costs
for all of our outcome variables here (Appendix Table A.5). Although the results are statistically less significant, the sign of the
estimated coefficients are generally negative. This indicates that treated students’ self-search throughout the four months after
the information treatment was likely to be limited for those with higher-than-usual search costs.
34
There are a few potential reasons why these numbers turned out to be lower than expected. First, the spring 2022 job market
was particularly disadvantageous to young job seekers due to the COVID-19 crisis (Park and Cho, 2022). Second, the survey was
administered at the time of graduation, and more employment contracts could have been finalized within a few months after
graduation. Third, attrition may have been more severe among those who had graduated from college or were employed because
the value of the time spent participating in the survey could be higher for these respondents.
27
Table 1.5: Treatment effects on Labor Market Engagement/Employment Outcomes ( 2
nd
follow-up)
Panel A. Labor market (1) (2) (3) (4) (5) (6)
engagement Graduated Applied Num. of Employed Emp.
(0/1) for jobs application (0/1) timing
(0/1) (months)
Treated (0/1) -0.010 0.049 0.166 -0.006 -0.634
(0.032) (0.036) (0.235) (0.027) (0.394)
Mean of Dep Var in C 0.433 0.447 1.715 0.147 2.773
Observations 582 582 582 582 80
Panel B. Characteristics
of jobs (if employed) Female Wages Work Welfare Job Job
proportion (annual, hours institution security quality
10,000
KRW)
(weekly) index
(0∼ 5)
index
(0∼ 5)
index
(Std.)
Treated (0/1) -0.025 317.4* -0.116 0.171 0.473 0.162
(0.025) (189.6) (2.249) (0.284) (0.470) (0.111)
Mean of Dep Var in C 0.548 2821.4 39.114 2.477 3.068 -0.073
Observations 79 80 80 80 80 80
Notes: This table reports the regression results of the treatment status on labor market engagement (Panel A) and employment
outcomes (Panel B) at the second follow-up survey. Robust standard errors in parentheses. Sample: 582 female students who
responded the second follow-up survey. The sample for Column (5) of Panel A and all columns of Panel B are the 80 students
who were employed. The dependent variable of Column (6) of Panel B is a standardized job quality index, which we compute by
aggregating the standardized characteristics (wages, work hours, welfare institutions index, and job security index) with equal
weights. The regressions are OLS with control variables chosen by the double LASSO method among the variables in Table 1.2.
The variables selected by the double LASSO method are provided in the Appendix.
differences in the likelihood of graduation or employment between the two groups. Although the results
are noisily estimated, students in Group T were more likely to apply for jobs and also applied for more job
openings. In addition, students in Group T were hired 0.6 months earlier than those in Group C.
Panel B compares the characteristics of the jobs where the students were employed. We asked the
employed graduates detailed questions about the characteristics of their jobs, including wages, work hours,
welfare institutions, and job security, defined to mirror the construction of the sector parameters for the
28
experiment. Then, we also matched the names of the firms with the two-digit sectors and their gender
composition to figure out whether each job was located in MDS or FDS.
35
Students in Group T were employed in jobs with 3% lower proportions of women on average than the
jobs secured by their counterparts in Group C, although the statistical power is limited due partly to the
limited number of total number of newly employed students. Columns (2) to (5) also show an interesting
pattern: the students in Group T were employed in jobs with higher wages, lower work hours, and better
welfare and job security. As a result, their jobs scored 0.16 standard deviations higher on the standardized
job quality index (p=0.146) than the jobs of Group C students (Column (6)).
This indicates that the change in composition of job search and applications observed among the seven
experimental sectors from our information treatment could shift the students’ actual employment out-
comes in broader settings. As shown in the previous sections, the female students did indeed benefit from
changing their beliefs on the key characteristics of potential employers in MDS. Then, they demonstrated
higher (lower) relative aspirations to enter MDS (FDS) jobs and searched for more information about MDS
than about FDS jobs. The equilibrium outcome, albeit estimated noise, turned out to be that they were able
to be employed earlier with better employment conditions in less female-dominated sectors.
36
1.4.2.5 ChecksonOtherComponentsoftheConceptualFramework
Our experiment was carefully designed to affect only the information component such that we could exam-
ine its role in occupational segregation in labor supply. Here, we check whether the information treatment
caused any unintended impacts on other components of our conceptual framework introduced in Section
35
Matching of the firm names and two-digit sectors was implemented with the sector classification provided by Incruit.com,
a Korean job information company.
36
The result on employment timing requires careful interpretation in that there could be two different drivers in opposite
directions. First, female students with an expanded job search scope could have higher reservation utility, contemplating job
offers longer. Second, they could receive more and better MDS job offers, which could not have occurred without their job search
in those sectors. Our result is more consistent with the second mechanism, suggesting that students achieve better employment
outcomes with mitigated information frictions.
29
Table 1.6: Treatment Effects on Other Components ( 1
st
and2
nd
follow-up)
(Log) Preference weights (allocation of 8 points)
Subjective
Prob.
Wages Work hours Welfare
institutions
Job security
(1) (2) (3) (4) (5)
Treated (0/1) -0.004 0.071 0.058 -0.059 -0.070
(0.047) (0.080) (0.061) (0.057) (0.070)
Cluster s.e. Individual No No No No
Round 1
st
follow-up
2
nd
follow-up
2
nd
follow-up
2
nd
follow-up
2
nd
follow-up
Observations 4,715 582 582 582 582
Notes: This table shows the regression results of the treatment status on the students’ subjective probabilities (Column (1)) and
preference weights for each of the four sector characteristics (Columns (2)∼ (5)). The regressions in Columns (2)∼ (5) are OLS
with control variables chosen by the double LASSO method among the variables in Table 1.2.
1.3. In equation (2), other components that might affect a student’s expected utility are subjective proba-
bilities (p) and preference weights (α ). Intuitively, provision of information could have confounding effects
with those components, hampering accurate identification of the impact of correcting bias on sector char-
acteristics.
To alleviate these concerns, we checked the differential changes in those components across groups.
First, in the first follow-up survey, we asked students about their subjective probabilities of successfully
entering each sector around graduation if they had started preparing then. We tested whether each student
i’s subjective probability for sectorj was different across groups by estimating a pooled OLS regression at
the individual–sector level:
logp
ij
=β 0
+β 1
Treat
ij
+Sector
j
+ϵ ij
where Sector
j
is sector fixed effects. Standard errors are clustered at the individual level. Column (1)
of Table 1.6 shows that there are no significant differences in the distribution of subjective probabilities
between the two groups.
30
Second, in the second follow-up survey, we asked students to assign a total of eight points across
our four core job characteristics according to their importance in students’ job search. This corresponds
to the preference weights (α ) in the model. Columns (2) to (5) show the regression results from the same
individual-level double LASSO specification used in the main analysis, in this case with the points assigned
to each job characteristic on the left-hand side. No significant treatment effects are found for any of these
preference weight components.
1.5 ADiscussionontheMechanism
The evidence drawn from our experiment reveals a hidden sequence of the decision-making process for
female students’ labor supply decision. For the information treatment to ultimately affect labor supply
decisions, (i) new information had to have changed the students’ beliefs and aspirations, (ii) the students
must have altered their job search behaviors according to the updated beliefs, and finally, (iii) the stu-
dents must have applied for/accepted job offers that provided higher reservation utility than their updated
thresholds.
To summarize our results, we found robust treatment effects on updating students’ beliefs (Figure
1.4) and aspirations (Table 1.3). Although we also found statistically significant effects on job search and
application behaviors, the effects were more concentrated on the sectoral allocation of search (Table 1.4)
than on the intensity and scope of search efforts.
37
In other words, students were more likely to reallocate
their restricted job search resources from FDS to MDS openings, instead of simply adding searches for
MDS jobs to their plates. To explain this, we exploit the richness of our data and discuss several pieces
of illustrative evidence on behavioral heuristics – confirmation bias and the anchoring effect – in the job
37
In the first follow-up survey, we directly asked students about the changes in weekly time spent on job search and in their
job search scope from the baseline. There were no significant differences in average time spent on job search across groups. In
addition, the proportion of the students responding that they had searched in more sectors over the baseline was only slightly
larger in Group T. The results are provided in the Appendix.
31
search process that could have undermined the impact of the dissolved information frictions on actual
behavioral changes.
In line with what has been consistently reported in the literature, the effects of the information treat-
ment on beliefs were salient in our experiment. However, if we come back to Figure 1.4, there still existed
notable differences between the population parameters and the beliefs of the treated students consistently
in the direction of underestimation of the gaps between MDS and FDS jobs. Moreover, the average beliefs
of the two groups about FDS characteristics were converging after four months (in the second follow-up
survey) relative to the baseline, reverting away from the true population parameters. This means that the
belief updating did not completely reflect Bayes’s law, indicating confirmation bias – i.e., belief updat-
ing in the direction of their original beliefs – at work. In other words, the students in Group T did not
fully accommodate the information given to them to override their priors in the first place, and the effects
dissipated with time despite their additional months of job search.
To provide further evidence on heuristics, we examined the reasons why students did not search for
information about each of the seven sectors if they did not. As presented in Appendix Table A.6, a sig-
nificantly smaller number of students in Group T chose wages and nonwage amenities as their reason for
not searching in MDS. Instead, the proportion of those who said that MDS jobs did not fit their aptitude
or interest increased by 5% points, while the proportions citing other reasons stayed the same. Although
this result has limitations from sample selection, it implies that the female students wereanchored in their
initial judgments on MDS and tried to find other, personal (idiosyncratic) reasons that could offset the
perceived benefits that could be derived from the updated information.
38
38
Although the operation of heuristics, as suggested by Tversky and Kahneman (1974), has been proven to be robust across
economic agents’ decision-making, its implications for labor supply decisions have yet to be understood in a systematic manner.
In their recent study, Conlon et al. (2018) document little evidence that learning in the job market is consistent with Bayesian
updating and suggest that the observed patterns of belief updating can be better explained in terms of “representative” heuristics.
Their results also suggest that the welfare implications of information frictions can be exacerbated when heuristics are at work.
32
1.6 Conclusion
This paper studies how information frictions in college students’ job search affect sectoral segregation
of female students. We examined this issue by administering a survey experiment specifically designed
to (i) assess students’ biased beliefs on sector characteristics and (ii) randomly provide the correct sector
statistics to the treatment group. Then, we followed how their beliefs, aspirations, and job search behaviors
changed during the crucial four months of transition from school to work.
Our main finding is that there exists large and systematic bias in favor of female-dominated sectors
in the beliefs of senior female college students. Providing correct information was effective for drawing
their beliefs closer to reality and boosting their aspirations to enter male-dominated sectors, and the effect
persisted throughout the job search process. We find evidence that the treated students allocated relatively
more search efforts toward jobs in male-dominated sectors and applied for more job openings in those
sectors.
We believe that our result has important policy implications for alleviating occupational segregation
from the labor supply side. If women’s skewed labor supply to a few female-dominated service sectors were
attributable only to gender differences in preferences, policy remedies for segregation should be focused on
understanding the fundamental drivers of those preference gaps. On the other hand, our results indicate
that there is an excessive concentration of labor supply in female-dominated sectors that can be solved
effectively with a low-cost information intervention.
33
Chapter2
ChangingGenderNormsandHouseholdResourceAllocation
2.1 Introduction
Gender norms have been understood as one important source of gendered household resource allocation.
Researchers reported how gender norms favoring the role of men as breadwinners resulted in a skewed
distribution of housework, as well as preventing wives from participating in the labor market. In most of
the literature, gender norms were considered fixed parameters of society because of their strong persistence
over time. Since norms are “not just attitudes in individual’s heads, but also embedded in social institutions,
including education, industry, and the public realm (Pearse and Connell, 2016),” their rapid changes are
rare. Only recently has the literature begun to shed light on examples of changing social norms and their
economic impact (Bursztyn, Egorov, and Fiorin, 2020).
Within this context, this article captures a rare occasion of a society-wide change in gender norms
prompted by a steep increase in the media coverage of feminism in Korea. Drawing on a set of empirical
analyses, we try to answer two key questions. (i) How does the media affect women’s perceptions of gender
norms? (ii) How do the changes in women’s perceptions of gender norms influence household resource
allocation and welfare?
This article begins by describing an explosive expansion of newspaper articles that cover feminism-
related issues in Korea. Until recently, Korea was known as a typical example of an East Asian society
34
with severely male-favoring gender norms, where male dominance, son preference, and missing girls were
prevalent (Das Gupta et al., 2003; Raymo et al., 2015). The sparse media coverage of feminism despite the
long history of the Korean feminist movement could also illustrate the traditional gender norms. However,
feminism suddenly became one of the most important keywords of the media in the mid-2010s after a series
of socially influential incidents, including the phenomenal success of a national-bestseller book against tra-
ditional gender roles (“Kim Ji-young, Born 1982”), large-scale protests against unfair police investigations
of female suspects, sex scandals and the Me Too movement involving the most popular politicians, and
appalling murder cases inspired by misogyny. Newspapers began to write about issues related to women’s
rights and social movements in support of gender equality and an egalitarian society at an unprecedented
speed and intensity.
To capture this moment of explosion in newspapers’ coverage of feminism, we collected all articles
that mentioned the word feminism at least once from the past ten years’ archives of the top six national
newspapers. Then we document that the increase in coverage was synchronized with precipitous changes
in women’s perceptions of gender norms in the late 2010s. We construct a measure of women’s perceptions
of gender norms (“Gender Norms Perceptions (GNP) score”) by averaging the responses to nine survey
questions in the Korean Longitudinal Survey of Women and Families (KLoWF). Both newspapers’ coverage
of feminism and the GNP scores showed a trend-deviant increase since 2015.
To separate the causal impact of newspapers on women’s perceptions of gender norms from a nation-
ally common trend toward egalitarian norms, we estimate a two-way fixed effects model using region-year
variations of women’s exposure to feminism-related newspaper articles. Furthermore, we overcome the
potential bias from the readers’ newspaper selection by exploiting two interesting features of the Ko-
rean newspaper market. First, conservative and liberal newspapers have shown systematically different
approaches to the issues related to feminism. For example, conservative newspapers have been writing
about feminism much more slowly and less frequently than their liberal counterparts. Second, partly due
35
to the historical regional divide, the market shares of the newspapers vary substantially by region. These
features allow us to extract external variations from exposure and construct a Bartik-style measure. More
specifically, we fix the regional market shares at a baseline year and draw variations from the national
growth of each newspaper’s feminism-related articles, which is free from the influence of local demand
for the newspaper. Regressions of this Bartik exposure on the GNP scores provide evidence that in the re-
gions where women were exposed to newspapers with faster-growing coverage of feminism, they showed
greater changes in perceptions of gender norms.
Drawing on this relationship, we turn to the second question: how did the changes in women’s per-
ceptions of gender norms influence household resource allocation and welfare? Instrumenting the GNP
scores by Bartik exposure, we estimate the 2SLS coefficients on a set of outcome variables, including time
use, labor market engagement, household expenditures, frequency of shared activities, and marital hap-
piness. The results show a substantial reduction in household labor among the wives whose perceptions
were more substantially changed by exposure to feminism-related newspaper articles. The wives’ time
use in household labor decreased significantly by 10 hours per week with a 1 standard deviation increase
in the GNP scores. On the other hand, the husbands’ household labor increased by about one hour and
with greater noise. Although husbands’ share of household labor increased as a result, the substantially
gendered allocation remained, and the wives were less likely to be satisfied with the time allocation. The
results were almost identical when we restricted our sample to dual-income households.
Further analyses of other outcome variables provide a bird’s-eye view of the changes in resource al-
location and their welfare implications. First of all, although we could not find a statistically significant
impact on the likelihood of the wives’ labor force participation, the weekly working hours of the employed
wives increased. In addition, by zooming in on the items of household expenditures, we could identify
substantial changes in their relative composition. While the total amount of expenditures remained the
same, the households with more influenced wives showed a substantial increase in expenditures on food
36
and children’s education. Considering that preparing food and taking care of children are heavily time-
consuming household tasks, increased spending on those items indicates that wives began to outsource
excessive household labor to the market. Finally, the boosted egalitarian perceptions were associated with
the increased frequency of shared activities with their husbands, especially in cultural events, exercise,
and family occasions on the wife’s side.
Taken together, this set of evidence implies that although women could not completely shift their ex-
cessive burden of household labor to their husbands even after changing their attitudes, they could achieve
different bargaining outcomes where they spent more time in their workplaces and more favorable activ-
ities. Husbands were not willing to bear the psychological costs of doing household labor that had been
considered women’s work but did accommodate the disutility from a changed consumption set. As a result,
the women reported a higher likelihood of feeling happy about their marriages in the new equilibrium.
This article contributes to three bodies of literature. The first is research on the relationship between
gender norms and the allocation of household labor between spouses. Following Akerlof and Kranton’s
(2000) idea, researchers have reported a close association between them in diverse societies (Álvarez and
Miles, 2003; Cunningham, 2008; Procher, Ritter, and Vance, 2018). This paper captures a rare instance of a
rapid change in gender norms that helps overcome endogeneity and enables a causal identification, which
is not plausible in static analyses. While efforts to measure the causal impact of gender norms – mostly
with the instrumental variables approach – have been made about women’s labor market engagement
(Dildar, 2015; Fortin, 2015; Cavapozzi, Francesconi, and Nicoletti, 2021), that about within-household labor
allocation have been sparse.
Secondly, it complements the literature on how gender norms are generated and evolve. For their
origins, researchers have suggested several vital factors, including contraceptive innovations (Goldin and
Katz, 2002), family and cultural background (Fernandez and Fogli, 2009; Brenøe, 2018), and education
(Dasgupta and Asgari, 2004). This research is in line with those focusing on the influence of the provision
37
of information (Bursztyn, González, and Yanagizawa-Drott, 2020), particularly from the media (Jensen and
Oster, 2009; Bassi and Rasul, 2017). Our findings are consistent with the findings of the literature in that
the media exerted a substantial influence on women’s perceptions of gender norms and induced actual
behavioral changes with economic implications.
Finally, it provides a novel perspective regarding the bargaining approach to household decision-
making (Ashraf, Field, and Lee, 2014; Doepke and Kindermann, 2019). We inherit Bina Agarwal’s (1997)
suggestion that the bargaining models should consider the potential roles of norms and perceptions. Our
results not only illustrate the implications of psychological costs induced by gender norms during the bar-
gaining process but also provide evidence of how changing gender norms can affect the outcomes of the
process.
The remainder of this article is organized as follows. Section 2.2 introduces the background and the
data. Section 2.3 examines the impact of exposure to feminism-related articles on women’s perceptions
of gender norms. Section 2.4 studies the impact of changed perceptions on household resource allocation
and welfare. Section 2.5 provides robustness checks. Section 2.6 concludes the article.
2.2 BackgroundandData
2.2.1 KoreanNewspapers’CoverageofFeminism
Since most Koreans read newspapers regularly,
1
newspapers demonstrate existing social norms and play
a crucial role in generating and altering them. This article focuses on the top six national newspapers –
Chosun, Joongang, Donga, Maeil Kyungje (hereafter Maekyung), Kyunghyang, and Hankyoreh – which a
1
According to the Media Audience Survey 2018, the integrated readership rate of newspapers is 79.6%. (The integrated
readership rate is defined as the proportion of the respondents who read newspaper articles via five media channels (paper, PC
internet, mobile internet, cellphone, and internet protocol TV)).
38
total of 66.3% of respondents to the Media Audience Survey 2018
2
identified as their primary source of
news articles.
Figure 2.1: Newspaper coverage of feminism and women’s perceptions of gender norms
(a) The number of feminism-related articles (b) The trajectory of the GNP scores
Notes: Panel (a) shows the number of articles mentioning the word feminism (the Korean loanword) at least once in the top six
newspapers from 2010 to 2019. The data were extracted from the digitized database on each newspaper’s website. Evernote web
clipper was used to refine the raw articles. Panel (b) shows the trajectory of the GNP scores (a smaller number represents more
traditional perceptions of gender norms, as defined in the following subsection) of women in our sample.
The left panel of Figure 2.1 shows that there were substantial changes in the number of articles men-
tioning the word feminism (hereafter “feminism-related articles”) produced by the six newspapers in the
past decade.
3
The number of feminism-related articles per year had been stable at around 30 for all six
newspapers from the 2000s to 2014.
4,5
However, newspapers began to substantially expand their coverage
of feminism in 2015. In 2018, the total number of these articles in the top six newspapers was 1,454, 11
times larger than in 2014. Also notable was a clear divergence in the acceleration of feminism-related
2
The survey is a nationally representative survey on the behaviors of media audience, conducted annually by Korea Press
Foundation. In 2018, 5,040 adult samples were selected by the stratified randomization process considering the population by
region. Respondents were interviewed by computer-aided personal interview.
3
The Korean vocabulary that we searched is the loanword for feminism. We also examined other similar words – (Korean
words for) feminist, gender, gender discrimination, and women’s right – and found no notable differences in the trends.
4
This indicates that the mainstream media’s approach to feminism was closer to ignorance than vilification , as typical in
countries where feminism has yet to flourish fully. Lind and Salo (2002) pointed out that the US media’s coverage of feminism
and academic research on the subject had also been sparse until the late 20th century.
5
To put this number into perspective, these newspapers’ average numbers of articles mentioning a word were 10,359 for
government, 5,186 for industry, 7,997 for culture, and 10,234 for society, respectively, in the same period. On the other hand, The
New York Times’ average numbers of articles mentioning those words were 906 for feminism, 28,634 for government, 13,455 for
industry, 11,628 for culture, and 8,122 for society, respectively.
39
articles between liberal (Hankyoreh and Kyunghyang) and conservative newspapers (Chosun, Joongang,
Donga, and Maekyung). While liberal newspapers led a social discourse on feminism, conservative ones
were slower to follow the issue. For example, the former doubled their feminism-related articles in 2015
compared to 2014, while the latter did so in 2017.
In the Appendix, we provide an in-depth analysis of the contents of those feminism-related articles.
Liberal newspapers reported on feminism-related issues in more diverse aspects and from a more support-
ive stance than conservative ones. For example, liberal newspapers covered more pro-feminist movements,
politics, and education issues, relatively more in social and op-ed sections. On the other hand, conservative
newspapers wrote more about celebrity scandals and controversies around feminist movements, more in
a culture section. Newspapers were different not only in their number of articles on feminism but also in
how they wrote about the issue.
2.2.2 Women’sPerceptionsofGenderNorms
Biennial panel data from the nationally representative Korean Longitudinal Survey of Women and Families
(KLoWF) were used to evaluate married women’s perceptions of gender norms. The first wave began in
2007 with a sample of 9,068 households and 9,997 women aged between 19 and 64. They were followed
from 2008 (Wave 2) to 2018 (Wave 7). Since this article focuses on the changes in social norms from the mid-
2010s, I mainly analyze Waves 3 to 7 (years 2010, 2012, 2014, 2016, and 2018), focusing on the latter three
waves. The sample consists of only women. Thus, responses about other household members’ behaviors
(e.g., husband, children) are reported by the interviewed women. The household-level effective attrition
rate was 28.7% in Wave 5 and 29.5% in Wave 7.
6
To assess Korean women’s perceptions of gender norms, we use the responses to questions in the
family-related attitudes section of the survey. This section includes nine questions (Table 2.1) asking
6
Further details are provided in the KLoWF User’s Guide.
(URL: https://klowf.kwdi.re.kr/portal/data/dataListPage.do?bbsId=ETCDATA)
40
Table 2.1: Components of the GNP Score
Category Original survey statements
Family Q1. Marriage is indispensable.
Q2. Having a child is indispensable.
Q3. I can get divorced even if I have a child.
Q4. My personal achievement is more important than marriage.
Q5. Marriage restricts my personal life.
Gender roles Q6. It is ideal that a husband works outside, and a wife takes care of the family.
Q7. A wife’s job helps build an equal relationship between a husband and wife.
Q8. A wife’s job is detrimental to children under school age.
Q9. A dual-income couple should be in charge of equal amounts of household labor.
how strongly each woman agrees with statements related to attitudes toward traditional gender norms
in Korean society – women are considered desirable when they get married, have children, and prioritize
maternity over career. The first five questions measure women’s attitudes toward family formation and
dissolution and their sense of identity in the family. The other four questions measure how women per-
ceive gender roles within and outside the household. We construct an index by taking an average of those
nine questions with equal weights after recoding each, such that 1 represents “strongly traditional” and 4
represents “strongly egalitarian”. This index (hereafter “Gender Norms Perceptions (GNP) score”) will be
used as a proxy for women’s perceptions of gender norms throughout this article.
7,8
The right panel of Figure 2.1 shows the trajectory of the GNP scores. We can observe a trend-deviant
escape from traditional attitudes in Wave 6 (2016), which coincides with the explosive increase in news-
paper coverage of feminism from the mid-2010s, as described in the left panel. Note that the sample is
7
Many researchers have similarly used survey responses (e.g., World Values Survey, European Social Survey, and International
Social Survey Programme, etc.) as a primary measure of women’s perceptions of gender norms. In the Appendix, we compare
the questions used in the literature with ours. The comparison shows that our questions are coherent with other researchers’
measures, except that they are focused more on the domestic sphere (family-related attitudes) rather than the public sphere
(political/societal) attitudes.
8
Note that this linear scale approach has limitations on oversimplifying a multi-dimensional spectrum of gender role attitudes.
For example, Knight and Brinton (2017) identified three different types of egalitarianism rising across postindustrial European
countries and found no convergence toward one dominant type. Considering this, in our analysis of the media’s impact on
perceptions of gender norms (Section 2.3.3), we zoom in on the GNP score and evaluate the effects on each component of the
index. The Cronbach’s alpha of the index is 0.6, and we report the pairwise correlation coefficients between the components in
the Appendix.
41
fixed throughout the period, which means that women began to alter their attitudes from traditional to
egalitarian, even when they were getting older.
2.2.3 HouseholdResourceAllocationandOtherOutcomesvariables
The KLoWF dataset provides rich ingredients for analyzing resource allocation within and outside the
household, as well as its implications for women’s welfare. As for allocation, our primary outcome variable
is each spouse’s time use in household labor. The survey records time spent on housework (e.g., cooking,
dishwashing, etc.) and childcare in minutes on each day of the week (weekdays, Saturdays, and Sundays).
We add up housework and childcare to construct one outcome variable, time use in household labor.
In addition, the survey provides women’s labor market status, detailed responses about household
expenditures in various categories, and the frequency of women’s daily activities, including household
labor and shared activities with husbands. Furthermore, the responses about satisfaction with household
labor allocation and women’s overall marital happiness provide measures of welfare. Table 2.2 details how
each outcome variable was constructed.
2.2.4 SampleandAttrition
Since our focus is on resource allocation between married couples, the sample of our main analysis is 4,422
women who stayed in marriage from Wave 5 (2014) to Wave 7 (2018). Cohabitation and out-of-wedlock
childbirth are relatively scarce in Korean society (Raymo et al., 2015); thus, there are limited concerns about
the external validity in terms of restricting the sample to married couples. Out of 5,158 married women
in Wave 5, 287 (5.6%) and 486 (9.4%) women did not respond to the surveys in Wave 6 and 7, respectively.
Among those who responded, 104 (2.1%) and 179 (3.8%) were excluded from the sample for each wave due
to divorce, separation, or the husband’s death. In the Appendix, we provide a selective attrition analysis
with observable characteristics. There were no variables correlated with attrition in Wave 6. In Wave 7,
42
the wife’s age and the husband’s work status were correlated with attrition at 10% significance. In Section
2.5, we check whether our treatment (changes in perceptions induced by exposure to the media) had a
differential impact on attrition.
Table 2.2: Construction of Outcome Variables
Category Variables Variable format Original survey questions
Household
labor
Time spent Continuous (mins/week)
allocation Husband’s share Continuous (0∼ 1)
Wife satisfied
with allocation
Binary
(=1 if response > med.)
How much are you satisfied with your
husband’s current share of
housework/childcare? (Likert (1∼ 5))
Wife’s labor
market
Labor force
participation
Binary
(=1 if participating)
engagement Weekly work hours Continuous
Expenditure
/Activities
Household
expenditure
Continuous
(monthly, 1,000 USD)
Frequency of shared
activities
Continuous
(num. of occasions/month)
How often did you do the following activities
with your husband last month? (Categorical
(e.g., once a week))
Welfare Wife’s marital
happiness
Binary
(=1 if response > med.)
How do you describe your feelings for your
marriage recently? (Likert (1∼ 10))
Notes: The household expenditures were inflation-adjusted by Consumer Price Index (2015 = 100) and converted to USD (1 USD
= 1,200 KRW). The welfare variables (wife satisfied with the allocation (0/1) and wife’s marital happiness (0/1)) were constructed
as binary variables that take value 1 if the original responses were greater than the median in the sample.
2.3 TheImpactofFeminism-relatedArticlesonPerceptions
This article tries to identify two key relations. The first is the impact of an increase in feminism-related
articles on women’s perceptions of gender norms (first-stage relation). The second is the changes in house-
hold resource allocation caused by altered attitudes of women (second-stage relation). In this section, we
look into the first-stage relation.
43
2.3.1 ExposuretoFeminism-relatedArticles
The coincidence of the changes in the number of feminism-related newspaper articles and women’s percep-
tions of gender norms shown in Figure 2.1 might be attributable to (unexplained) common trends toward
an egalitarian society, not necessarily indicating the influence of the former on the latter. One way to
distinguish the newspapers’ impact on the perceptions is to exploit region-year variations. More specif-
ically, the identifying assumption is that if the changes in the perceptions had been attributable only to
nationally common trends, all regions would have demonstrated similar changes. Conversely, if regions
with stronger exposure to feminism-related articles showed more substantial changes in the perceptions,
it is more likely to be interpreted as newspapers’ differential impact by region.
Two characteristics of the Korean newspaper market contribute to this strategy. First, the market
shares of six national newspapers account for 60∼ 70% of the total. Second, due to the historical re-
gional divide, the market share of each of the six newspapers varies substantially by region.
9
While
south-western regions (Chungcheong and Joella) show a strong readership of liberal newspapers (Han-
kyoreh and Kyunghyang), conservative newspapers (Chosun and Kyunghyang) have higher market shares
in south-eastern regions (Gyeongbuk andGyeongnam) throughout the analyzed period. Metropolitan areas
(Seoul and Gyeongin) are located in the middle of the ideological spectrum of the newspaper readership.
10
We define the (average) exposure to feminism-related articles in regionj in yeart as:
exposure
jt
=
X
k
(N
kt
× S
kjt
)× F
jt
(2.1)
where N
kt
is the number of feminism-related articles that newspaper k published in year t, S
kjt
is the
market share of newspaper k in region j in year t, and F
jt
is the average frequency of people reading
9
The market shares computed from the Media Audience Survey 2018 is provided in the Appendix.
10
For deeper understanding of the origin and implications of Korea’s historical regional divide, see Kwon (2004).
44
newspapers in regionj in yeart.
11
That is, exposure in a region is defined as the market share-weighted
number of feminism-related articles multiplied by the integrated newspaper readership rate.
Figure 2.2: The Relationship between the Changes in Exposure and the GNP Scores
(a) Exposure
(b) Average (predicted) GNP scores
Notes: Each woman’s GNP score is predicted after regressing the standardized GNP score on individual-level control variables
used in the main regressions. An average for each region is computed using the sampling weights. Colors are rescaled so that
the highest value is displayed as 100, while the lowest value is displayed as 0. The number of each region: 1-Seoul, 2-Gyeongin,
3-Choongchung, 4-Jeolla, 5-Gyeongbuk, 6-Gyeongnam, 7-Gangwon, and 8-Jeju.
Figure 2.2 graphically displays a close relationship between exposure and the GNP scores. From Wave
5 to Wave 7, regions that experienced more significant growth in exposure showed a deeper transition to
egalitarian attitudes among women. In 2016 (Wave 6), the metropolitan areas (Seoul and Gyeongin) and
Jeolla showed the highestexposure, mainly due to the relatively large market shares of liberal newspapers
11
The frequency of the newspaper readership by region was also retrieved from the Media Audience Survey 2018.
45
in those regions, resulting in the biggest increase in the average GNP score. Note that the gap in the number
of feminism-related articles between liberal and conservative newspapers was substantially larger in 2016
since the latter had yet to expand their coverage of feminism. On the other hand, in 2018 (Wave 7), the
gap became narrower andexposure was more likely to be determined by the average readership rate of the
region than the differences in market shares (the coverage of feminism was similar among newspapers),
which brought about a bigger increase in the average GNP scores in the Chungcheong and Gyeongnam
regions.
2.3.2 EmpiricalStrategy
One remaining concern in claiming causality in the relation is that exposure is not randomly assigned. In
other words, since market shares of newspapers in regionj are an aggregate result of choices about which
newspapers to read by people in that region, there might exist region-specific omitted variables that could
affect both the market shares of newspapers and the perceptions differentially by region. We overcome this
problem by extracting a purely external portion of the variations fromexposure by fixing each newspaper’s
regional market share – the result of a selection – at a baseline year. For this purpose, we adopt the idea
of Bartik’s shift-share instrumental variables.
Bartik’s canonical instrument is calculated by interacting local industry shares and national industry
growth (Bartik, 1991). For example, when local industry shares are fixed at baseline, the national growth
of each industry multiplied by the fixed local shares can provide externally determined variations in local
labor demand that are not affected by idiosyncratic shocks in local labor supply. Within the context of this
paper, the shift-share instrument can be constructed with newspapers’ market shares in each region and
their national growth in the number of feminism-related articles. After fixing the market share of each
newspaperk in regionj at a baseline year, the instrument for regionj in yeart is defined as:
46
Bartik
jt
=
X
k
(G
kt
× S
kj0
) (2.2)
whereG
kt
is the growth in the number of feminism-related articles that newspaperk published nationally
in yeart andS
kj0
is the market share of newspaperk in regionj in a baseline year (0). This instrument
provides a measure of exposure to feminism-related newspaper articles, which is not affected by people’s
newspaper choices in that region since market shares – the results of people’s choices – are fixed. There-
fore, the identification comes from the external variations of the newspapers’ market shares in each region
in that those regions with different initial market shares would have different exposure to the national
growth in feminism-related articles.
We henceforth call this variable “Bartik exposure,” indicating that it essentially extracts the external
portion of variations from exposure. We adopt Bartik exposure in a two-way fixed effects estimation to
argue that exposure to feminism-related articles has a causal impact on women’s perceptions of gender
norms (first-stage relation). Then we use it as an instrument for the GNP scores in the 2SLS regressions
(second-stage relation) that will be introduced in the next section. To calculateBartikexposure, we use the
number of feminism-related articles in the six national newspapers and their market shares in each region,
with 2013 as a baseline year.
Using Bartik exposure, we characterize the first-stage equation as:
GNP
ijt
=µ +Γ ij
+Υ t
+Λ ∗ Bartik
jt
+X
′
ijt
Θ +η ijt
(2.3)
where GNP
ijt
is the standardized GNP score of a woman i in region j in year t, Γ ij
is individual fixed
effects, Υ t
is year fixed effects, Bartik
jt
is Bartik exposure, andX
ijt
is control variables for time-variant
household-level characteristics. While individual fixed effects absorb all the time-invariant characteristics,
year fixed effects adjust for time-specific confounders. In addition, we control for potentially time-variant
47
observables (X
ijt
) – the husband’s and the wife’s age, their years of education, the number of children
under 19, a dummy for the husband’s employment status, and the annual household income. The KLoWF
data for 4,422 married women in Waves 5, 6, and 7 (years 2014, 2016, and 2018), which are posterior to
the baseline year, are used. Regressions are weighted by the sampling weights, and standard errors are
clustered at the individual level.
Note that in this model, the identification comes from the within-region variations across years, not
from differences between regions. Therefore, the differential level of average perceptions of gender norms
among regions at baseline does not affect the validity of the estimation. The estimated coefficient of interest
(
ˆ
Λ ) can be interpreted as the Intent-To-Treat (ITT) effect since the treatment status depends on the region
where each woman lives rather than whether she actually read the articles.
2.3.3 Results
Table 2.3: The First-stage Results
GNP score (Std.)
(1) (2)
Exposure (exposure
jt
) 0.0034***
(0.0007)
Bartik Exposure (Bartik
jt
) 0.0775***
(0.0148)
Individual F.E. Yes Yes
Controls Yes Yes
R
2
0.092 0.043
Num. of Obs. 13,230 13,230
Notes: *** p < 0.01, ** p < 0.05, * p < 0.1. OLS regression coefficients. Standard errors clustered at the individual level in
parentheses. Sample: 4,422 married couples (Observations with missing time use responses are omitted for consistency with the
second-stage regressions.)
Table 2.3 displays the results of the estimated first-stage relation. Column (1) shows the significant and
positive coefficient of exposure on the GNP scores. Although this number does not entirely overcome se-
lection bias, it helps us understand the approximate magnitude of the treatment effect, considering the
48
newspapers’ relatively stable regional market shares throughout the period. The magnitude of the esti-
mated coefficient, 0.0034, means that exposure to approximately 100 (weighted) feminism-related articles
was associated with substantial 0.34 standard deviations increase in the GNP score, significantly at the 1%
level. Considering that the average number of feminism-related articles in the six newspapers grew by 221
from 2014 to 2018, the average impact throughout this period can be estimated to be approximately 0.75
standard deviations.
Column (2) shows a positively significant coefficient for Bartik exposure after potential bias from the
newspaper selection is eliminated. Although it is difficult to directly compare the magnitude to that in Col-
umn (1) due to the instrument’s structure, this result confirms that women’s perceptions were significantly
affected by how the newspapers they read write about feminism.
Figure 2.3: Estimated Effects of Bartik exposure on Components of the GNP Score
Notes: This figure plots estimated coefficients for Bartikexposure with the components of GNP score as dependent variables from
the same regression specifications and the sample as Table 2.3. The questions Q1 ∼ Q9 are provided in Table 2.1. “Family” index
is a standardized aggregate index of the responses to Q1∼ Q5. “Gender roles” index is that of the responses to Q6∼ Q9. The lines
stand for 95% confidence intervals.
To verify whether the estimated effects demonstrate the changes in perceptions of gender norms in
various dimensions that might not be captured as an aggregated index, we ran the same regressions with
49
the components of the GNP score. We not only ran the regressions with each of the nine components but
also tested with the aggregated subcomponents of the attitudes towards family (Q1∼ Q5) and gender roles
(Q6∼ Q9), respectively. Figure C.5 shows that the estimated effects for attitudes toward family and gender
roles were similar in magnitude and both statistically significant. All nine components showed positive
coefficients, albeit with varying statistical significance.
2.4 TheImpactoftheChangesinPerceptionsonResourceAllocation
Based on the estimated impact of media exposure on women’s perceptions of gender norms, we examine
how these changes affected resource allocation and welfare in this section. Before introducing our empir-
ical strategy on this second-stage problem, we discuss the endogeneity problem in identifying the causal
relationship.
2.4.1 PreviousLiterature
The formal economic conceptualization of gender identity and its impact on household resource allocation
was suggested by Akerlof and Kranton’s (2000) seminal work. They focused on the asymmetry in the
observed relationship between each spouse’s share of outside work hours and household labor. Husbands
did not increase their share of household labor to higher than a certain level, even when the wives were
specialized enough to account for all of the household income. This phenomenon can be explained in
their framework by incorporating an identity component into both spouses’ utility functions. In other
words, the husband’s and the wife’s welfare are not only determined by resource allocation but also by
how this allocation interacts with their identity (or self-perceptions), which is influenced by gender norms
of society.
Inspired by their theory, economists have accumulated empirical evidence on the relationship between
gender norms and excessively disproportionate resource allocation by gender in both the public and the
50
domestic spheres, which cannot be fully understood within the Beckerian labor specialization perspective.
For example, traditional gender norms distinguishing the role of men and women prevent wives from
participating in the labor market (Dildar, 2015; Bursztyn, Fujiwara, and Pallais, 2017) and earning more
than their husbands (Bertrand, Kamenica, and Pan, 2015), even when wives have higher potential earnings.
Closer to the central theme of this paper, implications of gender norms for a disproportionate distri-
bution of household labor have been reported in diverse societal backgrounds. From cross-country per-
spectives, researchers have established that societies with more egalitarian gender norms demonstrate a
relatively equal distribution of household labor (Fuwa, 2004; Aassve, Fuochi, and Mencarini, 2014; Fahlén,
2016; Campaña, Giménez-Nadal, and Molina, 2018). On the other hand, the results from micro-level house-
hold surveys show substantial heterogeneity and a weaker level of statistical significance (Parkman, 2004;
Cunningham, 2005).
12
The mixed microeconomic findings are partly ascribable to the fact that most research has not ad-
dressed the endogeneity in the relationship. Agarwal (1997) is one of the first to raise this question. Not
only can gender norms affect the bargaining process between spouses, but “norms themselves can be the
subjects of bargaining.” She argued that the researchers should consider “the interactive nature of bargain-
ing in effectively challenging social norms.”
Within this context, estimating the causal impact of a woman’s perceptions of gender norms on house-
hold resource allocation is difficult because of the sparsity of external sources of variations that affect
bargaining outcomes only through the gender norms channel.
13
This article tries to overcome this prob-
lem by exploiting the external portion of variations in media exposure and instrumenting the endogenous
perceptions by Bartik exposure, as established in the previous section.
12
For example, Greenstein (1996) found that the husband’s gender ideology was significantly associated with the division of
labor only when he was married to a woman with an egalitarian ideology. Carriero and Todesco (2018) provided evidence that
the magnitude of the estimated relationship between those two variables varied by women’s level of education and housework
of women without college degrees was not affected by gender ideology.
13
One strategy that has been tried within this context is to use parents’ gender role attitudes as a proxy for gender norms.
For example, Hwang et al.’s (2019) analysis of Korean dual-earner couples showed that the wife’s household labor is significantly
greater when the husband is from a region with a higher sex ratio at birth.
51
Table 2.4: Descriptive Statistics (Averages) at Baseline (Wave 5)
GNP
< median
GNP
≥ median
p-value
(1) (2) (3)
Panel A. Household observable characteristics
Wife’s age 51.6 46.5 0.000
Wife’s years of education 11.2 12.7 0.000
Husband’s age 54.9 49.5 0.000
Husband’s years of education 12.4 13.3 0.000
Husband employed (0/1) 0.79 0.85 0.000
Number of children under 19 0.79 1.14 0.000
Household income (monthly, 1,000 USD) 2.93 3.45 0.000
Panel B. Outcome variables
Husband’s share of household labor 0.12 0.12 0.261
Wife’s time spent in household labor (mins/week) 1,347 1,466 0.018
Husband’s time spent in household labor(mins/week) 194 236 0.018
Wife satisfied with household labor allocation (0/1) 0.57 0.53 0.056
Wife’s labor market participation (0/1) 0.48 0.56 0.000
Wife’s work hours in workplace (weekly, conditional) 41.3 42.5 0.216
Household expenditure (monthly, 1,000 USD) 2.04 2.36 0.000
Wife’s marital happiness (0/1) 0.37 0.37 0.897
Num. of Obs. 2,098 2,324
Notes: This table compares the mean observable characteristics and outcome variables of 4,422 married couples according to
whether a woman’s GNP score is above or below the median in Wave 5. Column (3) reportst-testp-values. Statistics are weighted
by sampling weights.
2.4.2 DescriptiveStatistics
We first report descriptive statistics of our sample in relation to the GNP scores at baseline. Panel A
of Table 2.4 reports the average observable characteristics of the wives and the husbands that strongly
correlate with the wives’ GNP scores. More egalitarian women were younger and more educated, and
their households earned more. This is consistent with the literature and again confirms the viability of our
GNP score as a measure of perceptions of gender norms.
On the other hand, the correlation between the GNP scores and household resource allocation is less
clear. First of all, it is notable that the husbands’ share of household labor was equivalent in both types of
households. Both wives and husbands spent more time in household labor in the households with more
52
egalitarian women because they were younger and more likely to have children under 19. Despite the
same allocation, women with egalitarian perceptions were significantly less likely to be satisfied with it.
Egalitarian women were more engaged in the labor market and reported more household expenditures,
but there were no distinguishable differences in terms of marital happiness between the two types.
The descriptive statistics indicate that an econometrician should not only control for individual char-
acteristics at baseline but also overcome endogeneity that cannot be properly addressed by simple corre-
lations. In the next section, we elaborate on our empirical strategy.
2.4.3 EmpiricalStrategy
Based on the influence of newspaper articles on women’s perceptions of gender norms estimated in the first
stage, we examine how these changes affected household resource allocation and welfare. To overcome
endogeneity, we estimate the fixed-effects two-stage least squares (FE2SLS) model using Bartik exposure
as an instrument for the GNP scores instead of OLS specifications. The FE2SLS model is characterized as:
y
ijt
=α +λ ij
+γ t
+β ∗ \
GNP
ijt
+X
′
ijt
θ +ϵ ijt
(2.4)
wherey
ijt
stands for outcome variables of a womani in regionj in yeart,λ ij
is individual fixed effects,
γ t
is year fixed effects,
\
GNP
ijt
is predicted perceptions of gender norms in Equation (3), andX
ijt
is the
same household-level control variables used for the first-stage regressions. The inclusion of individual
fixed effects allows us to compare changes in outcome variables within each household across years and
alleviate concerns with regard to the different levels of those variables across households at baseline.
14
The
outcome variables are time use in household labor by spouses and the wife’s satisfaction with it, the wife’s
labor market engagement, household expenditures, the frequency of shared activities with the husband,
14
One remaining concern after including both individual and year fixed effects is that there might exist time-variant regional
factors that are correlated with the error term if women in the sample changed their regions of residence during the period of
analysis. However, the samples with internal migrations were around 1% (50 households in Wave 6 and 42 households in Wave 7).
The inclusion of region fixed effects did not affect the results of our main regressions (The results are provided in the Appendix).
53
and a binary indicator of marital happiness.
15
In the first-stage regression, the t-statistic ofBartikexposure
is 5.24, and theF -statistic of the estimated model is 25.9; thus, a concern for a weak IV problem in using
this instrument for the following 2SLS regressions is limited.
In this specification, the estimated coefficients are interpreted as the local average treatment effect
(LATE). More specifically, we estimate the impact of the changes in the perceptions inducedbyexposure to
feminism-related articles. In other words, the LATE is an average among those women whose perceptions
of gender norms were affected by exposure (compliers), and it excludes those who were not affected in
the first-stage (never-takers). In addition, the average is weighted by how much the first-stage effects
influenced the compliers’ perceptions. This instrumental variables approach is useful to extract the causal
impact from the spurious correlations, as well as having the additional advantage of alleviating concerns
about measurement errors in the GNP scores, as discussed in Section 2.2.2.
16
Although the fixed-shares nature of the instrument supports the exclusion restriction, one limitation
in this model should be noted. It assumes that the only mechanism through which differential exposure
affects resource allocation is the change in women’s perceptions, not men’s. This is due to the limitations
of our data since the survey only asked questions to wives, not husbands.
2.4.4 Results
TimeUseinHouseholdLabor
The first row of Table 2.5 reports the 2SLS estimates on each spouse’s time use in household labor. Column
(1) shows a significant increase in the husbands’ share in time use, with a magnitude of 6 percentage
15
For the construction of variables, see Table 2.2.
16
As explained in the previous section, the Bartik exposure use market shares of the top six newspapers rather than all news-
papers. This might cause the potential incomplete shares problem (Borusyak, Hull, and Jaravel, 2022). In canonical settings of
the Bartik instrument, the sum of the industries’ shares at baseline is assumed to be 1 (in this case,
P
k
S
kj0
should equal 1). When
the sum is neither 1 nor constant across the regions, regions with higher shares of the six newspapers can have systematically
different values of Bartikjt, leading to bias when the unobservables are also different across the regions. Following Borusyak et
al.’s (2022) remedy for this problem, we run IV regressions that include the summed market share (
P
k
S
kj0
) for each region as a
control variable.
54
Table 2.5: 2SLS Estimates on Time Use in Household Labor
All households Dual-income households
Husband’s
share of
time
Wife’s
time spent
(mins/
week)
Husband’s
time spent
(mins/
week)
Wife
satisfied
with
allocation
(0/1)
Husband’s
share of
time
Wife’s
time spent
(mins/
week)
Husband’s
time spent
(mins/
week)
Wife
satisfied
with
allocation
(0/1)
(1) (2) (3) (4) (5) (6) (7) (8)
2SLS
GNP score 0.060** -618.0*** 78.4 -0.261** 0.066 -442.5** -6.6 -0.321**
(Std.) (0.030) (205.0) (69.8) (0.109) (0.044) (222.4) (94.7) (0.161)
OLS
GNP score 0.003* -1.5 3.2 -0.008 -0.000 -6.5 4.2 -0.105**
(Std.) (0.002) (9.6) (3.8) (0.006) (0.003) (10.9) (5.7) (0.050)
Mean Dep. 0.133 1,414.3 217.1 0.841 0.165 1,201.0 237.8 0.870
Indiv. FE Yes Yes Yes Yes Yes Yes Yes Yes
Controls Yes Yes Yes Yes Yes Yes Yes Yes
Num. obs. 13,230 13,230 13,230 13,230 6,817 6,817 6,817 6,817
Notes: ***p< 0.01, **p< 0.05, *p< 0.1. Standard errors clustered at the individual level in parentheses. Sample: 4,422 married
couples (Observations with missing time use responses are omitted). The mean of dependent variables is reported for Wave 5.
For the construction of variables, see Table 2.2. The regressions with a binary variable of satisfaction (Columns (4) and (8)) are
estimated by the linear probability model. All regressions are weighted by sampling weights.
points associated with a 1 standard deviation increase in the (standardized) GNP score. If we look into
each spouse’s weekly time spent (Columns (2) and (3)), the estimated effect was a substantial reduction of
618 minutes (10 hours) per week for the wives. The reduced 618 minutes consisted of 385 minutes spent
on housework (cooking, dishwashing, etc.) and 233 minutes on childcare. On the other hand, the total
amount of time spent by the husbands increased only by 78 minutes and was estimated with noise. The
time spent on housework and childcare increased by 77 minutes and 1 minute, respectively.
17,18
Taken together, this indicates that the impact of the media’s explosive coverage of feminism con-
tributed to a steep reduction in the amount of women’s household labor by influencing their perceptions
17
To put these numbers into perspective, note that the total average GNP score increased by 0.36 standard deviations from
Wave 5 to Wave 7.
18
The allocation of housework had been more severely skewed than that of childcare at baseline (Wave 5). The average time
spent on housework and childcare was 1,088 and 327 minutes for the wives and 140 and 77 minutes for the husbands, respectively.
Considering the relative magnitudes of the reductions compared to the baseline, there was a relatively smaller impact on the
allocation of housework. This result is in line with Sullivan’s (2013) findings that the gendered allocation of routine housework
had the strongest persistence among diverse dimensions of household labor. The regression table is provided in the Appendix.
55
of gender norms. Although the husbands of the affected women began to consider substantial changes
in their wives’ attitudes when bargaining over household resource allocation, the magnitude of their in-
creased participation was not enough to overcome the severely disproportionate allocation. Consequently,
Column (4) shows that the wives who experienced a 1 standard deviation increase in the GNP score were
26% less likely to be satisfied with the allocation despite the improvement in the husbands’ contributions.
Columns (5) to (8) estimate the same 2SLS regressions with restricting the sample to dual-income
households. The magnitude of the estimated impact was similar to those from the whole sample. At
baseline, the wives in dual-income households spent 213 minutes less on household labor. In comparison,
the husbands spent 21 minutes more, resulting in allocation only slightly better (3 percentage points in
terms of the husbands’ share) compared to the average of all households. While the wives with changed
perceptions of gender norms reduced labor by 443 minutes, the husbands showed no changes in time
spent on household labor, resulting in a similar 6.6 percentage points increase in their shares. The wives’
satisfaction was more strongly aggravated.
On the other hand, the OLS results presented in the second row show limited correlations between
the changed perceptions and outcome variables. As discussed in the previous section, this is due to the
following reasons. First, while the 2SLS specification measures the local average treatment effects on
women who actually changed their perceptions after being influenced by feminism-related articles that this
article aims to estimate, the OLS specification measures the average correlation between the two variables
over the whole sample. Second, the measurement error in assessing women’s gender role perceptions by
the GNP scores could have caused an attenuation bias in the OLS estimates. Finally, there could exist a
reasonable simultaneity bias in the relationship. Since the GNP scores include women’s attitudes toward
gender roles, time allocation could also affect the GNP scores. For example, the wives who achieved a
decreased burden in household labor could have less strong opinions about the statements, including “A
dual income couple should be in charge of equal amounts of household labor”. In this case, the OLS
56
estimates could be biased upward and offset the negative impact of the increased GNP scores on the wives’
time spent in household labor.
OtherOutcomeVariables
Although the wives with changed perceptions were not successful in shifting enough burden of housework
to their husbands, they chose to reduce their workload at home anyway. How could they do that without a
corresponding increase in their husbands’ participation? A clearer image is illustrated by examining other
outcome variables, although our data do not provide detailed time use information outside the household.
Table 2.6 summarizes the changes in the outcome variables associated with a 1 standard deviation
increase in the GNP scores induced byexposure to feminism-related articles. Panel (A) shows that boosted
egalitarian perceptions did not bring about a statistically significant increase in the wife’s likelihood of
labor market participation. Still, they did cause an increase in weekly work hours by six hours, conditional
on employment. It means that the estimated effects on labor market engagement were in the intensive
margin and that the influenced working women could divert some of their housework hours into their
workplaces.
We next turn to the household expenditures in Panel (B). Although the total expenses were not affected
significantly, there were substantial composition effects. Among the items of expenses, those for food
(which includes eating at home and eating out) and children’s education increased by 24.5% and 31.9%,
respectively, and those for remaining activities decreased as much. Considering that food preparation and
childcare are heavily time-consuming activities, these results indicate that those women tried to find the
most effective way to outsource the burden of household labor to the market.
We can also find further evidence of the altered time allocation from the reported frequency of shared
activities with the husband in Panel (C). Positive effects were estimated across all shared activities mea-
sured. Particularly, women with a 1 standard deviation increase in the GNP score went to cultural events
57
with their husbands once more per month, with statistical significance at 1%. Moreover, the frequency
of joint exercise and participation in family occasions on the wife’s side also increased at close to 10%
significance ( p=0.178 and0.151, respectively).
Table 2.6: The 2SLS Estimates on Other Outcome Variables
(1) (2) (3) (4) (5)
Panel A: Wife’s labor market engagement
LFP
(0/1)
Weekly
Work
hours
(cond.)
GNP score (Std.) 0.086
(0.062)
6.737**
(3.384)
Panel B: (Log) Household expenditures (1,000 USD, annual)
Total Food Children Others
Edu.
GNP score (Std.) -0.096
(0.059)
0.220***
(0.077)
0.277
(0.200)
-0.323***
(0.100)
Panel C: Frequency of shared activities with the husband (monthly)
Cultural Exercise Social Family Family
Events Activities Occasions Occasions
(Husband’s) (Wife’s)
GNP score (Std.) 1.100***
(0.277)
0.524
(0.389)
0.138
(0.190)
0.410
(0.427)
0.616
(0.429)
Panel D: Wife’s marital happiness (0/1)
GNP score (Std.) 0.141
(0.091)
Individual F.E. Yes
Controls Yes
Num. of obs. 13,230
Notes: *** p < 0.01, **p < 0.05, *p < 0.1. Standard errors clustered at the individual level in parentheses. Sample: 4,422
married couples (Observations with missing time use responses are omitted). For the construction of variables, see Table 2.2. The
regressions with a binary variable of marital happiness is estimated by the linear probability model. All regressions are weighted
by sampling weights. Weekly work hours are conditional on employment. The corresponding OLS results are provided in the
Appendix.
58
Now we turn to the implications for women’s welfare, measured by self-reported marital happiness
in the survey in Panel (D). The estimated 2SLS coefficient means that a 1 standard deviation increase in
the GNP score was associated with 14% more likeliness to be (relatively) happy in marriage, close to 10%
significance ( p = 0.121). This is particularly interesting when viewed together with the previous results.
In sum, women more affected by feminism-related newspaper articles became happier in their marriages,
even when their satisfaction with the household labor allocation deteriorated. Instead of tolerating the
pressure of the excessive housework alone, they began to bargain with their husbands to outsource some
of the work and use that time to engage more actively in their jobs and enjoy more quality time with
the husbands, and achieved better marital happiness in consequence. This result not only suggests that
there might exist another important mechanism (than perceived fairness) through which the dynamics
of the household resource allocation affect marital happiness
19
but also necessitates more theoretical and
quantitative research on the relationship between egalitarian attitudes and marital quality.
20
2.5 RobustnessChecks
Parallelpre-trends(firststage)
The first-stage analysis of this article estimated a panel two-way fixed effects model using regional vari-
ations across time to show that (Bartik) exposure has a causal relationship with the changes in women’s
perceptions of gender norms. It is crucial to check whether there had been no differential trends in the
perceptions among regions before treatment, that is, before the variations of exposure were generated in
the mid-2010s.
19
One consistent findings of the sociology literature from the 1980s to 1990s is that the equal division of household labor is
associated with higher marital quality and less conflict or divorce. Perceptions of fairness or support were suggested as mediators
in the relationship (Pina and Bengtson, 1993; Frisco and Williams, 2003; Fleche, Lepinteur, and Powdthavee, 2020).
20
After Wilson (1982) argued that feminist women are less satisfied with their marriage, the literature focusing on this question
has been sparse, despite the prevalence of the notion that feminism contradicts traditional patriarchal values.
59
Figure 2.4: Trends in the (Standardized) Average GNP Scores across Regions
Notes: Sample: 4,422 married couples used in the main analysis. Each woman’s GNP score is standardized across a pooled sample
of the five waves.
Since our data allow for constructing comparable GNP scores from Wave 3 (2010), we can examine
the pre-trends from Wave 3 (2010) to Wave 5 (2014). Figure 2.4 shows that the trends before 2014 had
been similar across regions, except for Gangwon and Jeju. Since the two regions are the two smallest ones
in Korea (4% of our sample), their impact on the main analysis is limited. In the Appendix, we replicate
the main results of this article with samples who lived in other regions than these two throughout the
period of analysis. The first-stage coefficient changes from 0.0775 to 0.0746, and the estimated second-
stage coefficients are robust to the exclusion.
TherelationshipbetweenBartikexposureandobservablecorrelates(secondstage)
Since the second-stage analysis estimates the local average treatment effect with Bartik exposure defined
at the region level, it is necessary to examine whether this instrument is correlated with other region-level
confounders. More specifically, if Bartik exposure of each region was correlated with some idiosyncratic
characteristics of the newspaper readers in that region that are also correlated with outcome variables,
the exclusion restriction would be violated. Goldsmith-Pinkham, Sorkin, and Swift (2020) recommend
checking the correlations between the Bartik instruments and observable correlates based on the logic
that observable confounders also indicate the importance of unobservable ones (Oster, 2019).
60
In the Appendix, we show the results of linear regressions of local market shares of the newspapers
and Bartik exposure on the regional averages of observable characteristics used in the main analysis. Al-
though each newspaper’s market share (level) at baseline has a significant correlation with some of the
observables,Bartikexposure (growth) is not correlated with any of them at 10% significance, verifying that
the instrument is orthogonal to observable correlates in the model.
Thelikelihoodofdivorce(extensivemarginanalysis)
As explained in Section 2.2.4, the main analysis of this article used 4,422 women who stayed in marriage
throughout Waves 5, 6, and 7 as a sample. What if exposure to feminism-related articles also influenced
married women’s decisions to get divorced? In that case, the estimated impact on resource allocation and
welfare could be biased since those who were most affected might have dropped out of the sample. In the
Appendix, we provide the result of the same 2SLS estimation as equation (4), with a dummy for getting
divorced or separated as a dependent variable. The estimated coefficient is not significant at the 10% level,
indicating a limited concern with biased attrition. In addition, The results of the main regressions are
robust to the inclusion of the attrited women.
2.6 Conclusion
In this paper, we tried to illustrate what happened to household resource allocation on a rare occasion
of society-wide changes in gender norms, which were prompted by a phenomenal rise of feminism and
increased media coverage of it in the mid-2010s in Korea. To econometrically separate the impact of the
media from common trends toward gender equality, we collected all feminism-related newspaper articles
in the past ten years and examined the relationship between region-year variations of women’s exposure to
those articles and their perceptions of gender norms. As a next step, we discussed how this media-induced
61
changes in women’s perceptions affected resource allocation. To overcome endogeneity, we instrumented
perceptions by Bartik’s shift-share IV that exploits the media’s coverage of feminism.
The results first proved a statistically significant impact of feminism-related newspaper articles on
women’s perceptions of gender norms. On the other hand, those affected women failed to shift the burden
of household labor to their husbands completely. Instead, both spouses agreed to outsource it to the market
and spend more time together. In consequence, women’s marital happiness was ameliorated.
Several implications can be derived from this research. First, evidence suggests that a conventional
notion that implicitly posits women as passive victims of unfair household labor allocation should be
revisited. As the bargaining model of household assumes, resource allocation within the household is a
result of dynamic games between a wife and a husband. We could see that women were actively trying
to improve their marital quality, considering gender norms as psychological constraints for both spouses.
When gender norms began to change, women adjusted their perceptions accordingly and found another
equilibrium where they could maximize utility.
Second, it should be highlighted again that the analysis of intra-household resource allocation must
consider its interactions with the market outside the household. Although labor economists have long
studied the labor market implications of bargaining between spouses, perspectives on the market out-
sourcing of household labor have been sparse. While new technology enables unprecedented reductions
in household labor, the equilibrium allocation will keep transforming.
Finally, we note one limitation of this study that can also suggest a promising direction for future
research. Although this article is based on the still-powerful influence of legacy media through various
channels, it does not reflect how the new modes of communication, including social media, convey in-
formation and affect how people think. Studying the influence of these new media on gender norms and
economic allocation would provide an interesting topic for research.
62
Chapter3
Self-employmentwithintheFirm
3.1 Introduction
Production in developing countries is predominantly done by small firms (Hsieh and Olken, 2014).
1
The
literature has suggested a multitude of reasons for this fact. Firms could struggle to grow due to high costs
of expanding, such as credit or hiring frictions, or limited access to output markets.
2
In addition, firms
might be small because they operate with low returns to scale. Low scale economies may be the result of
the available production technologies or low human capital and managerial ability of entrepreneurs.
3
In this paper, we study limited specialization of labor as a source of low returns to scale. One essential
reason why individuals may work together in a firm is that this allows the most skilled ones – which usually
includes the entrepreneur – to focus on the most complex tasks. Firm productivity is then closely linked
to the entrepreneur’s talent and hiring workers is a way for skilled entrepreneurs to leverage their ability.
In the absence of specialization, instead, expanding the firm is similar to adding another self-employed
individual who works in the same premises but whose productivity is independent of the entrepreneur’s.
Any barriers to labor specialization within the firm hence lower the returns to setting up large firms.
1
This chapter is co-authored with Vittorio Bassi, Alessandra Peter, Tommaso Porzio, Ritwika Sen, and Esau Tugume.
2
Banerjee and Duflo (2014), De Mel, McKenzie, and Woodruff (2008), Hardy and McCasland, (2022), De Mel, McKenzie, and
Woodruff (2019), Akcigit, Alp, and Peters (2021), Jensen and Miller (2018), and Startz (2019), among others, provide empirical
evidence on the importance of these constraints.
3
See, for instance, Bloom et al. (2013), Bruhn, Karlan, and Schoar (2018), and Atkin et al. (2017).
63
To shed light on the internal organization of firms in developing countries, we collected unique time-
use data for entrepreneurs and their employees in Ugandan manufacturing firms. We document limited
labor specialization: in firms across the size distribution, there is significant overlap in the tasks performed
by the entrepreneur and her employees. Even in relatively large firms, entrepreneurs spend most of their
time on menial production tasks, rather than on complex, managerial ones. We interpret the facts through
the lens of a model in which entrepreneurs hire workers and re-assign tasks between members of the
firm subject to a cost. We estimate the model using the micro data and find that, quantitatively, Ugandan
manufacturing firms closely resemble a set of self-employed individuals who share a production space.
This internal organization of the firm generates low returns to scale and reduces the benefits of alleviating
any other frictions that might constrain firm size.
We surveyed a representative sample of Ugandan manufacturing firms in carpentry, welding, and grain
milling. These sectors make up 32% of total employment in manufacturing and 27% of manufacturing
employment in firms with five or more employees. The sample thus allows us to capture the organization
of production in both small firms and, by East African standards, relatively large ones.
We build on our previous work and start from the sample of firms in Bassi et al. (2022). For these
firms, we already collected detailed data on firm characteristics as well as on the production process for
specific products (e.g., which production steps are done) and the economic environment in which they
operate. We then develop a novel survey module designed to record detailed time use information within
the firm: for both the entrepreneurs and their employees, we know which tasks they were involved in for
each hour of their last working day. Tasks were classified using a list of pre-specified activities including
numerous categories within “production” tasks (that is, the specific production step worked on), “non-
production” tasks (e.g., interacting with customers) and “idle” time (e.g., resting). Finally, we implement
additional rounds of surveys to complement this data with qualitative information on the organization of
labor within the firm.
64
In line with what the literature has documented in similar settings (Hsieh and Olken, 2014), most firms
in our sample employ less than 10 workers. However, they are not micro-enterprises: the median firm has
five employees and monthly revenues of $1,400. In principle, there is scope for labor specialization in these
firms. We then rely on the time use data to describe the extent of labor specialization both between and
within firms, and how it varies along the firm size distribution. To our knowledge, this type of evidence is
unique, at least in a developing country context.
We find no evidence of specialization between firms. Firms produce similar products, and even spend a
similar share of time on each production step. The time allocation between production and non-production
activities is also similar across firms, with all businesses spending close to 20% of total employee and
owner time on non-production activities such as customer interaction, supervision, or input procurement.
Importantly, there is no relationship with firm size along any of these dimensions: larger firms simply
operate as replicas of smaller ones, doing more of the same tasks.
Having established the lack of specialization across firms, we turn to the division of labor within the
firm along two dimensions: (i) between production and non-production tasks; and (ii) within production
activities.
On the specialization between production and non-production activities, we find two key results. First,
there is significant overlap in the tasks performed by employees and the entrepreneur. Entrepreneurs spend
a large share of their time on simple production activities, nearly as much as employees do. On the flip
side, employees spend a considerable share of their time on non-production, or managerial, tasks.
Second, specialization of entrepreneurs in non-production tasks does increase with firm size, but only
in a limited way: even in firms with more than five employees, entrepreneurs spend most of their time in
simple production activities. To get a sense of the magnitude, we show that the observed specialization is
less than half of that implied by a counterfactual “full specialization” benchmark. This exercises emphasizes
that the reason behind low specialization is not that firms are simply too small to do so. The limited
65
specialization we document has an important consequence: in larger firms, most of the non-production
tasks, which we show to be more complex, are done by the employees rather than by the entrepreneur,
who is typically more skilled.
4
We then turn to labor specialization within production, and assess whether employees or the en-
trepreneur specialize in certain steps. This margin of specialization is less important in our data: on
average, 85% of employees in a given firm work on the typical production step, and this varies little with
firm size. Entrepreneurs are relatively more likely to be involved in complex production steps, which is
indicative of vertical specialization also within production.
In summary, we show two main empirical results. First, the division of labor within the firm is limited
and we do not find evidence that larger firms take advantage of their scale through sharp organizational
changes: even in firms with 10 employees, the internal organization resembles that of a set of self-employed
individuals. Second, we find some evidence of (vertical) specialization in larger firms – entrepreneurs
specialize on more complex non-production tasks–, while “Smithian” (horizontal) division of labor across
production tasks is less common throughout the size distribution.
Finally, we study whether our results vary across the three sectors in our sample. We find a striking
pattern: while carpentry and welding are nearly identical, in grain milling there is significantly higher
specialization across all dimensions. For example, in grain-milling firms with ten employees, entrepreneurs
spend almost all their time on non-production activities, and workers are almost fully specialized across
production steps.
This result highlights that specialization is possible in this setting, thus suggesting that institutional
features such as contract enforcement or lack of trust are not the only drivers of the lack of specialization.
Also, the sectoral heterogeneity helps us shed light on possible barriers to specialization. Leveraging our
detailed data on the output market and customer interactions, we show that products are customized in
4
In line with this, we use additional survey data to confirm that in larger firms employees play a more direct role in customer
interactions and in generating demand for the firm, and they are also more likely to take on independent orders where they
manage production as well as the entire relationship with the customers as if they were self-employed.
66
carpentry and welding, where customers typically ask for personalized details and finishing, but more
standardized in grain milling, where the production process simply involves turning the maize —brought
by the customer – into flour. We also show that customization may entail significant communication
and coordination costs within the firm, as the person producing the order must precisely know the spe-
cific product characteristics the customer asked for. This makes it difficult for the firm to “unbundle” the
production process into separate tasks that can be performed by separate, specialized, individuals. We
conclude that lack of standardization could be one important reason for limited specialization, and thus
hinder economies of scale.
5
Motivated by the data, we develop a model that serves two purposes. First, the model helps formalize
the two-way relationship between labor specialization and firm size and show how barriers to specializa-
tion impact firm-level and aggregate productivity. Second, it provides a framework to quantify the extent
of specialization costs in our setting as well as their implications for firm size, productivity, and selection
into entrepreneurship.
The heart of the model is an assignment problem of workers to tasks, which determines firm productiv-
ity. We embed this problem into an otherwise standard model of industry equilibrium with an occupational
choice.
Production of each unit of output requires completing a set of tasks, which differ in their level of
complexity. If self-employed, each individual must complete all the tasks herself. When working together
in a firm instead, individuals can unbundle the production process and delegate some tasks to others. Firm
productivity has two components. The first one depends purely on the entrepreneurial ability and reflects
the non-rival role of talent in production (e.g., a business idea). The second one is the average of the
abilities of all individuals producing, weighted by share of time each spends on the complex tasks: while
everyone can equally complete the simple tasks, high ability individuals are better at the complex ones.
5
The link between standardization and scale of operations has been established in the literature (Piore and Sabel, 1984; Holmes
and Stevens, 2014.
67
This second component depends on the allocation of talent within the firm and implies that delegating
complex tasks to more skilled individuals – i.e. specializing labor based on comparative advantage – can
increase firm productivity. Unbundling tasks, however, comes at a cost, which encapsulates the barriers
to labor specialization.
All individuals choose whether to start a firm or work as employees. Entrepreneurs choose how many
workers to hire, subject to a revenue wedge, which captures any external constraint that may keep firms
small (e.g., credit constraints). Workers are randomly assigned to firms – entrepreneurs merely choose
the mass of employees – and are compensated with a combination of a piece-rate component and a wage
level. The wage is endogenous and clears the labor market. Entrepreneurs then assign everyone (including
themselves) to tasks within the firm. The model yields a familiar talent segmentation by occupation. The
most skilled individuals sort into entrepreneurship, where the returns to ability are higher.
We first show that labor specialization and optimal firm size are closely intertwined. On the one hand,
there is limited scope for specialization in small firms: there are simply not enough complex tasks to fill
the entrepreneur’s time. On the other hand, a high cost of delegation reduces the returns to setting up
large firms: when delegation is costly, the owner cannot leverage her managerial ability by performing
the complex tasks of her workers and increasing their productivity.
We then characterize theoretically how the unbundling cost affects the allocation of talent within and
between firms, ultimately determining the firm size distribution and aggregate productivity.
The size of the unbundling cost determines the nature of the firm. When the cost is low, entrepreneurs
specialize on the most complex tasks and firm productivity is only a function of their ability. In this case,
entrepreneurial talent is scalable and firms are vehicles to leverage this talent. When the cost is large, each
worker is essentially self-employed within the firm , that is, performs tasks of all levels of complexity. Firm
productivity is equal to the average ability of all individuals and decreases in firm size, since any additional
worker is less skilled than the entrepreneur herself. In this case, firms are mere vehicles to share fixed
68
costs. The efficient firm size is smaller, since the lack of specialization generates strong decreasing returns
to scale.
The unbundling cost not only transforms the way in which firms are internally organized but has also
equilibrium effects that ripple through the economy. Higher labor specialization increases firm produc-
tivity, and thus the demand for labor. As a result, wages increase, leading some marginal entrepreneurs
to become workers, further increasing aggregate productivity through a classic selection effect. Overall,
when the unbundling cost is low, managerial ability is highly priced in the economy, talent can be leveraged
by taking over more and more complex tasks, and only large firms operate.
Before bringing the model to the data, we provide two qualitative tests supporting its theoretical pre-
dictions. First, given its higher specialization, grain milling should have larger firms, a bigger ability gap
between entrepreneurs and workers, and a stronger relationship between entrepreneurial ability and firm
size as well as revenues. We verify that all these predictions hold in the data. Second, we validate one
unique implication of our model: all else equal, entrepreneurial talent should be less relevant in larger
firms. In the absence of a credible instrument for firm size, we rely on cross-regional heterogeneity within
Uganda. We show that in regions with smaller firms, there are larger returns to managerial talent. This is
consistent with our model if we are willing to assume that regions with smaller firms face larger external
constraints to firm size – i.e. higher revenue wedges.
To make the model amenable to quantitative analysis, we then extend it along three dimensions: we
include an external sector, allow for preference shocks over occupations and for size-dependent distortions
by letting the revenue wedge be a function of firm size.
We estimate the model using data from carpentry and welding. The model yields a structural relation-
ship between firm size and the time spent on non-production tasks by entrepreneurs and employees. Using
our unique time-use survey, we can use this relationship to pin down the unbundling cost. Given this cost,
the joint distribution of firm sizes and revenues disciplines the heterogeneity in entrepreneurial ability as
69
well as the level of the revenue wedge and its relationship with firm size. To quantify the importance of
non-rival entrepreneurial talent, we use the relative importance of working in a high productivity firm
versus being a high skilled worker for wages.
With the estimated model, we perform two main exercises. First, we vary the size of the unbundling
cost and show that, in terms of average firm size and productivity, the current economy is closer to a
benchmark without labor specialization than to one in which labor can fully sort within firms based on
comparative advantage. This result provides a quantitative answer to the question: “Why do firms exist?”.
The empirical analysis pointed to vertical specialization between entrepreneur and employees as the main
source of labor specialization, and thus toleveragingentrepreneurialtalent as a plausible benefit of produc-
ing in firms. The quantitative analysis, however, shows that also this type of specialization is weak. The
current equilibrium closely resembles the case of self-employment within the firm , where workers operate
side by side, doing similar tasks and simply sharing the same premise and machines.
Second, we show that the benefits of relaxing other constraints to firm scale hinges on the estimated
unbundling cost. Relative to our benchmark estimates, calibrating the unbundling cost even just to match
the larger specialization observed in grain milling would almost double the effect of a reduction in the
revenue wedge on firm size and aggregate productivity. This result highlights a key takeaway of our
work. Barriers to within-firm labor specialization make traditional manufacturing a business model that
is difficult to scale. As a result, the returns from policy interventions aimed at spurring firm growth may
be limited.
Related literature. Our main contribution is to uncover that the internal organization of firms in de-
veloping countries features limited labor specialization, and to quantify the extent to which this limits
firm size and productivity. In doing so, we contribute to an established literature on the determinants of
why firm productivity is lower in developing countries and the size distribution is dominated by many
70
small firms (Bloom et al., 2010; Hsieh and Olken, 2014). This literature has focused on the role of manage-
ment (Bloom et al., 2013) and limits to delegation (Akcigit, Alp, and Peters, 2021), emphasizing that the
low adoption of managerial practices and frictions in the external labor market for managers impede firm
expansion and productivity growth. Our contribution is to highlight how barriers to labor specialization
inside the firm prevent entrepreneurs from fully leveraging the managerial talent they possess.
6
Second, we contribute to a classic theoretical literature in organizational economics, which has long
emphasized the importance of specialization of labor within the firm for productivity and growth (Chan-
dler, Hikino, and Chandler, 2009; Becker and Murphy, 1992; Bolton and Dewatripont, 1994; Yang and
Borland, 1991). More recently, this literature has focused on understanding the role of managerial hierar-
chies for firm growth, both theoretically and empirically (Garicano and Rossi-Hansberg, 2006; Caliendo,
Monte, and Rossi-Hansberg, 2015). We contribute by providing novel micro-data on task allocation be-
tween entrepreneurs and employees within developing country firms, which allows us to directly measure
specialization at a granular level. Our results point to the importance of vertical specialization between
entrepreneurs and employees as a more important reason for why firms exist than division of production
tasks among similarly skilled employees.
Our analysis shows that job roles are not as informative about specialization in small firms in devel-
oping countries, where entrepreneurs and workers conduct similar tasks. This highlights the importance
of collecting data on tasks to understand the internal organization of labor in small developing country
firms: inferring specialization from an internal hierarchy based on job roles would lead to misleading
conclusions.
7
Finally, A related literature studies the role of frictions in the output market as a barrier to growth
(Bold et al., 2022; Jensen and Miller, 2018; Startz, 2019; Vitali, 2022). We contribute by showing how a
6
A related literature studies the impact of interventions granting more autonomy to managers and bureaucrats inside large
organizations in developing countries, finding positive impacts on productivity (Kala, 2019; Bandiera et al., 2021).
7
A recent empirical literature leveraging detailed production data from large factories has examined the impact of managers
on worker productivity (Adhvaryu, Nyshadham, and Tamayo, 2019) and the sorting of workers to managers (Adhvaryu et al.,
2020). We add to this literature by studying the determinants of labor specialization within the firm.
71
specific feature of the output market – the characteristics of the products that are made by firms – impacts
firm productivity by affecting the internal allocation of labor.
8
Structureofthepaper. The rest of the paper is organized as follows. In Section 2 we describe the survey
and data collection. In Section 3 we present key facts on labor specialization both across and within firms.
Section 4 develops the model, Section 5 describes the estimation, and Section 6 reports our quantitative
results and counterfactuals. Section 7 concludes. Additional results are in the Online Appendix.
3.2 TheSurvey
This section describes the survey we conducted in Uganda to study labor specialization inside the firm. The
survey includes two waves: an initial round of data collected in 2018-19 and a follow-up survey conducted
in the spring of 2022. The main elements of the initial survey are described in detail in ASC. Here we briefly
summarize the key elements of the sampling strategy and survey design and then focus on those sections
of the survey instrument designed specifically to measure labor specialization inside the firm, which do
not feature in our previous work. We then describe the follow-up survey, which was designed specifically
for this paper.
3.2.1 Sampling
Our survey targeted firms in carpentry, welding and grain milling. We chose these sectors because: (i)
they are large, employing about 30% of workers in manufacturing; and (ii) they include both smaller and
8
Jensen and Miller (2018) is a particularly related paper in that they show that firms specialize labor as they grow larger
due to an exogenous increase in market share. While we also show that small firm size reduces specialization, the key focus of
our paper is to show that barriers to specialization hinder firm size in the first place, and to isolate and quantify each channel of
the two-way relationship between specialization and firm size. Holmes and Stevens (2014) show that in the US firms producing
customized products tend to be small as they serve a small share of the market. We contribute by linking product customization
to the internal organization of the firm and aggregate outcomes.
72
– for Ugandan standards – larger firms, which allows us to study labor specialization across the size dis-
tribution.
9
We selected a representative sample of 52 sub-counties for the survey, stratifying by total population
and whether the sub-county is in the broader area of Kampala, the capital city.
10
Within each sub-county,
we first conducted a complete firm listing in these sectors, identifying close to 3,000 firms. We then ran-
domly sampled about 1,000 firms from the listing, oversampling firms employing more than five workers
to ensure enough observations among relatively larger firms. We interviewed the entrepreneur and all
employees working on pre-specified core products that are common in the three sectors: doors in carpen-
try, windows in metal fabrication, and maize flour in grain milling.
11
Our final sample includes the 1,115
entrepreneurs and 2,883 employees who answered the initial survey.
12
3.2.2 MainSurvey
The main survey was conducted through in-person visits at the firms in 2018-19. The aim of the survey
was to collect detailed information onhow firms produce and on the economic environment in which they
operate. Specifically, we asked detailed questions on: (i) the entire production process for the core products;
(ii) demographics of entrepreneurs and employees; (iii) output and labor market; (iv) labor specialization
inside the firm.
Starting from the measurement of the production process, we collected information on the hours of
machine and labor time used in each production step, as well as on the characteristics of the machines and
workers involved in each step. Moving to the details of entrepreneurs and employees, we have information
on age, years of education, as well as a range of measures of individual ability, including a standardized
9
The latest Census of Business Establishments from the Uganda Bureau of Statistics from 2010 shows that these three sectors
comprise 32% of total manufacturing employment and 27% of manufacturing employment in firms with five or more employees.
10
A sub-county is the second largest sub-national administrative unit in Uganda. The average sub-county consists of 5,285
households and spans 4.4 square miles.
11
If a firm did not produce the pre-specified core product, we interviewed all employees working on the main product of the
firm. See ASC for details.
12
Compliance with the initial survey was high at over 90% and all our results are appropriately weighted to reflect our sampling
strategy.
73
index of managerial ability for firm entrepreneurs (based on multiple survey questions (McKenzie and
Woodruff, 2017) and measures of cognitive and non-cognitive skills for employees. On output market
characteristics, we collected information on how firms find customers and on perceptions about constraints
to growth, as well as on prices of the core product. On the labor market, we know the mode of payment
of all employees, their tenure, and various indicators of labor market frictions, such as the time taken by
the entrepreneur to replace employees who leave. Finally, to measure labor specialization inside the firm,
we used time diaries of entrepreneurs and employees, as described in more detail in the next paragraph.
13
Measuringlabor specialization. To measure labor specialization, we asked the entrepreneur and ev-
ery employee to indicate all the hours in which they worked for the firm in the last day worked. For each
hour, we then asked to indicate the specific tasks they performed, choosing from a pre-specified list of
16 possible tasks (plus an option to specify other tasks not on the list). The list included tasks related to
“production” (e.g. working on the main product of the firm) but also “non-production” tasks encompassing
all other managerial/organizational activities that typically need to be carried out to run a business, such
as managing customer interactions, book-keeping and financial management, sourcing of inputs, mainte-
nance of machines, management of stock, or supervision and training. Finally, we know the amount of
time that workers and owners spend being inactive, that is, away from the firm for non-business reasons
or eating/resting.
In addition to time allocation across production and non-production tasks, we also collected informa-
tion on the share of production time that owners and employees spend on the various production steps
for the core product (e.g., doors in carpentry).
13
ASC primarily use the data on the production process to study the rental market for machines in carpentry. The data on
labor specialization is not used in this previous paper.
74
Table 3.1 summarizes all the tasks that we measure in our time use survey. This unique data allows
us to construct the share of time entrepreneurs and employees spend on different tasks, which is the key
input for our measures of labor specialization between and within firms.
3.2.3 Follow-upSurvey
In the spring of 2022, we conducted a follow-up survey of the sample of entreprenerus and employees
interviewed in the main survey.
14
The main aim of the follow-up survey was to gather additional details
on labor specialization inside the firm.
On labor specialization inside the firm, we collected additional information on interactions with cus-
tomers, to shed light on how product customization may lead to production activities and non-production
activities such as customer management activities being bundled together. For instance, we asked whether
employees effectively perform “independent orders” at the firm, that is, whether they take on orders that
are conducted at the firm’s premises but are managed entirely by the employee, who is the residual claimant
on the profits from the order.
Attrition from the follow-up survey is described in the Appendix. In short, we successfully managed
to trace and interview about 68% of entrepreneurs and 59% of employees. Attrition is higher in the grain
milling sector, but not differentially so by firm size and managerial ability, which are two key variables of
interest in the empirical analysis of the next section. This survey is used to provide additional qualitative
evidence on possible barriers to specialization, and none of the data moments used for estimation uses this
follow-up survey.
15
14
The follow-up survey was conducted by phone. In total across the three sectors, we had phone numbers of 1,101 en-
trepreneurs (out of the 1,115 interviewed in the initial survey) and 2,177 employees (out of the initial 2,883), so this is the sample
targeted in the follow-up survey.
15
Uganda was subject to two strict Covid-19 related lock-downs, the first in April-May 2020 and the second in July 2021.
Our follow-up survey was collected in spring 2022, and so well after the Covid-19 related restrictions were lifted. Our follow-up
survey asked about firm closures around the time of the second lockdown, and the data shows that while 46.3% of firms were
either closed or only partially open during the lockdown, 88.5% were fully operational quickly after, so concerns about firms
exiting from our panel due to the Covid-19 shock are not first order. In a separate survey of the same sample of firms conducted
after the first lockdown (and not used for this paper), Bassi et al. (2021) find a similar pattern of a sharp but largely transitory
increase in firm closures during the first lockdown.
75
Table 3.1: Measuring Time Use
Panel A: All Tasks
(i)Production Supervising other workers Looking for workers
Preparation (Main product) Training other workers Managing loans
Processing (Main product) Book-keeping Other non-prod. tasks
Finalizing (Main product) Maintanence
Producing other products Organizing stock (iii)Idle
Procuring inputs Eating/Resting
(ii)Non-productionTasks Looking for input suppliers Waiting for customers
Interacting with customers Looking for new machines Away not for business
Panel B: Production Steps (Main Product)
(i)Carpentry (ii)Welding (iii)Grainmilling
Design Design Cob shelling
Drying (before production) Cutting Drying
Cutting Bending Cleaning/Destoning
Planing Grinding Conditioning
Thicknessing Welding De-hulling
Edging Polishing Milling
Sanding Painting Sealing
Mortising
Finishing
Drying (after painting)
Notes: The table reports the complete list of tasks reported in our time use data. Panel B breaks down production into the pre-
specified steps. The steps are listed in the typical order of implementation. Steps classified as “Preparation (Main product)” in
Panel A correspond to the following steps listed in Panel B: (i) Carpentry: Design∼ Drying (before production), (ii) Welding:
Design, (iii) Grain milling: Cob shelling∼ Conditioning. Steps classified as “Processing (Main product)” in Panel A correspond
to the following steps listed in Panel B: (i) Carpentry: Cutting∼ Mortising, (ii) Welding: Cutting∼ Welding, (iii) Grain milling:
De-hulling∼ Milling. Steps classified as “Finalizing (Main product)” in Panel A correspond to the following steps listed in Panel
B: (i) Carpentry: Finishing∼ Drying (after painting), (ii) Welding: Polishing∼ Painting, (iii) Grain milling: Sealing. Table C.1 in
the Appendix reports the overall share of firm-level time spent in each of these tasks, where the firm-level time is computed by
summing the time spent in a given task by the entrepreneur and all employees in a firm.
3.3 TheOrganizationofProductionInsidetheFirm
In this section, we use our survey data to describe the organization of labor inside the firm and how this
varies across the size distribution. We begin by presenting basic descriptive statistics about the firms in our
sample and the economic environment where they operate, which are important to understand the context
of our study. Then, we use our novel time use data to studywhatfirmsdo (i.e., “task composition”) andwho
does what within the firm (i.e., “task allocation”). This allows us to characterize two possible margins of
76
specialization: (i) firms specializing in different activities, and (ii) individuals within the firm specializing
in different tasks. We first show that (i) does not play a significant role, which then allows us to focus
on (ii) in detail. We conclude this section by presenting heterogeneity in specialization by firm size, to
show that even within the same institutional environment there are differences in specialization, and to
highlight possible barriers to specialization.
3.3.1 Firmcharacteristics,andmarketsforoutputandlabor
We use our survey data to describe basic characteristics of the firms in our sample and their entrepreneurs,
the output market where they operate, as well as the characteristics of their employees and the labor market
in urban Uganda.
Firmandentrepreneurcharacteristics. Figure 3.1 shows firm size distributions in our three sectors.
In line with the literature on the firm size distribution in developing countries ([79]), most firms employ
less than ten workers. However, most manufacturing firms in these sectors are not micro-enterprises: the
median firm employs five workers in addition to the entrepreneur. So, in principle, there is be scope for
labor specialization in most of the firms in our sample. Figure 3.1 further shows that firms are slightly
larger in grain milling and the dispersion in firm size is also larger in grain milling. We will come back to
this result in the model section where we show that our model is able to explain this difference in size and
dispersion in size across sectors.
77
Figure 3.1: Firm Size Distribution
0
.1
.2
.3
Density
1 2 3 4 5 6 7 8 910+
Firm Size
Carpentry
0
.1
.2
.3
1 2 3 4 5 6 7 8 9 10+
Firm Size
Welding
0
.1
.2
.3
1 2 3 4 5 6 7 8 910+
Firm Size
Grain Milling
Notes: Sample: all surveyed firms. All figures are weighted by sampling weights within each sector. Firms of size larger than 10
are grouped together in the category “10+”.
Panel A of Table 3.2 shows that the firms in our sample are well-established entities: the average
firm generates around $1,400 in revenues per month and has been in business for 10 years. In addition,
the large majority are registered with the local authority. The average firm uses about 20% of the pre-
specified modern machines listed in our survey, showing that the production process of the average firm
is at least partly mechanized. Panel A also shows that firms tend to produce similar products within each
sector: about 70%-90% of the firms across sectors make the core product that we asked about. The average
entrepreneur has 10 years of education and scored about 11 points on a battery of questions measuring
managerial ability and ranging from -1 to +27. In all these dimensions, there are little differences between
sectors.
78
Output market. Panel B shows that firms struggle to grow their customer base and establish a wider
brand reputation through marketing: the majority of firms cite lack of demand as a critical constraint to
growth and only 9% of firms spend any money on advertising. As a result, they operate in an informal
and highly local product market where personal interactions with customers are key: 58% of firms report
that directly talking to customers is their main strategy to signal product quality. In line with this, most
sales are made through walk-ins. Importantly, most sales are made to order, rather than the customer
buying a ready-made product. These features of how firms find customers and sell to them are consistent
with significant output market frictions driven by difficulties in accessing wider markets, and is in line
with a growing literature on output market frictions in developing countries. Panel B shows a remarkable
similarity of these features of the output market across sectors.
16
Taken together, the evidence in Panels A and B shows that manufacturing firms in this context do
two core activities: (i) they produce goods for sale; (ii) they engage with final customers through personal
interactions to generate demand for the firm.
Employeecharacteristicsandlabormarket. Panel C shows that employees are slightly less educated
than entrepreneurs on average. In terms of the labor market, the average tenure of employees is about
3.5 years. This confirms that manufacturing firms offer stable employment in this labor market. The
most common type of payment is by far piece-rate: over 90% of employees are paid at least a piece-rate
component (the share paid exclusively piece-rate is around 80%), thus indicating that workers likely share
some of the productivity benefits of working at more productive firms. Finally, the average entrepreneur
reports that it would take them over five weeks to replace a worker who leaves. This is indicative of
sizeable labor market frictions, the prevalence of which is documented in more detail in the same context
by ASC. In line with this, Appendix Figure C.1 shows that there is a substantial amount of overlap in the
16
These are sectors with very limited imports and exports. Using the 2019 VAT and customs data for Uganda, we show that
imports and exports in the carpentry sector amount to just about 3% and 1% of domestic sales. Table C.2 in the Appendix shows
that the price of carpentry goods in Uganda is low relative to that of similar IKEA products across the world, which can explain
why penetration of large multinationals through imports is limited.
79
education distribution of workers employed by entrepreneurs above and below median managerial ability,
confirming that sorting in the labor market is limited.
Table 3.2: Descriptives on Firm Characteristics and Output Market
All Carpentry Welding Grain
milling
(1) (2) (3) (4)
Number of firms 1,115 522 433 160
PanelA.FirmandEntrepreneurChars.
Number of employees 4.8 4.5 4.9 6.0
Monthly revenues (USD) 1,437 1,222 1,548 1,916
Firm age (Yrs.) 10.1 10.4 8.9 12.0
Formal license 82% 76% 86% 91%
Mechanization (rel. to max machine utilization) 19% 23% 17% 19%
Makes core product 80% 68% 92% 93%
Managerial ability score of entrepreneur (-1; +27) 11.2 11.2 10.9 11.8
Entrepreneur education (Yrs.) 10.0 9.8 10.0 10.9
PanelB.OutputMarket
Lack of Demand as a main constraint to growth 55% 55% 54% 57%
Any marketing expenditure 9.3% 7.7% 12.0% 7.8%
Sales are done through walk-ins 80% 80% 75% 94%
Sales are made to order 79% 75% 89% 69%
Talks to customers to communicate product quality 58% 61% 57% 54%
PanelC.EmployeeChars. andLaborMarket
Employee education (Yrs.) 9.3 8.9 10.2 7.9
Employee tenure (Yrs.) 3.5 3.5 3.3 3.9
Employees paid piece-rate 91% 93% 92% 83%
Expected time taken for finding a replacement (Days) 37.5 38.8 41.4 23.6
Notes: Means are reported. Sample: surveyed firms. All statistics are at the firm level, weighted by sampling weights. 1 USD =
3,800 UGX for monetary amounts. Monthly revenues: average revenues in the three months preceding the survey (trimmed at
top 1%). Mechanization: types of machines used relative to all potential types of machines for producing main product. Main
product: doors in carpentry, windows in welding, and maize flour in grain milling. Managerial ability score: index going from -1
to 27 and based on multiple survey questions.
3.3.2 TaskComposition: WhatDoFirmsDo?
In this section, we use our time use data to study which tasks are done by firms and whether different
firms specialize in different tasks. That is, we study whether there is evidence of specialization acrossfirms
80
in what firms do. We present the analysis pooled across sectors, and then return to heterogeneity across
sectors in the final part of this section.
Figure 3.2: Task Composition across the Size Distribution
(a) Average Production Step
0
20
40
60
80
100
Average Share of Firms Doing a Step (%)
1 2 3 4 5 6 7 8 9 10+
Firm Size
Average Share
(b) Total Firm Hours
0
20
40
60
80
100
Share of time (%)
1 2 3 4 5 6 7 8 9 10+
Firm Size
Idle
Production
Non-production Tasks
(c) Hours in Production
0
20
40
60
80
100
Share of time (%)
2 3 4 5 6 7 8 9 10+
Firm Size
Preparation
Processing
Finalizing
(d) Hours in Non-production Tasks
0
20
40
60
80
100
Share of time (%)
1 2 3 4 5 6 7 8 9 10+
Firm Size
Customer interaction
Supervision/training
Operations/logistics
Notes: Sample: all surveyed firms. Time use reported by interviewed entrepreneurs and employees. All figures are weighted
by sampling weights within each sector and the relative number of surveyed firms for sectors. Panel (a): For each pre-specified
production step, a dummy was generated about whether a firm performs the step. Then the average was computed across the
firms for each size. Panel (b) reports the share of firm-level time in the three categories of Production, Non-Production and Idle
Time. Panel (c) shows the breakdown of the firm-level time in production into the sub-categories of Preparation, Processing
and Finalizing. Panel (d) breaks down non-production time into customer interaction, supervision and operations/logistics. See
Table 3.1 for more details on task categories. The category operations/logistics in Panel (d) includes all tasks listed between
book-keeping and Other non-prod tasks from Table 3.1.
Which production steps firms do. As we have shown in Table 3.2 that most firms engage in the
production process of the pre-specified core product, we begin by studying whether firms specialize in
some of the production steps needed to make the core product. We limit the sample to firms engaged
in the production of the core product, and for each production step we compute the share of firms that
81
perform that step. We then aggregate across steps weighting by the average share of time spent on that
production step in the data, so that steps that represent a larger fraction of total production time get a
higher weight. We average across sectors weighting by the share of firms in each sector. The results are
in Figure 3.2, Panel (a). The Figure shows that: (i) each step is done by most firms: the average share of
firms doing the representative step is higher than 80%; (ii) this does not vary across the size distribution.
These results show that there is no significant specialization across production steps, neither in smaller
nor larger firms.
How firms allocate time between production and non-production. Having shown that firms do
not specialize across production steps, we use our time use data to study whether firm specialize in the
tasks performed to produce goods and sell them to customers, i.e., to run the firm activities. While clearly
firms need to spend time in both production and non-production, or managerial, activities, some firms
could specialize in marketing or building customer relationships, while others mostly in producing goods.
To shed light on this, we start by examining the share of total firm time spent in: (i) production; (ii) non-
production and (iii) idle time, following the classification of tasks described in Table 3.1. To do so, we sum
up all the time spent on each of these three groups of tasks by the entrepreneur and all employees in the
firm, and divide this by total firm-level time spent in all activities. The results are in Figure 3.2, Panel (b),
which shows that the main activity that firms do is production, amounting to around 60% of total firm
time. This is not surprising given that these are manufacturing firms. Non-production activities are also
important however, accounting for slightly less than 20% of time. The rest of the time is spent idle (either
resting or waiting for customers). The Figure further shows that this breakdown of firm activities does
not vary across the firm size distribution, indicating that there is no specialization into these three broad
classes of tasks.
Of course, it could be that firms specialize in different activities within production: while it’s true that
most firms engage in most production steps, some firms may specialize in certain steps by spending a
82
higher share of time there. Similarly, some firms may specialize in customer interactions, while others in
sourcing inputs. We turn to these possibilities next.
Howfirmsallocatetimeacrosstaskswithinproduction. Figure 3.2, Panel (c), shows the breakdown
of firm time within production activities. To be able to pool across sectors, we harmonize production steps
by dividing them into the three categories of preparation, processing and finalizing described in Table
3.1. Not surprisingly, most of the production time is spent on processing steps. We clearly see that the
share of time in the various production tasks is remarkably stable across the size distribution: firms do not
specialize in different tasks within production.
17
Howfirmsallocatetimeacrosstaskswithinnon-production. Finally, Figure 3.2, Panel (d), studies
how firms spend their time within non-production activities. To do so, we focus on three main groups
of activities within non-production: (i) customer interactions; (ii) supervision and training; and (iii) op-
erations,finance, and logistics.
18
The Figure shows that about half of non-production time is spent on
operations, finance, and logistics. Customer interactions are also important, accounting for about 30% of
non-production time. The Figure shows that the share of time in the different activities is once again fairly
constant across the size distribution.
19
Taken together, the results in Figure 3.2 show a remarkable similarity in how firms spend their time
across the size distribution. As firms grow, they do not specialize in different activities: large firms just
do more of the same tasks. In other words, there are no economies of scale driven by changes in task
composition.
17
Appendix Figure C.2 shows that this remains true when we look at the share of firm time spent in each individual production
step, by sector.
18
This last category combines time spent in non-production activities listed between “book-keeping” and “other non-prod.
tasks” from Table 3.1.
19
Figure 3.2, Panel (d), does show that there is slightly more supervision/training in larger firms while firms of size one spend
almost no time supervising/training. This is reassuring as we would naturally expect this activity to scale with firm size. Figure C.3
in the Appendix shows that the share of time in the different components of idle time is also constant across the size distribution.
83
3.3.3 TaskAllocation: WhoDoesWhatWithintheFirm?
In this section, we study the division of labor inside the firm. Labor specialization within the firm could
generate scale economies and be at the core of why individuals come together in a firm. As we have shown
that firms do similar tasks along the size distribution, this allows us to study the division of labor inside the
firm without potential selection concerns related to smaller and larger firms engaging in different tasks.
Both horizontal and vertical labor specialization could be important. Horizontal (or “Smithian") spe-
cialization refers to division of labor across equally complex tasks: like on the assembly line of Henry
Ford, by specializing in a given task, individuals can increase their proficiency in that task, leading to an
overall increase in productivity. Vertical specialization refers to the division of labor based on comparative
advantage, whereby individuals more skilled in certain tasks are assigned to those tasks (e.g., more able
individuals are assigned to more complex tasks).
We study two dimensions of labor specialization. First, we provide evidence on specialization between
production and non-production tasks. This is a vertical type of specialization, as we will show that non-
production tasks are on average more complex. Second, we study the division of labor within production
across steps. As we find evidence that some production steps are more complex than others, this corre-
sponds to a combination of horizontal and vertical specialization. For each of these two dimensions, we
study specialization between the entrepreneur and employees, as well as between employees.
Laborspecializationbetweenproductionandnon-productiontasks
We begin by studying labor specialization between production and non-production activities. To do so,
in Figure 3.3, we compare the time spent on each task by the entrepreneur with the time spent on the
same task by the average employee. The y-axis reports the different tasks from our survey, where we color
in blue those related to production, in red those related to non-production, and in grey those related to
idle time. Each bar then reports the sum (normalized to 100%) of the time spent on that activity by the
84
entrepreneur and the average employee. We color in dark the (normalized) time spent by the entrepreneur
on that task, and in light the (normalized) time spent by the average employee: if the entrepreneur and the
average employee spend the same amount of time on a given task, the dark and light bars each amount
to 50%. The Figure shows that: (i) there is substantial overlap between entrepreneurs and the average
employee in the time they spend in all tasks; (ii) entrepreneurs specialize in non-production tasks. That is,
we do find some evidence of division of labor between entrepreneurs and employees along this dimension
of specialization. The fact that entrepreneurs spend more time in all non-production tasks also validates
our classification of tasks into production and non-production.
Figure 3.3: Task allocation between Production and Non-production Tasks
0 20 40 60 80 100
Share of time (%)
eat/rest/wait
away not for business
organize stock
other managerial tasks
look for workers
maintanence
train other workers
interact with customers
manage loans
procure inputs
look for new machines
look for input suppliers
book-keeping
supervise other workers
managerial tasks
production (other)
production (main final)
production (main process)
production (main prep)
production (main total)
production
Notes: Dark bars: entrepreneurs. Light bars: average employee. Blue bars: production tasks. Red bars: Non-production tasks.
Grey bars: Idle time. Sample: all surveyed firms. Time use reported by interviewed entrepreneurs and employees.
This result is meaningful because we also find evidence that: (i) non-production tasks are more complex
and (ii) entrepreneurs are more skilled than employees. On the first point, Appendix Table C.3 shows that
employees assigned to non-production tasks earn substantially more. On the second point, Appendix Table
C.4 shows that entrepreneurs on average are the most skilled person within the firm. Figure 3.3 therefore
85
uncovers evidence of vertical specialization between entrepreneurs and employees. In Appendix Figure
C.4 we replicate Figure 3.3 but comparing higher and lower skilled employees (splitting by median salary
within the firm): we find much higher overlap and lower vertical specialization between employees on
this dimension.
Figure 3.4: Task Allocation between Production and Non-production by Firm Size
(a) Employee
0
.2
.4
.6
.8
1
Share of Time in Non-prod. Tasks
1 2 3 4 5 6 7 8 9 10+
Firm Size
High earnings Low earnings
(b) Entrepreneur
0
.2
.4
.6
.8
1
Share of Time in Non-prod. Tasks
1 2 3 4 5 6 7 8 9 10+
Firm Size
(c) Employee + Entrepreneur
0
.2
.4
.6
.8
1
Share of Overall Firm Time in Non-prod. Tasks
1 2 3 4 5 6 7 8 9 10+
Firm Size
(d) Employee vs. Entrepreneur
0
.2
.4
.6
.8
1
Within-Firm Share of Time in Non-prod. Tasks
1 2 3 4 5 6 7 8 9 10+
Firm Size
Employees Entrepreneur
Notes: Sample: all surveyed firms. Shaded areas: 95% confidence intervals. The size of dots and squares represent the number
of firms in each size group. Time use reported by interviewed entrepreneurs and employees. Panel (a): Employee share of
time in non-production tasks. Employees are classified as high and low earnings workers by reported salaries within each firm.
Panel (b): Entrepreneur share of time in non-production tasks. The red squares represent the benchmark specialization, where
all available time (except for idle time) of entrepreneurs are hypothetically reassigned to non-production tasks. Panel (c): The
total (entrepreneur + employee) share of firm time in non-production tasks. Panel (d): Comparison of relative time use in non-
production tasks between entrepreneur and average employee.
86
Largerfirmsaremorespecialized. In Figure 3.4, we study how the vertical specialization documented
in Figure 3.3 varies across the size distribution. To do so, we report the average share of time spent in
non-production tasks by firm size together with the line of best fit, both for employees (Panel (a)) and
entrepreneurs (Panel (b)). The Figure confirms first of all that specialization among employees is limited
and does not vary with firm size: employees spend about 20% of their time in non-production activities, and
while high-skilled employees (as measured by earnings) spend a little more time on non-production tasks,
the gap with low-skilled employees is small and constant across the size distribution. On the other hand,
entrepreneurs specialize in non-production tasks, and the gap in time allocation with employees increases
in firm size: larger firms are more specialized. Still, specialization is limited even in large firms: to show
this, we compute for each firm the (counterfactual) share of time that the entrepreneur would spend in
non-production tasks if entrepreneurs were fully specialized in non-production tasks.
20
This empirical full
specialization benchmark is reported in pink, and shows that even in firms with more than five employees
is just spending 50-60% of their time in non-production tasks, while in principle there would be enough
non-production tasks for the entrepreneur to spend their entire time in non-production activities. This
makes clear that the reason why firms in our data do not exhibit more specialization is not just that they
are too small to specialize.
Inlargerfirms,mostnon-productiontasksaredonebyemployees. As a consequence of the lim-
ited specialization of entrepreneurs in non-production tasks, in larger firms most of the non-production
activities are done by employees, not the entrepreneur. This is show in Figure 3.4, Panels (c) and (d). Panel
(c) confirms again that the share of firm-level time in non-production tasks is constant across the size dis-
tribution at around 20% (in line with Figure 3.2, Panel (b)). However, Panel (d) shows that who does the
non-production tasks varies dramatically with firm size: in firms of size one and two, naturally most of
20
To do so, we reassign the time spent by employees in a firm on non-production tasks to the entrepreneur. The counterfactual
share of time in non-production tasks stays at 100% in firms of size larger than six workers.
87
the non-production time in the firm is supplied by the entrepreneur. However, in larger firms, most of the
non-production tasks are in fact done by employees: for instance, in firms of size eight, 70% of the time
spent by the firm in non-production activities is coming from employees. This is a direct consequence
of the fact the engagement of the entrepreneur in non-production tasks does not rise steeply with firm
size. This is a striking result: even though the entrepreneur is usually the most skilled individual in the
firm and non-production tasks are more complex, in larger firms most non-production tasks are done by
employees. In line with this, in Appendix Figure C.5 we use a separate set of survey questions on customer
interactions from the follow-up survey to show that in larger firms employees play a relatively larger role
in dealing with customer relationships, which are a key component of non-production tasks, as discussed
above. Through the model, we will be able to quantify the implications for firm size and productivity of
this misallocation of labor inside the firm.
Laborspecializationwithinproductionbetweensteps
Finally, we look at specialization within production, across steps, to study whether this margin of divi-
sion of labor is also meaningful. In Figure 3.5, Panel (a) and (b), we look at the share of employees and
entrepreneurs performing a production step. To do so, for each firm we compute the share of employees
performing each production step. We then aggregate across steps weighing by the time intensity of the
step, following a procedure similar to Figure 3.2, Panel (a). This creates a measure of the average share of
employees performing a representative production step. We do the same for entrepreneurs to create the
share of entrepreneurs performing a production step.
Starting from employees (Panel (a)) we see that the share of employees working on a step is high
and barely decreases with firm size – even in firms of size 8-10, more than 70% of employees work on
the representative production step. To better interpret this magnitude, we build an empirical benchmark
corresponding to the share of employees that would work on a production step under full specialization.
88
Figure 3.5: Task Allocation within Production across the Size Distribution
(a) Employees
0
.2
.4
.6
.8
1
Average Share of Employees Performing a Production Step
2 3 4 5 6 7 8 9 10+
Firm Size
Full Specialization
(b) Entrepreneurs
0
.2
.4
.6
.8
1
Share of Entrepreneurs Working on a Production Step
2 3 4 5 6 7 8 9 10+
Firm Size
(c) High & Low Earning Employees
0
.2
.4
.6
.8
1
Share of Production Time on Complex Tasks
1 2 3 4 5 6 7 8 9 10+
Firm Size
High earnings Low earnings
(d) Entrepreneurs
0
.2
.4
.6
.8
1
Share of Production Time on Complex Tasks
1 2 3 4 5 6 7 8 9 10+
Firm Size
Notes: Figures 3.5 (a) and 3.5 (b) depict the share of employees and entrepreneurs (respectively) working on a production step. In
each firm we calculate the share of persons working on each production step performed, and construct an average measure over
steps, weighing by their time intensity in the production process. The red diamond markers in Figure 3.5 Panel (a) represent the
share of employees that would work on a production step under full specialization (reported for firms of size 6 and 10 only). To
compute this, we simply reassign employees across production steps to minimize the overlap in steps between employees, while
keeping the total amount of time worked by employees in production within the firm fixed. Figures 3.5 (c) and 3.5 (d) represent
the share of worker and entrepreneur production time spent on complex production steps. Production steps are categorized as
complex (or simple) using employee assessments of their ability to perform each production step conducted by the firm on a 1-5
scale. These measures were collected for all employees, regardless of whether they actually work on the production step. We
aggregate across employee reports in each sector, to compute step-level measures of difficulty. Steps that are ranked as having
above median levels of difficulty are regarded as ‘complex’.
To do so, we simply reassign employees across production steps to minimize the overlap in steps
between employees, while keeping the total amount of time worked by employees in production within
the firm fixed. We report this benchmark for firms of size six and ten. Comparing the actual allocation
with the full specialization benchmark highlights that this type of specialization is indeed very limited, and
does not increase with firm size. The same analysis for entrepreneurs (Panel (b)) reveals a similar pattern:
89
as we have shown that entrepreneurs are less likely to work on production (Figure 3.3), we naturally find
that the share of entrepreneurs working on the typical step is lower than for employees. However, again
there is no strong evidence of specialization increasing with firm size (i.e., the gap between the share of
employees and entrepreneurs performing the typical step is constant in firm size).
Laborspecializationbetweencomplexandsimplesteps. Finally, we narrow in on vertical special-
ization across production steps, by exploiting a survey question where each employee was asked to rank
(on a 1 to 5 scale) their ability to perform each production step conducted by the firm (regardless of whether
the particular employee performs that step). We use this information to rank production steps and then we
split them by above/below median complexity. In Figure 3.5, Panels (c) and (d), we study how employees
and entrepreneurs allocate their production time to simple and complex steps. If an individual only works
on complex steps, the share of time in complex steps would be 100%. The Figure shows that: (i) higher
skilled employees are more likely to work on complex steps than low skilled employees, but the gap be-
tween the two groups of employees is small and completely flat with firm size; (ii) entrepreneurs spend
slightly more time on complex steps, but again their share of time in complex tasks is close to 50% and
there is no gradient with firm size. We conclude that while there is some evidence of vertical specialization
also within production activities, there is no organizational change with firm size in this dimension.
Islaborspecializationanimportantreasonwhyfirmsexist? Taken together, the evidence in this
section shows that the division of production tasks between individuals within the firm is not a main
reason why firms exist: specialization across production steps is limited, both among employees, and
between employees and entrepreneurs, and does not vary with firm size. On the other hand, we find
more evidence of vertical specialization between production and non-production tasks: larger firms allow
entrepreneurs to leverage their talent by specializing in more complex non-production tasks. This type
of vertical specialization is therefore a more important reason why individuals get together in a firm.
90
Still, even this type of specialization is far from perfect. Using the model will allow us to more precisely
quantify the limits to specialization, and the resulting implications for aggregate productivity and the size
distribution of firms.
3.3.4 SectoralHeterogeneityandBarrierstoSpecialization
Finally, we study whether these results on labor specialization vary across our three sectors. This allows
us to shed light on potential barriers to labor specialization faced by firms in this context.
Table 3.3: Heterogeneity in Labor Specialization by Sector
Carpentry Welding Grain
milling
(1) (2) (3)
PanelA.AcrossFirmSpecializationinProd. Steps
Avg. share of firms doing a step 0.870 0.839 0.816
Slope of share of firms doing a step with firm size -0.001 -0.003 0.053
PanelB.SpecializationinProductionvs. Non-prod.
Avg. diff. in entr. and empl. share of time in non-prod. 0.319 0.352 0.618
Slope of share of empl. time in non-prod. with firm size 0.004 -0.007 0.019
Slope of share of entr. time in non-prod. with firm size 0.025 0.016 0.063
PanelC.SpecializationwithinProductionSteps
Avg. share of empl. performing a step 0.831 0.880 0.705
Slope of share of empl. performing a step with firm size -0.023 -0.019 -0.064
Avg. share of entr. performing a step 0.686 0.647 0.306
Slope of share of entr. performing a step with firm size -0.012 -0.024 -0.054
Notes: Sample: surveyed firms and employees. The rows starting with “slope” report the OLS coefficient of a regression of the
corresponding variable on firm size. All statistics and regressions are weighted with sampling weights.
Heterogeneityinlaborspecializationacrosssectors. In Table 3.3 we summarize the main statistics
and relationships by sector. Panel A reports the average share of firms doing the typical production step,
and is slope with firm size. While carpentry and welding are similar, we find some evidence of higher
specialization across firms in grain milling, where smaller firms specialize in fewer steps. In panel B instead
we look at specialization between production and non-production steps. We find a striking pattern. While
91
carpentry and welding are almost identical, in grain milling there is significantly higher specialization:
the difference in the share of time spent in non-production by employees and the entrepreneur is twice as
large in grain milling, and the slope of specialization with firm size is also much larger. We find a similar
pattern in Panel C, which looks at specialization within production across steps: the share of employees
performing the typical step is lower in grain milling, and it decreases more steeply with firm size; the same
is true when looking at entrepreneurs. Carpentry and welding instead continue be show almost identical
numbers.
This result is important because it shows that specialization is possible in this setting, thus suggesting
that institutional features such as contract enforcement or lack of trust are not the only drivers of lack of
specialization. This heterogeneity also reassures us that the limited specialization uncovered on average
in this section is not confounded by measurement error, as the measurement of time use was exactly the
same across the three sectors.
Table 3.4: Descriptives on Customization and Customer Interactions by Sector
Carpentry Welding Grain
milling
(1) (2) (3)
PanelA.ProductCharacteristics
80-20 price dispersion for main product (Std.) 1.46 1.61 1.15
Charges different prices due to customization 42% 52% 17%
Customers buy on order to customize products 65% 65% 26%
Customers buy on order to bring own inputs 5% 5% 52%
PanelB.ProcessCharacteristics
Days to complete typical order 4.0 4.0 0.6
Customer pays fully upfront (quality uncertanty) 26% 30% 54%
Potential number of machine types for main product 24 20 13
PanelC.Worker-CustomerInteraction
Customers buy on order to discuss details with person producing 52% 48% 22%
Workers perform independent orders 47% 49% 30%
Notes: Means are reported. Sample: surveyed firms. All statistics are at the firm level, weighted by sampling weights.
92
Barriers to specialization. In Table 3.4, we leverage our detailed data on the output market and cus-
tomer interactions to explore differences in product characteristics and demand across the three sectors.
We find again striking differences across sectors. Panel A shows evidence suggesting that in carpentry
and welding products are highly customized, while in grain milling the output is more standardized. For
instance, consistent with customization and quality differences across firms, we find substantial residual
dispersion in the price of the main product in carpentry and welding, even after accounting for sub-county
fixed effects. Price dispersion is instead more limited in grain milling. Also, when asked about why they
charge different prices for the same product, about half of the firms in carpentry and welding report that
customization is an important reason for this, while the corresponding share is only about one fourth in
grain milling. In addition, firms in carpentry and welding report that the main reason why customers buy
on order is to customize products, while in grain milling the main reason for doing so is that customers
bring their maize to the firm, to be processed into flour.
In line with customization and quality uncertainty being more important in carpentry and welding,
Panel B shows that in carpentry and welding: (i) it takes much longer to complete an order, (ii) it is less
common for customers to pay fully upfront and (iii) firms use many more machines for the production
of the same type of product. Throughout Panel A and B, the similarity between welding and carpentry is
remarkable.
The customization of production may entail significant communication and coordination costs within
the firm, as the person producing the order must have clear the specific product characteristics demanded
by the customer. This may make it more costly for the firm to “unbundle” the production process into
separate tasks that can be performed by different specialized individuals. In line with this, Panel C shows
that in carpentry and welding it is much more common that the reason why customers buy on order is
that they want to discuss the details of the product with the person producing. Also, in these sectors it is
93
more common that employee perform independent orders, whereby the employee manages production as
well as the entire relationship with the customers as if they were self-employed.
Taken together, the evidence in Table 3.4 shows that lack of standardization can be an important reason
behind the more limited specialization in carpentry and welding.
3.4 Model
Motivated by the data, we now develop a simple model. We use it in this Section to formalize the two-way
relationship between labor specialization and firm size and show how barriers to specialization impact
firm-level and aggregate productivity. In the next Sections, it will then provide a framework to quantify
the extent of specialization costs in our setting as well as their implications for firm size, productivity, and
selection into entrepreneurship.
We consider a static, closed economy with one sector, manufacturing. This should be interpreted as
either carpentry, welding, or grain milling.
21
Each agent chooses whether to be a worker or start a firm.
Each firm owner chooses employment as well as the allocation of workers to tasks, thereby determining
firm productivity.
3.4.1 Environment
We first describe the environment of our economy.
Agents and Demographics. The economy is populated by a measure 1 of agents who differ in their
ability in the manufacturing sectorz ∈ [0,z
max
], which we refer to simply as ability. The distribution of
ability in the population is given by G(z). Agents supply one unit of labor inelastically and have linear
utility over consumption.
21
In the quantitative model, we allow for an outside sector toi appropriately capture the rest of the economy.
94
Individuals can start a firm and become entrepreneurs (owners) o, or be employed as workers in the
manufacturing sectorw or an outside sectorx. The distribution of ability in the two occupations is given
byF
o
andF
w
withG(z)=F
o
(z)+F
w
(z). The equilibrium share of workers is then given byF
w
(z
max
).
Aggregate Output. Manufacturing firms produce differentiated varieties, which are combined using a
CES aggregator to produce outputY . The elasticity of substitution across individual varieties is given by
γ .
Y =
Z
Y(z)
γ − 1
γ dF
o
(z)
γ − 1
γ (3.1)
We normalize the aggregate price index to 1,P =
R
p(z)
1− γ dF
o
(z)
1
1− γ = 1 , so that cost minimization
implies
Y(z)=Y p(z)
− γ (3.2)
Labor Market. There is a spot market for labor. Importantly, ability z of individuals – owners and
employees alike – is private information at the time of hiring. Therefore, there is a single labor market
that randomly matches owners and employees. When a firm owner chooses employments, she choose
only the mass of workers, the composition is given by the distribution of abilityF
w
(z).
Upon being matched, employee and owner productivity are publicly observed, and the two parties
agree on a wage using a standard Nash Bargaining framework. The worker’s outside optionw is endoge-
nous and adjust to clear the labor market. The owner’s outside option are her profits when producing with
one fewer worker. The wage of an employee therefore depends both on his own ability and the ability of
the entrepreneur he works for.
95
Manufacturing Production. Total output of a firm is equal to the sum of the output produced by ev-
eryone in the firm – the owner ˆ z as well as her mass of n− 1 workers whose distribution of ability is
F
w
(z)/F
w
(z
max
). We refer to the output produced by one member of the firm as their production line.
Each production line takes one unit of time, i.e. one worker, to complete and consists of a fractionD of
complex tasks. The remainder 1− D are the simple tasks. In the context of carpentry, simple tasks are
sweeping the floor or cutting the wood, whereas complex tasks include negotiating with customers and
suppliers, or working on the finishing of a door.
In order for the production line to yield any output, all simple and complex tasks must be completed.
The value of each production line is a function of two components. First, it depends on the ability of the
firm owner, ˆ z, capturing the quality of the “idea” (the blueprint), or also the reputation of the shop. The
contribution of entrepreneurial ability is non-rival: everybody’s output benefits from the ability of the
entrepreneur, irrespective of who does what within the firm.
Second, the value of the output depends on net task productivity, which is a measure of the average
ability levelz with which the complex tasks have been performed. In the absence of labor specialization,
this is simply equal to the individualz orˆ z corresponding to the production line. With labor specialization,
tasks within each production line can be traded. For instance, the firm owner could take over negotiating
with customers while her employee cuts the wood for his as well as the owner’s door. In that case, the
employee’s net task productivity would be a function of the owner’s ability as well.
We refer to it as net task productivity since trading tasks is costly. In order to spread different parts
of a production process across multiple individuals, tasks must beunbundled. For example, if the owner is
the one negotiating with customers on all orders, she must then communicate exactly what the customer
wants to the employee. The cost of unbundling depends on the share of tasks in each production line that
are delegated to others.
96
Task assignment in the firm is summarized by µ , which consists of 4 functions.µ c
(z,z
′
) andµ c
(z,z
′
)
specify – for all pairs of employees{z,z
′
} in the firm – the fraction of time z
′
spends onz’s complex (c)
and simple (s) tasks. The other two functions, ˆ µ c
(z) and ˆ µ s
(z) specify the share of time the owner of the
firm spends on simple and complex tasks of each of her employees z.
Formally, the output of a firm of size n, owned by an individual with abilityˆ z and with task assignment
µ is given by Equation (3.3). It covers the most general setup and is therefore heavy on notation. We include
an empirically relevant special case that provides all the economic intuition right below.
Y(ˆ z,n,µ )=y(ˆ z,ˆ z,µ )+(n− 1)
Z
y(z,ˆ z,µ )
dF
w
(z)
F
w
(z
max
)
(3.3)
where
y(z,ˆ z,µ )= ˆ z
λ ˜ z(z,ˆ z,µ )
1− λ I
y(z,ˆ z,µ )
˜ z(z,ˆ z,µ )=exp

µ C
(z,ˆ z)
D
log(ˆ z)+(n− 1)
Z
µ C
(z,z
′
)
D
log(z
′
)
dF
w
(z
′
)
F
w
(z
max
)

1− κ
µ C
(z,z)
D

I
y(z,ˆ z,µ )
=I
[ˆ µ C
(z,ˆ z)+(n− 1)
R
µ C
(z,z
′
)
dFw(z
′
)
Fw(z max)
≥ D]
I
[ˆ µ S
(z,ˆ z)+(n− 1)
R
µ S
(z,z
′
)
dFw(z
′
)
Fw(z max)
≥ 1− D]
whereλ is the relative importance of the non-rival entrepreneurial idea andκ (.) is the unbundling cost.
The indicator function simply enforces that for each production line, allD complex and1− D simple tasks
must be completed.
97
In order to be feasible, an assignment must satisfy the time constraint of each individual in the firm.
No one can spend more than their time endowment – one unit of time – on any tasks, simple or complex,
their own or others’. Formally, an assignmentµ is feasible if and only if:
∀z : µ C
(ˆ z,z)+µ S
(ˆ z,z)+(n− 1)
Z

µ C
(z
′
,z)+µ S
(z
′
,z)

dF
w
(z
′
)
F
w
(z
max
)
≤ 1 (3.4)
ˆ µ C
(ˆ z)+ ˆ µ S
(ˆ z)+(n− 1)
Z

ˆ µ C
(z
′
)+ ˆ µ S
(z
′
)

dF
w
(z
′
)
F
w
(z
max
)
≤ 1 (3.5)
Asimplecase. In order to more clearly lay out the intuition behind the assignment problem within the
firm, we now consider a simplified version of the problem. In particular, we make two assumptions. First,
the firm owner is the highest skilled individual in the firm. This will be true in equilibrium as we show
below. Second, the owner’s time constraint is slack, meaning she optimally spends less than her full unit
of time on her and her employee’s complex tasks. This is a joint assumption on parameters of the model.
It is motivated by the data, where we see that owners spend a significant fraction of their time on simple
tasks.
When the owner is more skilled than her employees and not time-constrained, she is the only one who
takes on complex tasks of others. To see this, note first that production line output is linear in the average
of the abilities z of all those performing complex tasks. Given the linearity, it is always better to assign
complex tasks to the most skilled individual. Second, the unbundling cost only depends on µ c
(z,z), i.e.
what fraction of complex tasks are done by the individual herself rather than unbundled. In particular, the
cost does not depend on who these tasks are performed by once they are unbundled.
The assignment problem therefore reduces to choosingµ c
(z,z)≡ µ (z), the share of time each worker
spends on her own complex tasks. The remainder is done by the owner, who also performs all of her own
complex tasks, ˆ µ (ˆ z) = D. Suppressing the indicator function for simplicity, we can rewrite the output of
each production line as
98
y(z,ˆ z,µ )= ˆ z
λ |{z}
non-rival


z
µ (z)
D
ˆ z
1− µ (z)
D
| {z }
task productivity
[1− κ (µ (z)/D)]
| {z }
unbundling cost



1− λ (3.6)
The output produced by a worker z is a geometric mean of the ability of the firm owner ˆ z and the
net task productivity (task productivity inclusive of any unbundling cost). Whenλ = 0, there is perfect
symmetry between owners and workers in the production function. The bigger is λ , the more the non-
rival entrepreneurial idea matters for productivity. Net task productivity is itself a geometric mean of the
worker’s and the owner’s ability. This time, the weight on the owner is endogenous, it is the share of
complex tasks taken over by the owner.
Aggregating to the firm-level, we can re-write Equation (3.3) as
Y(ˆ z,n,µ )=nZ(ˆ z,n,µ ) (3.7)
where
Z(ˆ z,n,µ )= ˆ z
λ
1
n
ˆ z+
n− 1
n
Z
z
µ (z)
D
ˆ z
1− µ (z)
D
[1− κ (µ (z)/D)]
dF
w
(z)
F
w
(z
max
)

1− λ 3.4.2 Choices
We next turn to the choices that economic agents in this model make. Entrepreneurs, or firm owners,
choose how many workers to hire, assignment of all individuals to tasks, as well as output and the price
to maximize profits.
π (ˆ z)= max
{µ,n ≥ 1}
(1− τ )pY(ˆ z,n,µ )− (n− 1)
Z
w(z,ˆ z,µ )
dF
w
(z)
F
w
(z
max
)
(3.8)
s.t.(3.2), (3.3), (3.4), (3.5)
99
The revenue wedge τ is a stand-in for any other frictions or features of the market that might keep
firms small, such as credit constraints, labor market frictions, transport costs, or variable markups.
Occupational Choice Each agent observes their ability z and chooses whether to be a worker or an
entrepreneur. Workers make no further choices since there is random matching in the labor market and
tasks within the firm are assigned by entrepreneurs. Profits conditional on entering are known, since firm
owners hire a representative sample of workers. Wage earnings on the other hand depend on who the
worker happens to match with. An individual with abilityz therefore starts a firm if and only if her profits
are higher than her expected wage in the labor market
π (z)≥ Z
w(z,ˆ z,µ )
dF
o
(ˆ z)
F
o
(z
max
)
(3.9)
3.4.3 Equilibrium
Finally, we describe the notion of equilibrium in our setting, which simply requires that all agents maximize
and that the wage level is such that the total labor demanded by entrepreneurs is equal to the mass of
individuals choosing not to start a firm – i.e. the labor market clears.
Definition of Competitive Equilibrium The competitive equilibrium is given by a wage levelw, firm
sizesandoutputpriceforeachfirmownerabilitytype {p(z
′
),n(z
′
)}
∀z
′,taskassignmentchosenbyeachfirm
ownerz
′
foreachpair(z,z
′′
)ofindividualsinthefirm(includingthemanagerherself) {µ (z,z
′′
;z
′
)}
∀(z,z
′
,z
′′
)
,
occupational choice functionI
o
(z), and distributions
Fo(z)
Fo(zmax)
,
Fw(z)
Fw(zmax)
such that:
1. firm owners choose firm size and task assignment to maximize profits solving (3.8);
2. individuals choose their occupation according to (3.9);
3. output prices for each firm is p(ˆ z)=

y(ˆ z)
Y
1
γ 100
4. the labor market clears:
P
r
P
s
R
n(z)g
o
(z,r,s)dz =
P
r
P
s
R
g
w
(z,r,s)dz;
5.
Fo(z)
Fo(zmax)
,
Fw(z)
Fw(zmax)
areconsistentwiththeoccupationalchoice–i.e.F
w
(z)=
R
(1− I
o
(z))dG(z)and
F
o
(z)=
R
I
o
(z)dG(z)
3.4.4 Characterization
We study the allocation of talent within and between firms, how it is affected by the unbundling cost κ (.),
and how it shapes firm size and productivity. Towards this goal, we first describe the occupational choice,
then analyze the within firm organization and its implication for firm size and productivity. Finally, we
focus on the overall equilibrium of the economy. All proofs of the results are relegated to the Appendix.
Assumptions. Throughout this section, we work under two assumptions.
Assumption 1. The cost to unbundle a share x of complex tasks and delegate them is κ (x) = 1− exp{− ˆ κ (x)}, where ˆ κ (x)=κ 0
1/κ
1
x
1+1/κ
1
(1+1/κ
1
)
.
Assumption2. The parameters of the model are such that in each firm ˆ z the entrepreneur spends at least
some time on simple tasks: ˆ µ S
(ˆ z)≥ 0∀ˆ z.
Assumption 1 is useful to provide simple closed-form solutions and to parameterize the unbundling
cost by a key parameter of interest,κ 0
, which modulates the size of the cost. Assumption 2 is motivated
by the empirical evidence showing that even the entrepreneurs in the largest firms spend some of their
time performing simple tasks. It is a joint assumption on the parameters of the model, which simplifies
the exposition in this section, but will be relaxed in the quantitative analysis.
22
Occupational Choice. The model yields a familiar sorting of talent into occupations as a function of
their skill-sensitivity.
22
In the estimated model in the next section, this assumption will hold at our estimates, but not necessarily in the counterfac-
tuals.
101
Lemma1 (Occupational Choice). Thesolutiontotheoccupationalchoiceproblemisgivenbyacutoff z
0
such
that an individualz chooses to become an entrepreneur if and only ifz≥ z
0
.
Lemma 1 shows that talent is segmented by occupation. The reason is that while the earnings of both
workers and entrepreneurs are increasing in their abilityz, the relationship is steeper for entrepreneurs.
For entrepreneurs, their ability not only affects their own output, but also the one of their workers (as long
asλ > 0 andκ 0
<∞). For workers instead, their wages are a function of both their ability, and the one
of the owners they are matched with.
Lemma 1 implies that the entrepreneur is, in all firms, the most skilled individual. This result simplifies
the analysis of the within-firm allocation of talent to which we turn next.
Labor Specialization. The entrepreneur allocates labor to tasks within the firm to maximize output.
Given the properties of the production function, the key margin is the choice of which individuals to
assign to complete the complex tasks, which determine the firm-level productivity.
In making this assignment decision, the entrepreneur faces a simple trade-off between the productivity
gains from allocating individuals based on their comparative advantage – i.e. of having the complex tasks
assigned to the most skilled individuals – and the cost of unbundling and delegating the relevant tasks.
Consider an individual with abilityz. The benefit of delegating a marginal complex task from z to any
individual with abilityz
′
> z is given by the percentage difference in abilities: logz
′
− logz.The cost is
increasing in the overall share of complex tasks that are unbundled and delegated byz – i.e. in the share
of tasks not performed byz himself: 1− µ C
(z,z)
D
. At the optimum, the marginal benefits and costs must
offset each other, giving the condition
logz
′
− logz
| {z }
marginal benefit if complex task is assigned to z
′
= κ 1/κ
1
0
(1− µ C
(z,z))
1/κ
1
| {z }
marginal cost of delegating complex tasks
, (3.10)
which must hold for anyz
′
who performs complex tasks forz.
102
Equation 3.10 highlights a key property of the solution. Since the unbundling cost depends only on the
total share of tasks individual z has delegated, but not on the identity of the individualz
′
to whom these
tasks are delegated, it is always optimal to allocate all the complex tasks to the most skilled individuals in
the firm.
23
Due to Lemma 1, the entrepreneur is the most skilled individual in the firm. Therefore, all the
"unbundled" complex tasks should be assigned to her as long as she has enough time to complete them,
which is guaranteed by Assumption 2.
The potentially complicated assignment problem has therefore a very simple solution, which high-
lights the specific nature of labor specialization in our model. Specialization is purely along the vertical
dimension and leads the most skilled individual in the firm, the entrepreneur, to specialize on the complex
tasks.
This feature of our model is consistent with the evidence presented in the empirical Section 3.3. While
we have considered a rich set of possibilities, the data pointed towards vertical specialization between
employees and entrepreneurs as the most salient one. Moreover, it also implies that the level of labor
specialization within the firm can be simply summarized by the share of complex tasks that are assigned
to the entrepreneur, as formalized in the next definition.
Definition 1 (Average Labor Specialization). Let the total times spent on complex tasks by the en-
trepreneur ˆ z and by one of her workerz be
ˆ
θ (ˆ z)≡ ˆ µ C
(ˆ z,ˆ z)+
(n− 1)
Fw(zmax)
R
ˆ µ C
(ˆ z,z
′
)dF
w
(z
′
) andθ (z,ˆ z)≡ µ C
(z,ˆ z)+
(n− 1)
Fw(zmax)
R
µ C
(z,z
′
)dF
w
(z
′
). Further, let the average time spend on complex tasks across all
employees beθ w
(ˆ z)≡ 1
Fw(zmax)
R
θ (z
′
,ˆ z)dF
w
(
′
). We define the average labor specialization in the firm,
θ (ˆ z), to be the difference between the time spent on complex tasks by the entrepreneur and the average
employee:
θ (ˆ z)
|{z}
specialization
≡ ˆ
θ (ˆ z)
|{z}
entrepreneur
− θ w
(ˆ z)
|{z}
employees
.
23
This result holds because the productivity benefits of delegating are increasing in z
′
.
103
Equipped with this definition, Lemma 2 summarizes the discussion so far by describing the extent of
labor specialization and its relationship with firm size.
Lemma2 (Labor Specialization). The solution to the profit maximization problem for an entrepreneur ˆ z of a
firm size n gives
1. The time spent on complex tasks by a worker of abilityz and by the entrepreneur ˆ z are
θ (z,ˆ z)=D

1− 1
κ 0
(logˆ z− logz)
κ 1

(3.11)
ˆ
θ (ˆ z)=D

1+
n− 1
κ 0
F
w
(z
max
)
Z
(logˆ z− logz)
κ 1
dF(z)

. (3.12)
2. The average labor specialization is then
θ (ˆ z)=D

n
κ 0
F
w
(z
max
)
Z
(logˆ z− logz)
κ 1
dF(z)

(3.13)
which is declining in the unbundling costκ 0
, and increasing in firm size n at a rate which decreases in
κ 0
.
Lemma 2 is visualized in the left panel of Figure 3.6 which plots the time on complex task as a function
of individual ability for different values of κ 0
. As the figure shows, the smaller is κ 0
, the more complex
tasks are concentrated towards the entrepreneur.
Figure 3.7 shows the relationship between firm size and labor specialization for the case in which
κ 1
→ ∞, hence when the time spent on complex tasks is identical for all employees, irrespective of
the ability gap with the firm entrepreneur. This is a useful tractable benchmark which highlights the
tight connection between our theoretical model and the empirical analysis. The model shows that the
relationship between the time spent in the (more complex) non-production tasks and firm size is the key
moment to pin down the size of the unbundling costκ 0
. When this cost is small, each new employees brings
104
a lot of "unbundled" complex tasks that could be assigned to the entrepreneur, and thus labor specialization
increases steeply with firm size.
The figure also highlights that, in our model, there is a causal relationship from firm size to labor
specialization. The model thus allows for firms in low income countries to be not specialized because they
are too small. In small firms, in fact, there are simply not enough complex task to specialize, and most
individuals spend similarly their time doing simple, production tasks. In larger firms, instead, there are
more "low-hanging" complex tasks that can be unbundled and delegated to the more skilled entrepreneurs.
FirmProductivity. The assignment of talent to tasks and the resulting labor specialization, determine,
together with the distribution of talent, the productivity of the firm.
Lemma3 (Firm Productivity). The output of a firm of size n isY(ˆ z,n)=Z(ˆ z,n,µ )n where
Z(ˆ z,n,µ )= ˆ z
λ |{z}
non-rival
˜
Z(ˆ z,n,µ )
1− λ | {z }
aggregate task-productivity
˜
Z(ˆ z,n,µ )=





1
n
ˆ z
1− λ +
n− 1
n
Z
˜ z(z,ˆ z,µ )
1− λ dF(z)
| {z }
dilution from firm size





1
1− λ ,
andZ(ˆ z,n,µ ) is strictly declining inn as long as bothλ< 1 andκ 0
>0.
Lemma 3 shows the key forces at play. The aggregate firm productivity inherit the same structure,
previously discussed, of each individual production line. It has two components, with relative weights
given byλ : i. a non-rival component ˆ z
λ , which captures the unique role of entrepreneurial talent and is
independent from the allocation of labor to task; ii. the aggregate task-productivity
˜
Z(ˆ z,n,µ )
1− λ , which
is a weighted average of the ability of the individuals completing the complex tasks, and thus a function
of the assignment.
105
This second component encapsulates the roles of labor specialization and firm size in productivity. As
long as talent is partly rival and there is not full specialization, the task-productivity of each employee is
below the one of the entrepreneur – i.e. ˜ z(z,ˆ z,µ ) < ˆ z. As a result, increasing the size of the firm would,
ceteris paribus, decrease its productivity since it gives a larger weight to the less productive workers.
This mechanism generates decreasing returns to scale originating from the limited internal specialization
and from the fact that, consistent with results shown in the empirical Section, in large firms employees
complete a bigger share of the skill-sensitive production tasks. This inability to efficiently leverage the
entrepreneurial talent stems from the limited specialization, and, as such, it is more severe the larger is the
unbundling cost (largeκ 0
) and the stronger is the role of "rival" talent for productivity (lowλ ).
Optimal Firm Size. Labor specialization and firm size are closely intertwined. Lemma 2 showed one
side of this two-way causal relationship: small average firm size limits labor specialization. Lemma 4 shows
that the opposite is also true: barriers to labor specialization make firms smaller.
Lemma4 (Firm Size). The optimal firm size n of each entrepreneur ˆ z solves
(1− τ )
"
∂p(n)
∂n
Z(ˆ z,n,µ )n+p(ˆ z)Z(ˆ z,n,µ )
| {z }
change in revenues
+p(ˆ z)
∂Z(ˆ z,n,µ )
∂n
n
| {z }
prod. dilution<0
#
=
∂
∂n
h
(n− 1)w(ˆ z,µ,p (n))
i
| {z }
change in labor cost
,
and it is declining in the wedgeτ and, as long asλ< 1, in the unbundling costκ 0
.
As usual, profit maximization implies optimal firm size is such that the marginal labor cost from an
additional worker is equal to the marginal revenue gain. The marginal cost is relatively standard and
simply a function of the total wage bill.
24
The first component of the marginal benefit is also standard and
given by the change in firm revenues, which takes into account the price effect. The second component,
instead, is unique to our framework. As shown in Lemma 3, firm productivity is diluted by each additional
24
Notice one, possibly unusual, feature. In our model, the labor cost depends on firm size itself since part of wage payment is
a piece-rate and thus a function of firm level output person, which depends on size through the price.
106
worker as they are less skilled than the entrepreneur. In choosing firm size, the entrepreneur takes into
account this source of decreasing returns to scale.
There are two kinds of frictions that can keep firms small. The wedge τ reduces firm size directly by
lowering firm revenues. The unbundling cost κ 0
reduces firm size both through the production dilution
term just discussed and since it reduces average productivity for any firm size, as Lemma 3 shows. There
is also an apparent complementarity between the frictions. The returns from relaxing the external wedge
τ are limited if internal barriers to labor specialization hinder firm productivity and generates strong de-
creasing returns to scale.
WhyDoFirmsExist? TwoPolarCases. Before turning to the aggregate implications, we describe in
Lemma 5 two polar firm organizations that can emerge in our model.
Lemma 5 (Self-Employment within the Firm). Depending on the weight of non-rival entrepreneurial talent
in production (λ ) and the size of the unbundling cost (κ 0
) the model spans two polar firm organizations
1. ScalableEntrepreneurialTalent. Ifλ = 1 orκ 0
= 0; thenY(ˆ z,n,µ ) = ˆ zn, and optimal firm size
is increasing in ˆ z.
2. Self-Employment Within the Firm. If λ = 0 and κ 0
→ ∞; then Y(ˆ z,n,µ ) = z(ˆ z) n, with
z(ˆ z)≡ 1
n

ˆ z+
n− 1
Fw(wmax)
R
zdF
w
(z)

, and optimal firm size is constant for all ˆ z.
One extreme is obtained when either delegation is free (κ 0
=0), so that there can be full specialization
of labor, or entrepreneurial talent is completely non-rival (λ = 1). This benchmark resembles the typical
firm problem as in Lucas ‘78, in which labor is a commodity and all workers inherit the productivity given
by the entrepreneurial ability. Firms are vehicles to leverage talent and the more skilled entrepreneurs
would have larger firms.
The opposite extreme, instead, considers the case when delegation is so costly (κ 0
−→ ∞ ) that all
individuals within the firm behave as if they are self-employed: everyone completes all the tasks needed
107
to produce the output and there is no scope for specialization of labor. If in addition talent is purely rival
– i.e. the productivity of production is purely a function of the ability of the individual performing the
complex tasks andλ =0 – then the productivity of the firm is simply the average ability of all its workers.
In this benchmark, the only reason why firms could exist is that they allow individuals to share the fixed
costs of entry. For this reason, all firms have identical size, and entrepreneurs have no means to leverage
their ability.
Overall, Lemma 5 shows that in our model, the notion of what a firm is and why firms emerge crucially
depends on the unbundling cost and the resulting internal organization of the firm.
EquilibriumandAggregateImplications. So far, we have considered the solution to the problem of
one entrepreneur. Next, we turn to the overall economy.
Proposition 1 (Aggregate Effects of the Unbundling Cost κ 0
). Suppose thatλ < 1, a decline inκ 0
yields
an increase of:
1. average labor specializationθ (ˆ z) for all ˆ z;
2. the slope of the relationship between average labor specialization and firm size;
3. the average firm size n≡ R
n(z)
dFo(z)
Fo(z
max)
, wheren(z) is the optimal firm size;
4. the average ability of firm owners: z
0
increases;
5. the average firm productivity Z≡ R
Z(z,n(z),µ (z))n(z)
dFo(z)
Fo(z
max)
;
6. the average wagew =
RR
w(z,ˆ z,µ )
dFo(ˆ z)
Fo(zmax)
dFw(z)
Fw(zmax)
;
7. the return to managerial ability:
∂Z(z,n(z),µ (z))
∂z
increases for allz.
Proposition 1 summarizes the effect of a decline in the unbundling cost κ 0
on the economy. When
the cost is high, the economy is made of many small firms, owned by low productivity managers, and
108
internally organized with limited specialization. The result is low aggregate productivity, low demand for
workers, and consequently low wages. The returns to managerial ability are limited as firm owners are
not able to leverage their talent.
Reducing the delegation cost transforms the way firms are organized internally with effects that ripple
through the economy in equilibrium. Higher labor specialization increases firm productivity, and thus the
demand for labor. As a result, wages increase, leading some marginal firm owners to become workers.
This further increases aggregate productivity through a classic selection effect. Overall, managerial ability
is highly priced in the economy, as talent can be leveraged by taking over more and more complex tasks.
Figure 3.6: Within-Firm Organization and Unbundling Costκ 0
(a) Task Allocation
θ(z,z′ )
z
z′ f(z)
Firm Owner
D
Workers
1
κ
0
∈ (0,∞)
κ
0
= 0
κ
0
→∞
(b) Firm Size
n(z′ )
z′
κ
0
∈ (0,∞)
κ
0
= 0
κ
0
→∞
(c) Firm Productivity
ℤ(z′ )
z′
κ
0
∈ (0,∞)
κ
0
= 0
κ
0
→∞
Figure 3.7: Labor Specialization, Firm Size, and Unbundling Costκ 0
(a) Workers Time in Complex Tasks
θ
w
(n)
2
n
D(1−
1
κ
0
)
κ
0
(b) Entrepreneurs Time in Complex Tasks
θ(n)
n
D(1+d)
1
κ
0
D
κ
0
109
3.5 BringingtheModeltotheData
Next, we bring our model to the data. As a first step, we use heterogeneity across sectors and regions to
validate some of the theoretical predictions of Section 3.4. We then extend and parameterize the model
to make it amenable to a quantitative analysis. Finally, we discuss identification and the results from the
estimation.
3.5.1 EmpiricalValidationoftheTheoreticalPredictions
We provide two qualitative tests to support the theoretical predictions of the model shown in Section 3.4.4.
AcrossSectors:TestingProposition1 Proposition 1 provides empirical predictions which are testable
if could find variation in the unbundling cost κ 0
. In the absence of credible exogenous and market-level
variation in κ 0
, we rely on cross-sectoral heterogeneity to at least provide supporting evidence. As dis-
cussed in Section 3.3, the larger extent of standardization in grain-milling suggests that the unbundling
cost is there lower, consistent with the higher documented labor specialization.
In Table 3.5, we show that also the other predictions of Proposition 1 hold. While carpentry and
welding are remarkably similar, Grain milling which is more specialized, has larger firms, larger returns
from managerial ability and a bigger skill gap between entrepreneurs and their employees.
25
25
The one prediction which we do not test is the average wage. The reason is that we do not observe in the data a wage-level
which is comparable to the one of the model. A simple comparison of average wage would not hold since employees in grain
milling are less skilled.
110
Table 3.5: Cross-Sectoral Hetoreogeneity to Test Proposition 1
All Carpentry Welding Grain
milling
(1) (2) (3) (4)
PanelA.AverageSpecialization&FirmSize
Specialization 0.31 0.27 0.29 0.54
Firm Size 5.8 5.5 5.9 7.0
PanelB.Reg. Coeff’sonMan. Ability(Std.)
Log Revenues 0.30 0.24 0.25 0.58
Log Revenues per Worker 0.18 0.14 0.15 0.35
Log Size 0.12 0.10 0.10 0.23
PanelC.Reg. Coeff’sonEntrepreneur(0/1)
Years of Education 1.34 1.41 -0.04 4.33
Age 12.0 9.9 11.6 18.9
Log Earnings 0.82 0.74 0.79 1.12
Notes: Panel A: Sample: all firms. Average specialization: the gap in the average share of time in non-production tasks between
entrepreneurs and employees. / Panel B: Sample: all firms. / Panel C: Sample: all interviewed entrepreneurs and employees.
Regressions for Panel B. and C. include region fixed effects and are weighted by sampling weights.
Across Regions: Testing Lemma 3 Our model has one relatively unique implication which is shown
in Lemma 3: all else equal, entrepreneurial talent should be less relevant in larger firms. The reason is that,
in larger firms, employees do effectively a large share of the "firm management". To test this prediction,
ideally we would like to find a credible instrument for firm size which operate at the regional level. In
the absence of such exogenous variation, we provide suggestive evidence exploiting heterogeneity across
sub-counties.
We next describe our procedure. We first drop all firms in grain-milling since we have shown that
they have different returns to managerial ability in Table 3.6.
26
We then calculate for each sub-county the
average firm size, and rank sub-counties from the one with smallest to the one with the largest firms on
average. We divide the group of sub-county in two groups with roughly equal number of firms.
27
. Finally
we compute, within each group of sub-counties, the return to managerial ability. Specifically, we regress
26
Results are not affected by this restriction, but we keep it to avoid discussing that sectoral composition across regions is not
driving our estimates.
27
The "marginal" sub-county is one of the biggest one, implying that we end up with 40% of the firms in one group, and 60%
in the other
111
log of revenues on sector dummies, and either the previously used managerial score, or simply the years
of education of the entrepreneurs.
The results for both sets of regressions are shown in Table 3.6. We notice that the set of sub-counties
with the smallest firms have substantially higher returns from managerial ability, however we measure it.
This result is consistent with Lemma 3. If firms are smaller, the talent of the entrepreneur has a bigger
impact on productivity. In large firms, instead, the average productivity is more a function of the workers
ability, rather than the entrepreneurial one.
Table 3.6: Returns to Managerial Ability in Locations with Different Firm Size
Dep. Var: (Log) Revenues
(1) (2) (3) (4)
Manager Ability (Std.) 0.388 0.177
(0.056) (0.041)
Yrs. of Education 0.060 0.036
(0.016) (0.012)
Subcounty by Firm Size Small Large Small Large
(Average Firm Size) (4.80) (6.15) (4.80) (6.15)
Sector FE Yes Yes Yes Yes
AdjustedR
2
0.152 0.044 0.081 0.029
Observations 360 583 360 583
Notes: OLS regression coefficients with carpentry and welding firms (grain milling excluded.) Robust standard errors in paren-
theses. Regressions are weighted by sampling weights. The 49 subcounties are classified into tertiles by the average size of the
interviewed firms.
3.5.2 Model’sExtensionandParameters
We extent the model along two dimensions. We introduce an external sector which captures the rest of
the economy, and we allow the revenue wedge to be a function of firm size, to capture size-dependent
distortion. Specifically, we use
τ (n)=τ 0
n
τ 1
112
.
We then assume that talent is distributed following a log-normal distribution, with mean normalized
to one and standard deviation given by σ z
. Finally, we allow for a cost of entry into entrepreneurship,
which captures both monetary and non-monetary values, and we allow each firm to need an overhead
amountd of non-production tasks which must be supplied by the entrepreneur.
With these assumptions, we are left with 11 parameters to be pinned down.
3.5.3 TargetedMomentsandIdentification
Our benchmark estimates use pooled data for carpentry and welding since we have shown those two
sectors are very similar. We also provide an alternative estimation targeting moments from grain-milling.
The model yields a structural relationship between firm size and the time spent on non-production
tasks by entrepreneurs and employees. Using our unique time-use survey, we can use this relationship to
pin down the unbundling cost. Given this cost, the joint distribution of firm sizes and revenues disciplines
the heterogeneity in entrepreneurial ability as well as the level of the revenue wedge and its relation-
ship with firm size. To quantify the importance of non-rival entrepreneurial talent, we use the relative
importance of working in a high productivity firm versus being a high skilled worker for wages.
3.5.4 EstimationResults
Estimation is still work in progress, and the results shown in the next Section are from roughly calibrated
quantitative explorations which help us to get a sense of the overall magnitudes.
Importantly, however, Figure 3.8 shows that the model is able to perfectly match the allocation of time
to production and non-production tasks, and how it varies across sectors.
113
Figure 3.8: Model Fit for Within-firm Allocation of Time to Tasks
2 4 6 8 10
Firm Size
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Managerial Time
Workers
Carpentry/Welding Model
Grain Milling Model
Carpentry/Welding Data
Grain Milling Data
2 4 6 8 10
Firm Size
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Managerial Time
Firm Owners
Notes: The figure shows the time-allocation to production and non-production tasks, as in panels (a) and (b) of Figure 3.4. It
includes both data (average for carpentry/welding and grain-milling), and the model predictions.
3.6 QuantifyingwithinFirmOrganization
We next use the estimated model to explore, from a quantitative perspective, how are firms organized in
our setting and what are the aggregate implications.
3.6.1 ClosetoSelf-EmploymentWithintheFirm
As first exercise, we consider the carpentry/welding calibration and vary the size of the unbundling cost
κ 0
from∞ to 0. In practice, this means that we are changing the allocation of talent within the firm
from a benchmark with no specialization at all – i.e. individuals are self-employed within the firm –, to
the opposite extreme with full specialization. This exercise is useful to quantify the aggregate role of the
unbundling cost and to assess how close is our estimated economy to the two benchmark cases discussed
in Lemma 5.
114
Figure 3.9: Aggregate Effect of Changing the Unbundling Cost κ 0
As we vary the unbundling costκ 0
, we recompute the equilibrium of the economy keeping all the other
parameters constant. We plot the (very preliminary) results of this exercise in Figure 3.9. We show the
impact on firm size, on aggregate consumption, on our usual measure of labor specialization, and on the
selection of the entrepreneurs relative to the employees. There are two features of the results worthwhile
to notice.
First, changes in the unbundling cost have large effects. Even just reducing κ 0
to change the within
firm organization from the level of specialization observed in carpentry and welding to the one of grain
milling would increase average firm size by more than one employee on average. Firms, therefore are, at
least in part, small due to lack of specialization.
115
Second, and possibly most important, even shutting down completely any specialization would not
reduce much firm size, nor output, relative to the baseline equilibrium. This result provides a quantitative
answer to the question: “Why do firms exist?”. Given the current (preliminary) estimates, our baseline
economy is very close to the polar case withself-employmentwithinthefirm . Therefore, we conclude that,
in our setting, the role of firms are vehicles to leverage talent is minor, and firms are merely a means to
share fixed production costs, such as the cost of the premise, or machines that can be shared by many
individuals.
3.6.2 ReturnsfromOtherInterventionsAreMuted
As a second exercise, we study the effect of changing the revenue wedge τ and show how it varies as
a function of the estimated unbundling cost. The idea of this exercise is to show that the return from
typical development interventions aimed at making firm grow, such as policies to relax credit or hiring
constraints, depends on how firms are internally organized since their internal organization ultimately
affects their returns to scale.
Specifically, we start from the two alternative model calibrations: the one for carpentry/welding and
the one from grain milling. We then reduce the revenue wedge τ to increase firm size. The (very pre-
liminary) results of this exercise are shown in Figure 3.10, where we include in four different panels the
same four statistics previously shown: size, consumption, specialization, and entrepreneurial section. We
normalize each statistic relative to its benchmark value.
The (preliminary) results are striking. Relative to our benchmark, calibrating the unbundling cost to
match the larger specialization observed in grain milling would almost double the effect of a reduction
in the revenue wedge. This result highlights a key takeaway of our work. Barriers to within-firm labor
specialization make traditional manufacturing a business model that is difficult to scale. As a result, the
116
returns from policy interventions aimed at spurring firm growth may be limited. The reason is that en-
trepreneurs face strong decreasing returns to scale coming from the inability to leverage their talent by
specializing on the most complex tasks.
Figure 3.10: Returns from a Reduction in the Revenue Wedgeτ 3.7 Conclusion
In this paper, we have used novel survey data on time use within Ugandan manufacturing firms, together
with a model, to study the extent of labor specialization and quantify its implications for firm size and
productivity.
117
Our key contribution is to uncover limited specialization of labor inside the firm as a significant source
of low returns to scale. Even though most firms in our data are large enough to potentially specialize
labor, they do so only to a limited extent: the internal organization of labor largely resembles that of self-
employed individuals sharing a production space. The limited specialization reduces the extent to which
entrepreneurs can leverage their talent by hiring more workers, thus yielding smaller and less productive
firms.
Our results shed new light on the role of entrepreneurs and firms in the economy. As we do find
some evidence of specialization between entrepreneurs and employees in managerial and production tasks,
respectively, our results suggest that vertical specialization is a more important reason why firms exist in
this context than horizontal or “Smithian” specialization of employees across production tasks. Overall
however, the extent of labor specialization is limited and the main reason why individuals get together
in a firm appears to be that firms allow individuals to share a premise and production costs. Our results
also shift the focus away from “who starts” a firm and towards “who does what” within the firm: as
entrepreneurs do not fully specialize in complex managerial tasks, the identity of the entrepreneur is not
as relevant for understanding the implications of the size distribution for the allocation of talent between
firms.
Our results also have important implications for the design of policies aimed at increasing firm pro-
ductivity and size in developing countries. They provide an explanation for why the returns to inter-
ventions incentivizing firms to hire more workers have shown low returns for the firm: the returns to
hiring additional labor are muted if firms cannot specialize the additional workers. Removing other ex-
ternal constraints, such as credit constraints, will also only have limited impacts if the organization is not
“scalable”. Therefore, finding ways to facilitate labor specialization and business scalability is likely to
increase the returns to removing external constraints. We have shown that customization of production
is a potentially important barrier to labor specialization in this context, which points to the importance of
118
promoting product standards as a possible policy lever. Providing additional evidence on what generates
barriers to the specialization of labor and what policy interventions can be effective at alleviating them
remains a promising area for future research.
119
Bibliography
[1] Arnstein Aassve, Giulia Fuochi, and Letizia Mencarini. “Desperate housework: Relative resources,
time availability, economic dependency, and gender ideology across Europe”. In: Journal of
Family Issues 35.8 (2014), pp. 1000–1022.
[2] Girum Abebe, Stefano Caria, Marcel Fafchamps, Paolo Falco, Simon Franklin, Simon Quinn, and
Forhad Shilpi. “Marching frictions and distorted beliefs: Evidence from a Job fair experiment”. In:
Department of Economics, Oxford University (mimeo) (2020).
[3] Achyuta Adhvaryu, Vittorio Bassi, Anant Nyshadham, and Jorge A Tamayo. No line left behind:
Assortative matching inside the firm . Tech. rep. National Bureau of Economic Research, 2020.
[4] Achyuta Adhvaryu, Anant Nyshadham, and Jorge A Tamayo. Managerial quality and productivity
dynamics. Tech. rep. National Bureau of Economic Research, 2019.
[5] Bina Agarwal. “Bargaining and gender relations: Within and beyond the household”. In: Feminist
economics 3.1 (1997), pp. 1–51.
[6] Ufuk Akcigit, Harun Alp, and Michael Peters. “Lack of selection and limits to delegation: firm
dynamics in developing countries”. In: American Economic Review 111.1 (2021), pp. 231–275.
[7] George A Akerlof and Rachel E Kranton. “Economics and identity”. In: The Quarterly Journal of
Economics 115.3 (2000), pp. 715–753.
[8] Begoña Álvarez and Daniel Miles. “Gender effect on housework allocation: Evidence from
Spanish two-earner couples”. In: Journal of Population Economics 16.2 (2003), pp. 227–242.
[9] Richard Anker. Gender and jobs: Sex segregation of occupations in the world. International Labour
Organization, 1998.
[10] Richard Anker, Helinä Melkas, Ailsa Korten, et al. Gender-based occupational segregation in the
1990s. Vol. 16. International Labour Office Genève, 2003.
[11] Nava Ashraf, Erica Field, and Jean Lee. “Household bargaining and excess fertility: an
experimental study in Zambia”. In: American Economic Review 104.7 (2014), pp. 2210–37.
120
[12] David Atkin, Azam Chaudhry, Shamyla Chaudry, Amit K Khandelwal, and Eric Verhoogen.
“Organizational barriers to technology adoption: Evidence from soccer-ball producers in
Pakistan”. In: The Quarterly Journal of Economics 132.3 (2017), pp. 1101–1164.
[13] Oriana Bandiera, V Bassi, R Burgess, I Rasul, M Sulaiman, and A Vitali. “Worker heterogeneity
and job search: Evidence from a six-year experiment in Uganda”. In: PEDL Research Paper (2020).
[14] Oriana Bandiera, Michael Carlos Best, Adnan Qadir Khan, and Andrea Prat. “The allocation of
authority in organizations: A field experiment with bureaucrats”. In: The Quarterly Journal of
Economics 136.4 (2021), pp. 2195–2242.
[15] Abhijit V Banerjee and Esther Duflo. “Do firms want to borrow more? Testing credit constraints
using a directed lending program”. In: Review of Economic Studies 81.2 (2014), pp. 572–607.
[16] Timothy J Bartik. “Who benefits from state and local economic development policies?” In: WE
Upjohn Institute for Employment Research (1991).
[17] Vittorio Bassi, Raffaela Muoio, Tommaso Porzio, Ritwika Sen, and Esau Tugume. “Achieving scale
collectively”. In: Econometrica 90.6 (2022), pp. 2937–2978.
[18] Vittorio Bassi, Tommaso Porzio, Ritwika Sen, and Esau Tugume. The impact of the COVID-19
lockdown on SMEs and employment relationships in Uganda. Tech. rep. International Growth
Centre Policy Brief UGA-20112, 2021.
[19] Vittorio Bassi and Imran Rasul. “Persuasion: A case study of papal influences on fertility-related
beliefs and behavior”. In: American Economic Journal: Applied Economics 9.4 (2017), pp. 250–302.
[20] Gary S Becker and Kevin M Murphy. “The division of labor, coordination costs, and knowledge”.
In: The Quarterly journal of economics 107.4 (1992), pp. 1137–1160.
[21] Alexandre Belloni, Victor Chernozhukov, and Christian Hansen. “Inference on treatment effects
after selection among high-dimensional controls”. In: The Review of Economic Studies 81.2 (2014),
pp. 608–650.
[22] Michele Belot, Philipp Kircher, and Paul Muller. “How wage announcements affect job search-a
field experiment”. In: CEPR Discussion Paper No. DP13286 (2018).
[23] Michele Belot, Philipp Kircher, and Paul Muller. “Providing advice to jobseekers at low cost: An
experimental study on online advice”. In: The review of economic studies 86.4 (2019),
pp. 1411–1447.
[24] Marianne Bertrand, Emir Kamenica, and Jessica Pan. “Gender identity and relative income within
households”. In: The Quarterly Journal of Economics 130.2 (2015), pp. 571–614.
[25] Julian R Betts. “What do students know about wages? Evidence from a survey of
undergraduates”. In: Journal of human resources (1996), pp. 27–56.
[26] David M Blau. “Search for nonwage job characteristics: a test of the reservation wage
hypothesis”. In: Journal of labor Economics 9.2 (1991), pp. 186–205.
121
[27] Francine D Blau, Peter Brummund, and Albert Yung-Hsu Liu. “Trends in occupational
segregation by gender 1970–2009: Adjusting for the impact of changes in the occupational coding
system”. In: Demography 50.2 (2013), pp. 471–492.
[28] Francine D Blau and Lawrence M Kahn. “The gender wage gap: Extent, trends, and explanations”.
In: Journal of economic literature 55.3 (2017), pp. 789–865.
[29] Nicholas Bloom, Benn Eifert, Aprajit Mahajan, David McKenzie, and John Roberts. “Does
management matter? Evidence from India”. In: The Quarterly Journal of Economics 128.1 (2013),
pp. 1–51.
[30] Nicholas Bloom, Aprajit Mahajan, David McKenzie, and John Roberts. “Why do firms in
developing countries have low productivity?” In: American Economic Review 100.2 (2010),
pp. 619–23.
[31] Tessa Bold, Selene Ghisolfi, Frances Nsonzi, and Jakob Svensson. “Market access and quality
upgrading: Evidence from four field experiments”. In: American Economic Review 112.8 (2022),
pp. 2518–52.
[32] Patrick Bolton and Mathias Dewatripont. “The firm as a communication network”. In: The
Quarterly Journal of Economics 109.4 (1994), pp. 809–839.
[33] Mary Borrowman and Stephan Klasen. “Drivers of gendered sectoral and occupational
segregation in developing countries”. In: Feminist Economics 26.2 (2020), pp. 62–94.
[34] Kirill Borusyak, Peter Hull, and Xavier Jaravel. “Quasi-experimental shift-share research designs”.
In: The Review of Economic Studies 89.1 (2022), pp. 181–213.
[35] Thomas Breda, Julien Grenet, Marion Monnet, and Clémentine Van Effenterre. “Do female role
models reduce the gender gap in science? Evidence from French high schools”. In: IZA Journal of
Labor Economics (2020).
[36] Anne Brenøe. “Origins of gender norms: Sibling gender composition and women’s choice of
occupation and partner”. In: University of Zurich, Department of Economics, Working Paper 294
(2018).
[37] Miriam Bruhn, Dean Karlan, and Antoinette Schoar. “The impact of consulting services on small
and medium enterprises: Evidence from a randomized trial in Mexico”. In: Journal of Political
Economy 126.2 (2018), pp. 635–687.
[38] Leonardo Bursztyn, Georgy Egorov, and Stefano Fiorin. “From extreme to mainstream: The
erosion of social norms”. In: American Economic Review 110.11 (2020), pp. 3522–48.
[39] Leonardo Bursztyn, Thomas Fujiwara, and Amanda Pallais. “Acting wife: Marriage market
incentives and labor market investments”. In: American Economic Review 107.11 (2017),
pp. 3288–3319.
122
[40] Leonardo Bursztyn, Alessandra L González, and David Yanagizawa-Drott. “Misperceived social
norms: Women working outside the home in saudi arabia”. In: American economic review 110.10
(2020), pp. 2997–3029.
[41] Lorenzo Caliendo, Ferdinando Monte, and Esteban Rossi-Hansberg. “The anatomy of French
production hierarchies”. In: Journal of Political Economy 123.4 (2015), pp. 809–852.
[42] Juan Carlos Campaña, Jose Ignacio Giménez-Nadal, and José Alberto Molina. “Gender norms and
the gendered distribution of total work in Latin American households”. In: Feminist Economics
24.1 (2018), pp. 35–62.
[43] Renzo Carriero and Lorenzo Todesco. “Housework division and gender ideology: When do
attitudes really matter?” In: Demographic Research 39 (2018), pp. 1039–1064.
[44] Danilo Cavapozzi, Marco Francesconi, and Cheti Nicoletti. “The impact of gender role norms on
mothers’ labor supply”. In: Journal of Economic Behavior & Organization 186 (2021), pp. 113–134.
[45] Alfred Dupont Chandler, Takashi Hikino, and Alfred D Chandler. Scale and scope: The dynamics
of industrial capitalism. Harvard University Press, 2009.
[46] Lucas C Coffman, Clayton R Featherstone, and Judd B Kessler. “Can social information affect
what job you choose and keep?” In: American Economic Journal: Applied Economics 9.1 (2017),
pp. 96–117.
[47] John J Conlon. “Major malfunction: A field experiment correcting undergraduates’ beliefs about
salaries”. In: Journal of Human Resources (2019), 0317–8599R2.
[48] John J Conlon, Laura Pilossoph, Matthew Wiswall, and Basit Zafar. Labor market search with
imperfect information and learning. Tech. rep. National Bureau of Economic Research, 2018.
[49] Patricia Cortes and Jessica Pan. “Occupation and gender”. In: The Oxford handbook of women and
the economy (2018), pp. 425–452.
[50] Patricia Cortés, Jessica Pan, Laura Pilossoph, and Basit Zafar. Gender differences in job search and
theearningsgap:Evidencefrombusinessmajors. Tech. rep. National Bureau of Economic Research,
2021.
[51] Mick Cunningham. “Gender in cohabitation and marriage: The influence of gender ideology on
housework allocation over the life course”. In: Journal of Family Issues 26.8 (2005), pp. 1037–1061.
[52] Mick Cunningham. “Influences of gender ideology and housework allocation on women’s
employment over the life course”. In: Social science research 37.1 (2008), pp. 254–267.
[53] Monica Das Gupta, Jiang Zhenghua, Li Bohua, Xie Zhenming, Woojin Chung, and Bae Hwa-Ok.
“Why is son preference so persistent in East and South Asia? A cross-country study of China,
India and the Republic of Korea”. In: The Journal of Development Studies 40.2 (2003), pp. 153–187.
123
[54] Nilanjana Dasgupta and Shaki Asgari. “Seeing is believing: Exposure to counterstereotypic
women leaders and its effect on the malleability of automatic gender stereotyping”. In: Journal of
experimental social psychology 40.5 (2004), pp. 642–658.
[55] Suresh De Mel, David McKenzie, and Christopher Woodruff. “Labor drops: Experimental evidence
on the return to additional labor in microenterprises”. In: American Economic Journal: Applied
Economics 11.1 (2019), pp. 202–35.
[56] Suresh De Mel, David McKenzie, and Christopher Woodruff. “Returns to capital in
microenterprises: evidence from a field experiment”. In: The Quarterly Journal of Economics 123.4
(2008), pp. 1329–1372.
[57] Yasemin Dildar. “Patriarchal norms, religion, and female labor supply: Evidence from Turkey”. In:
World Development 76 (2015), pp. 40–61.
[58] Matthias Doepke and Fabian Kindermann. “Bargaining over babies: Theory, evidence, and policy
implications”. In: American Economic Review 109.9 (2019), pp. 3264–3306.
[59] Juan José Dolado, Florentino Felgueroso, and Juan F Jimeno. “Recent trends in occupational
segregation by gender: a look across the Atlantic”. In: Available at SSRN 320108 (2002).
[60] Pascaline Dupas and Jonathan Robinson. “Savings constraints and microenterprise development:
Evidence from a field experiment in Kenya”. In: American Economic Journal: Applied Economics
5.1 (2013), pp. 163–92.
[61] Stefan Eriksson and Jonas Lagerström. “The labor market consequences of gender differences in
job search”. In: Journal of Labor Research 33.3 (2012), pp. 303–327.
[62] Susanne Fahlén. “Equality at home–A question of career? Housework, norms, and policies in a
European comparative perspective”. In: Demographic Research 35 (2016), pp. 1411–1440.
[63] Raquel Fernandez and Alessandra Fogli. “Culture: An empirical investigation of beliefs, work, and
fertility”. In: American economic journal: Macroeconomics 1.1 (2009), pp. 146–77.
[64] Sarah Fleche, Anthony Lepinteur, and Nattavudh Powdthavee. “Gender norms, fairness and
relative working hours within households”. In: Labour Economics 65 (2020), p. 101866.
[65] Jonas Fluchtmann, Anita M Glenny, Nikolaj Harmon, and Jonas Maibom. “The gender application
gap: Do men and women apply for the same jobs?” In: mimeo (2020).
[66] Nicole M Fortin. “Gender role attitudes and women’s labor market participation: Opting-out, aids,
and the persistent appeal of housewifery”. In: Annals of Economics and Statistics/Annales
d’Économie et de Statistique 117/118 (2015), pp. 379–401.
[67] Michelle L Frisco and Kristi Williams. “Perceived housework equity, marital happiness, and
divorce in dual-earner households”. In: Journal of Family Issues 24.1 (2003), pp. 51–73.
[68] Makiko Fuwa. “Macro-level gender inequality and the division of household labor in 22
countries”. In: American sociological review 69.6 (2004), pp. 751–767.
124
[69] Yana Gallen and Melanie Wasserman. “Informed choices: Gender gaps in career advice”. In: CEPR
Discussion Paper No. DP15728 (2021).
[70] Luis Garicano and Esteban Rossi-Hansberg. “Organization and inequality in a knowledge
economy”. In: The Quarterly journal of economics 121.4 (2006), pp. 1383–1435.
[71] Claudia Goldin and Lawrence F Katz. “The power of the pill: Oral contraceptives and women’s
career and marriage decisions”. In: Journal of political Economy 110.4 (2002), pp. 730–770.
[72] Paul Goldsmith-Pinkham, Isaac Sorkin, and Henry Swift. “Bartik instruments: What, when, why,
and how”. In: American Economic Review 110.8 (2020), pp. 2586–2624.
[73] Theodore N Greenstein. “Husbands’ participation in domestic labor: Interactive effects of wives’
and husbands’ gender ideologies”. In: Journal of Marriage and the Family (1996), pp. 585–595.
[74] Matthew Groh, David McKenzie, Nour Shammout, and Tara Vishwanath. “Testing the importance
of search frictions and matching through a randomized experiment in Jordan”. In: IZA Journal of
Labor Economics 4.1 (2015), pp. 1–20.
[75] Ingar Haaland, Christopher Roth, and Johannes Wohlfart. “Designing information provision
experiments”. In: (2020).
[76] Susan Hanson and Geraldine Pratt. “Job search and the occupational segregation of women”. In:
Annals of the Association of American geographers 81.2 (1991), pp. 229–253.
[77] Morgan Hardy and Jamie McCasland. “Are small firms labor constrained? experimental evidence
from ghana”. In: American Economic Journal: Applied Economics (2022).
[78] Thomas J Holmes and John J Stevens. “An alternative theory of the plant size distribution, with
geography and intra-and international trade”. In: Journal of Political Economy 122.2 (2014),
pp. 369–421.
[79] Chang-Tai Hsieh and Benjamin A Olken. “The Missing "Missing Middle"”. In: The Journal of
Economic Perspectives 28.3 (2014), pp. 89–108.
[80] Jisoo Hwang, Chulhee Lee, and Esther Lee. “Gender norms and housework time allocation among
dual-earner couples”. In: Labour Economics 57 (2019), pp. 102–116.
[81] Rahmah Ismail, Maryam Farhadi, and Chung-Khain Wye. “Occupational segregation and gender
wage differentials: Evidence from Malaysia”. In: Asian Economic Journal 31.4 (2017), pp. 381–401.
[82] Simon Jäger, Christopher Roth, Nina Roussille, and Benjamin Schoefer. Worker beliefs about
outside options. Tech. rep. National Bureau of Economic Research, 2022.
[83] Vaishali Jain, Lala Rukh Khan, Sarah Xu, and Julia Wu. “Understanding gender barriers in the
pakistani labour market”. In: IGC Policy Brief (2021).
125
[84] Robert Jensen. “Do labor market opportunities affect young women’s work and family decisions?
Experimental evidence from India”. In: The Quarterly Journal of Economics 127.2 (2012),
pp. 753–792.
[85] Robert Jensen and Nolan H. Miller. “Market integration, demand, and the growth of firms:
Evidence from a natural experiment in India”. In: American Economic Review 108.12 (2018),
pp. 3583–3625.
[86] Robert Jensen and Emily Oster. “The power of TV: Cable television and women’s status in India”.
In: The Quarterly Journal of Economics 124.3 (2009), pp. 1057–1094.
[87] Namrata Kala. The impacts of managerial autonomy on firm outcomes . Tech. rep. National Bureau
of Economic Research, 2019.
[88] Carly R Knight and Mary C Brinton. “One egalitarianism or several? Two decades of gender-role
attitude change in Europe”. In: American Journal of Sociology 122.5 (2017), pp. 1485–1532.
[89] Michael S Kofoed et al. “The effect of same-gender or same-race role models on occupation choice
evidence from randomly assigned mentors at west point”. In: Journal of Human Resources 54.2
(2019), pp. 430–467.
[90] Peter Kuhn and Kailing Shen. “Gender discrimination in job ads: Evidence from china”. In: The
Quarterly Journal of Economics 128.1 (2013), pp. 287–336.
[91] Keedon Kwon. “Regionalism in South Korea: Its origins and role in her democratization”. In:
Politics & Society 32.4 (2004), pp. 545–574.
[92] Thomas Le Barbanchon, Roland Rathelot, and Alexandra Roulet. “Gender differences in job
search: Trading off commute against wage”. In: The Quarterly Journal of Economics 136.1 (2021),
pp. 381–426.
[93] Rebecca Ann Lind and Colleen Salo. “The framing of feminists and feminism in news and public
affairs programs in US electronic media”. In: Journal of Communication 52.1 (2002), pp. 211–228.
[94] Ilse Lindenlaub and Anja Prummer. Gender, social networks and performance. Tech. rep. 2016.
[95] Ioana Marinescu and Daphné Skandalis. “Unemployment insurance and job search behavior”. In:
The Quarterly Journal of Economics 136.2 (2021), pp. 887–931.
[96] Alexandre Mas and Amanda Pallais. “Valuing alternative work arrangements”. In: American
Economic Review 107.12 (2017), pp. 3722–59.
[97] David McKenzie, Nabila Assaf, and Ana Paula Cusolito. “The additionality impact of a matching
grant programme for small firms: experimental evidence from Yemen”. In: Journal of Development
Effectiveness 9.1 (2017), pp. 1–14.
[98] David McKenzie and Christopher Woodruff. “Business practices in small firms in developing
countries”. In: Management Science 63.9 (2017).
126
[99] David McKenzie and Christopher Woodruff. “Experimental evidence on returns to capital and
access to finance in Mexico”. In: The World Bank Economic Review 22.3 (2008), pp. 457–482.
[100] David McKenzie and Christopher Woodruff. “What are we learning from business training and
entrepreneurship evaluations around the developing world?” In: The World Bank Research
Observer 29.1 (2014), pp. 48–82.
[101] Xin Meng and Paul Miller. “Occupational segregation and its impact on gender wage
discrimination in China’s rural industrial sector”. In: Oxford Economic Papers (1995), pp. 136–155.
[102] OECD. “Education at a Glance”. In: OECD Indicators (2020).
[103] OECD. “Gender wage gap”. In: OECD Indicators (2021).
[104] Emily Oster. “Unobservable selection and coefficient stability: Theory and evidence”. In: Journal
of Business & Economic Statistics 37.2 (2019), pp. 187–204.
[105] Saejung Park and Joonmo Cho. “Young people are medically invulnerable to COVID-19 but
vulnerable in the labor market: Korean evidence”. In: Health Economics Review 12.1 (2022),
pp. 1–16.
[106] Allen M Parkman. “Bargaining over housework: The frustrating situation of secondary wage
earners”. In: American Journal of Economics and Sociology 63.4 (2004), pp. 765–794.
[107] Rebecca Pearse and Raewyn Connell. “Gender norms and the economy: insights from social
research”. In: Feminist Economics 22.1 (2016), pp. 30–53.
[108] Barbara Petrongolo. “Gender segregation in employment contracts”. In: Journal of the European
Economic Association 2.2-3 (2004), pp. 331–345.
[109] Darlene L Pina and Vern L Bengtson. “The division of household labor and wives’ happiness:
Ideology, employment, and perceptions of support”. In: Journal of Marriage and the Family (1993),
pp. 901–912.
[110] Michael J Piore and Charles F Sabel. “The second industrial divide: possibilities for prosperity”. In:
(1984).
[111] Vivien Procher, Nolan Ritter, and Colin Vance. “Housework allocation in Germany: The role of
income and gender identity”. In: Social Science Quarterly 99.1 (2018), pp. 43–61.
[112] James M Raymo, Hyunjoon Park, Yu Xie, and Wei-jun Jean Yeung. “Marriage and family in East
Asia: Continuity and change”. In: Annual Review of Sociology 41 (2015), pp. 471–492.
[113] Myungcheol Sa. “The change of occupational segregation by gender: Focus on trend and
decomposition”. In: Quarterly Journal of Labor Policy 15.4 (2015), pp. 1–24.
[114] Sarah A Schnitker. “An examination of patience and well-being”. In: The Journal of Positive
Psychology 7.4 (2012), pp. 263–280.
127
[115] Meredith Startz. “The value of face-to-face: Search and contracting problems in Nigerian trade”.
In: mimeo (2019).
[116] Oriel Sullivan. “What do we learn about gender by analyzing housework separately from child
care? Some considerations from time-use evidence”. In: Journal of Family Theory & Review 5.2
(2013), pp. 72–84.
[117] Christopher Taber and Rune Vejlin. “Estimation of a Roy/search/compensating differential model
of the labor market”. In: Econometrica 88.3 (2020), pp. 1031–1069.
[118] Nikolas Tromp. “The narrowing gender wage gap in South Korea”. In: Journal of the Japanese and
International Economies 53 (2019), p. 101032.
[119] Amos Tversky and Daniel Kahneman. “Judgment under uncertainty: Heuristics and biases: Biases
in judgments reveal some heuristics of thinking under uncertainty.” In: Science 185.4157 (1974),
pp. 1124–1131.
[120] Anna Vitali. “Consumer search and firm location: Theory and evidence from the garment sector
in Uganda”. In: mimeo (2022).
[121] Glenn D Wilson. “Feminism and marital dissatisfaction”. In: Personality and Individual Differences
3.3 (1982), pp. 345–347.
[122] Matthew Wiswall and Basit Zafar. “Determinants of college major choice: Identification using an
information experiment”. In: The Review of Economic Studies 82.2 (2015), pp. 791–824.
[123] World Bank. World development report 2012: Gender equality and development. The World Bank,
2011.
[124] Xiaokai Yang and Jeff Borland. “A microeconomic mechanism for economic growth”. In: Journal
of political economy 99.3 (1991), pp. 460–482.
128
129
Appendices
A AppendixtoChapter1
A.1 AppendixTablesandFigures
Table A.1: Employment Shares and Gender Composition of the Top 20 Sectors in GOMS
Sectors Employment Proportion
shares women
(%) (%)
Education 13.3 67.9
Wholesaletrade 6.4 38.7
Professional services 5.8 57.9
Retailtrade 5.0 49.3
Financialserviceactivities 4.9 44.6
Socialworkactivities 4.8 70.2
Businesssupportactivities 4.1 62.7
Publishing activities 3.4 52.2
Food and beverage service 3.0 57.3
Insurance and pension funding 2.6 42.4
Activities auxiliary to financial services 1.7 36.3
Warehousing and transportation 1.7 46.0
Human health activities 1.7 66.3
Manuf. ofelectroniccomponents 1.5 42.6
Computer programming, etc. 1.5 43.6
Creative, arts, recreation activities 1.4 61.0
Manuf. of other machinery, etc. 1.4 36.3
Manuf. of motor vehicles, trailers 1.3 31.8
Manuf. of chemical products 1.2 40.0
Research and development 1.1 50.9
Sample size 33,542
Notes: This table reports the employment shares and proportions of women employees of the top 20 employers of college grad-
uates with humanities and social sciences majors, computed from the most recent five waves (2015 ∼ 2019) of the Graduates
Occupational Mobility Survey. The seven experimental sectors are in bold.
130
Table A.2: Construction of Main Outcome Variables
Category Variable Construction Description Results
Sectoral choices (Std.) Relative aspirations
to MDS
Rel.Asp
i
=
P
m∈MDS
Aspim
P
f∈FDS
Asp
if
Studenti’s aspiration to enter
sectorj, measured on a 1∼ 10
Likert scale
(All three rounds)
Table 1.3,
Col (1)∼ (5)
Sectoral choices Proportion of MDS
searched
among seven sectors
Prop.MDSSearched
i
=
P
m∈MDS
Num.SectorsSearchedim
P
j∈7
Num.SectorsSearchedij
Studenti’s search for jobs in
sectorj among the seven
experimental sectors
(Second follow-up)
Table 1.4,
Col (3)
Sectoral choices Proportion of MDS
applied for
among seven sectors
Prop.MDSApplied
i
=
P
m∈MDS
Num.SectorsAppliedim
P
j∈7
Num.SectorsAppliedij
Studenti’s applications for jobs in
sectorj among the seven
experimental sectors
(Second follow-up)
Table 1.4,
Col (6)
Employment outcomes Proportion of women in
the new job,
if employed
Femaleproportion
i
The proportion of women in the
sector where studenti is employed
(second follow-up)
Table 1.5,
Panel B Col
(1)
Checks on other
components
(Log) Subjective
probabilities for the seven
sectors
logp
ij
Studenti’s subjective probability
of receiving a job offer in sector j
around graduation (first
follow-up)
Table 1.6,
Col (1)
Checks on other
components
Preference weights on
four job characteristics
α ic
Studenti’s allocation from the
total of 8 points to job
characteristicc (second follow-up)
Table 1.6,
Col (2)∼ (5)
131
Table A.3: Attrition and Balance in the Follow-up Surveys
1
st
Follow-up 2
nd
Follow-up
Group
C
Group
T
p
value
Norm.
Diff.
Group
C
Group
T
p
value
Norm.
Diff.
(1) (2) (3) (4) (5) (6) (7) (8)
Panel A. Demographics & background
Age 22.61 22.52 0.402 -0.059 22.61 22.55 0.625 -0.041
Num. of semesters in college 7.17 7.16 0.910 -0.008 7.19 7.18 0.843 -0.016
GPA (/4.0) 3.59 3.60 0.783 0.019 3.59 3.61 0.484 0.058
Num. of certificates 1.59 1.59 0.962 -0.003 1.62 1.55 0.506 -0.055
College in metropolitan area (0/1) 0.59 0.59 0.915 0.008 0.58 0.63 0.241 0.097
Parents’ monthly income (10,000 KRW) 555.9 524.4 0.092 -0.119 533.8 512.4 0.324 -0.082
Mother’s yrs. of education 14.5 14.5 0.954 0.004 14.4 14.7 0.084 0.145
Panel B. Attitudes & personal traits
Risk preference in financial matters (1 ∼ 7) 3.43 3.50 0.477 0.050 3.48 3.50 0.908 0.010
Risk preference in daily life (1∼ 7) 3.37 3.48 0.222 0.086 3.44 3.49 0.643 0.039
Patience (1∼ 7) 4.64 4.64 0.933 0.006 4.60 4.62 0.843 0.016
Panel C. Job search history at baseline
Started job search (0/1) 0.95 0.97 0.129 0.107 0.95 0.97 0.362 0.076
Ever applied for job openings (0/1) 0.29 0.25 0.265 -0.078 0.28 0.23 0.120 -0.129
Planning to apply in one year (0/1) 0.56 0.59 0.554 0.042 0.57 0.59 0.645 0.038
Kolmogorov–Smirnovp value (major
dist.)
0.092 0.043
Joint orthogonalityp value 0.308 0.326
Differential attrition p value 0.379 0.249
Sample size 409 398 300 282
Notes: This table shows the attrition and balance of observable characteristics between the treatment and control groups in
the two rounds of follow-up surveys.p values smaller than 0.1 are in bold. All variables are measured at baseline. Kolmogorov–
Smirnovp value is the result of the Kolmogorov–Smirnov test of identical major distribution within humanities and social sciences
categories. Joint orthogonalityp value is the result of anF test with the null hypothesis that the variables are jointly orthogonal
to treatment status. Differential attrition p value is the result of anF test with the null hypothesis that the attrition rates are
orthogonal to treatment status.
132
Table A.4: Parameters vs. Students’ Beliefs
1
st
Follow-up 2
nd
Follow-up
Group C Group T Group C Group T
Parameter
(Mean)
Beliefs
(Mean,
SD)
Mean
Percent
error
Beliefs
(Mean,
SD)
Mean
Percent
error
p value Beliefs
(Mean,
SD)
Mean
Percent
error
Beliefs
(Mean,
SD)
Mean
Percent
error
p value
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11)
Annual salary
(10,000 KRW)
MDS avg. 3,121 2,800
(674)
-10.3 2,880
(628)
-7.7 0.079 2,850
(664)
-8.7 2,919
(683)
-6.5 0.217
FDS avg. 2,147 2,432
(500)
13.3 2,338
(444)
8.9 0.005 2,385
(442)
11.1 2,365
(390)
10.1 0.547
Work hours (weekly)
MDS avg. 44.0 43.7
(6.4)
-0.6 43.8
(6.0)
-0.4 0.885 43.6
(5.9)
-0.9 43.1
(4.7)
-2.0 0.307
FDS avg. 40.0 43.2
(6.3)
8.2 42.5
(5.9)
6.2 0.073 42.7
(5.9)
6.9 42.5
(5.1)
6.3 0.546
Welfare inst. index
(0∼ 5)
MDS avg. 4.49 3.78
(0.64)
-15.9 4.07
(0.64)
-9.4 0.000 4.01
(0.69)
-10.7 4.07
(0.61)
-9.5 0.300
FDS avg. 3.80 3.61
(0.81)
-5.2 3.60
(0.80)
-5.2 0.965 3.64
(0.81)
-4.4 3.53
(0.73)
-7.1 0.099
Job security index
(0∼ 5)
MDS avg. 4.51 3.79
(0.77)
-16.0 4.09
(0.65)
-9.3 0.000 3.90
(0.68)
-13.7 4.08
(0.62)
-9.7 0.001
FDS avg. 3.36 3.71
(0.73)
10.4 3.56
(0.76)
5.9 0.004 3.80
(0.75)
13.1 3.69
(0.71)
9.9 0.074
Notes: The table compares wage and nonwage parameters (mean) of the seven sectors and students’ elicited beliefs in the two follow-up surveys. Percent error is defined as (belief
- truth)/truth *100. Columns (6) and (11) reports thep value of thet test with the null hypothesis of the same percent errors across the two groups.
133
Table A.5: Analyses of Heterogeneity in Treatment Effects on the Main Outcome variables
7 sectors Num. MDS Prop. MDS 7 Sectors Num. MDS Prop. MDS Female prop.
ever searched (0/1) searched searched ever applied (0/1) applied applied of the job
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) (13) (14)
Treated (0/1) 0.094** 0.087* 0.146** 0.112 0.101** 0.111** 0.057 0.060 0.080* 0.069 0.128* 0.140* 0.001 -0.007
(0.044) (0.049) (0.066) (0.073) (0.047) (0.052) (0.042) (0.048) (0.048) (0.050) (0.073) (0.077) (0.031) (0.030)
Search Cost (Wage) 0.006 0.012* -0.001 0.015*** 0.009 -0.004 -0.002
(0.004) (0.007) (0.003) (0.005) (0.008) (0.004) (0.005)
Treated x Search Cost -0.002 -0.018* -0.001 -0.006 -0.011 -0.001 -0.004
(Wage) (0.007) (0.011) (0.007) (0.008) (0.010) (0.010) (0.005)
Search Cost (Nonwage) 0.008 0.004 -0.005 0.013** 0.003 -0.013*** -0.003
(0.006) (0.008) (0.005) (0.006) (0.005) (0.005) (0.004)
Treated x Search Cost 0.001 -0.005 -0.004 -0.004 -0.006 -0.004 -0.004
(Nonwage) (0.008) (0.013) (0.008) (0.009) (0.008) (0.009) (0.005)
Observations 582 582 582 582 407 407 582 582 582 582 188 188 79 79
Notes: This table shows the regression results for all our main outcome variables on the interaction terms between the treatment status and estimated search costs for wage and
nonwage information, as explained in Table 1.3. The dependent variables of Columns (1) to (12) correspond to those in Table 1.4. The dependent variables of Columns (13) and
(14) correspond to those in Panel B Column (1) in Table 1.5.
134
Table A.6: Reasons for Not Searching in Specific Sectors ( 2
nd
follow-up, aggregated)
MDS FDS
C T C T
(%) (%) (%) (%)
Do not like wages 2.1 1.3 7.9 6.4
Do not like work hours, welfare, or security 4.8 2.5** 8.8 8.1
Does not fit my aptitude or interest 54.6 59.8 48.5 51.8
Do not like workload or difficulties 7.0 7.4 5.8 6.0
Do not like future prospects/reputation 3.6 3.2 4.7 4.9
Do not think I will have a chance even if I try 15.9 17.5 13.1 13.6
Do not like female proportion/friendliness 3.3 3.6 1.8 0.6
Do not know which firms belong to the sector 8.6 4.7** 9.4 8.7
Total 100 100 100 100
Notes: This table shows the distributions of the reasons why the female students did not search for information about MDS/FDS
jobs, if applicable. Sample: Those who reported having not searched for information about each sector among the 582 female
students in the second follow-up survey. The distributions for each of the seven sectors are provided in the Online Appendix.
Asterisks represent the results of the OLS regressions of the dummy for choosing each reason on the treatment status and control
variables used in the main analysis (**:p< 0.05).
135
Figure A.1: Occupational Segregation by Gender among College Graduates (2-digit sectors)
(a) Cdf of sectoral distribution (b) Share of employment by gender
Notes: The distribution of 48,277 four-year college graduates’ first jobs for the past five years (2015 ∼ 2019) in 99 two-digit sectors,
reported in the GOMS. Sector classifications: manufacturing (10 ∼ 34), wholesale trade (46), retail trade (47), food and beverage
service activities (56), professional services (71), business service activities (75), public administration and defense (84), education
(85), human health activities (86), social work activities (87).
136
Figure A.2: Histogram of Female Graduates’ Sector Distributions by Stated Preferences
(a) Wages (b) Work hours
(c) Welfare institutions (d) Job security
Notes: The sector distributions by the stated preferences for each characteristic, calculated from the 48,277 college graduates
responses in the past five waves (2015 ∼ 2019) of the GOMS. The definitions of the four characteristics are presented in Table
1.1. Stated preference is a dummy taking 1 if each characteristic is reported as one of the two priorities in a student’s job choice
criteria.
137
A.2 AnExperimentof“TailoredJobInformationNewsletter”
Overview. The main outcome variables assessing the sectoral allocation of job search in Section 1.4 are
at the sector level. In the second follow-up survey, we asked whether the students searched or applied for
any of the seven experimental sectors in the past four months and constructed the variables with those
responses. However, this sector-level analysis has the limitation in that it cannot capture within-sector
variation. For example, a student might have searched for more firms in the wholesale trade sector (MDS)
and searched for fewer firms in the education sector (FDS). To address this problem, we embedded a survey
experiment in the first follow-up survey designed to test the differential changes in sectoral allocation,
measured at the firm level.
Experimental design. Students were asked to participate in a demand survey for the “tailored job in-
formation newsletter”, which was explained to provide information on firms and job characteristics about
which students self-select into receiving information. For example, if a student chooses “ABC electronics”
as a firm name and “welfare institutions” as a job characteristic of interest, the newsletter provides this
information. We provided a list of a total of 49 firms in our seven experimental sectors (seven firms for
each sector) that posted their actual job openings on one of Korea’s representative job matching websites,
Incruit.com. (The actual material shown in the experiment is provided in the Online Appendix). Then, the
students were to choose a minimum of three to a maximum of ten firms in order of preferences.
Results. Table A.7 demonstrates the results with the same regression specifications as in the main anal-
ysis. Columns (1) and (2) show that the treatment effects on the number of total sectors and firms chosen
were positive but small and statistically insignificant. On the other hand, the students in Group T chose 0.3
more firms that belong to the three MDS (Column (3)). Considering that Group C students chose 2.6 firms
in MDS on average, the treatment effect was an approximately 10% increase in the number of MDS firms,
which is substantial but smaller than the magnitude computed in the sector-level analysis (approximately
138
Table A.7: Treatment Effects in the Newsletter Experiment ( 1
st
follow-up)
Num. of Num. of Num. of Num. of Prop. of
sectors firms MDS firms FDS firms MDS firms
(1) (2) (3) (4) (5)
Treated (0/1) 0.070 0.166 0.275* -0.004 0.026
(0.087) (0.171) (0.152) (0.119) (0.024)
Mean of Dep Var in C 2.795 4.914 2.597 1.543 0.524
Observations 807 807 807 807 807
Notes: This table shows the regression results of the treatment status on the students’ choices of firms in the tailored job informa-
tion newsletter experiment at the first follow-up survey. Robust standard errors in parentheses. Sample: 807 female students who
responded to the second follow-up survey. Dependent variables: (1) The number of sectors chosen (among the seven sectors) / (2)
The number of firms chosen (among the 49 firms) / (3) The number of firms chosen in the three MDS / (4) The number of firms
chosen in the three FDS / (5) The proportion of firms chosen in the three MDS. The regressions are OLS with control variables
chosen by the double LASSO method among the variables in Table 1.2.
20%, Column (2) of Table 1.4). As a result, the proportion of FDS firms was smaller and that of MDS firms
larger for Group T students, albeit estimated with noise.
139
B AppendixtoChapter2
B.1 AppendixTablesandFigures
Table B.1: Market Shares of the Big Six Newspapers (%)
CS JA DA MK KH HKR
2014
Seoul 35.5 21.6 11.9 7.3 3.7 2.4
Gyeongin 29.9 19.7 16.1 6.8 5.6 3.2
Chungcheong 25.6 9.7 13.1 3.7 6.1 4.1
Jeolla 5.0 5.3 12.8 4.1 4.4 10.3
Gyeongbuk 26.7 5.4 14.8 3.6 1.1 1.9
Gyeongnam 29.9 12.3 8.0 1.8 2.4 2.1
Gangwon 24.7 6.1 12.8 4.1 7.9 3.8
Jeju 7.9 9.2 7.9 7.7 7.3 3.8
Average 28.0 14.8 12.8 5.2 4.2 3.4
2018
Seoul 24.5 21.8 11.2 6.5 6.2 3.8
Gyeongin 32.1 15.1 16.1 9.5 6.1 3.7
Chungcheong 19.0 5.3 14.8 2.2 5.9 2.1
Jeolla 4.3 3.4 12.7 3.2 5.3 12.0
Gyeongbuk 32.5 14.3 13.3 4.8 2.6 1.7
Gyeongnam 33.8 9.0 6.9 4.4 2.3 1.7
Gangwon 17.7 13.9 0 0 0 0
Jeju 17.5 4.6 3.1 2.7 0 5.1
Average 26.2 13.4 12.2 5.9 4.9 3.7
Notes: The market shares of the six national newspapers in 2014 and 2018, computed from the Media Audience Survey 2018. A
newspaper’s market share is defined as the proportion of respondents who chose it as the main newspaper they read the previous
week. The names of newspapers are abridged (CS: Chosun, JA: Joongang, DA: Donga, MK: Maekyung, KH: Kyunghyang, HKR:
Hankyoreh) for legibility.
140
Table B.2: Survey Questions Used to Measure Perceptions of Gender Norms in the Literature
Literature Source Questions used Format
Fuwa (2004) International Social
Survey
A job is all right, but what most women really want is
a home and children.
Likert scale
(0∼ 4)
Programme Being a housewife is just as fulfilling as working for
pay.
A man’s job is to earn money; a woman’s job is to
look after the home and family.
All in all, family life suffers when the woman has a
full-time job.
A preschool child is likely to suffer if his or her
mother works.
Fahlén (2016) European Social
Survey
Women should be prepared to cut down on paid work
for the sake of the family.
Likert scale
(0∼ 8)
Men should have more right to jobs than women
when jobs are scarce.
Campaña,
Giménez-Nadal,
and Molina
(2018)
World Values Survey When a mother works for pay, the children suffer. Likert scale
(1∼ 4)
On the whole, men make better political leaders than
women do.
A university education is more important for a boy
than for a girl.
On the whole, men make better business executives
than women.
Being a housewife is just as fulfilling as working for
pay.
Cavapozzi,
Francesconi, and
Nicoletti (2021)
UK Household
Longitudinal Study
Pre-school child suffers if mother works. Likert scale
(1∼ 5)
Family suffers if mother works full time.
Husband and wife should contribute to household
income.
Husband should earn, wife should stay at home.
Employers should help mothers combine jobs and
childcare.
141
Table B.3: Pairwise Correlation Coefficients between the Components of the GNP Score
Q1 Q2 Q3 Q4 Q5 Q6 Q7 Q8 Q9
Q1 1.000
Q2 0.480 1.000
Q3 0.273 0.172 1.000
Q4 0.106 0.012 0.118 1.000
Q5 0.065 -0.100 0.151 0.417 1.000
Q6 0.208 0.200 0.080 -0.016 -0.116 1.000
Q7 -0.118 -0.160 0.039 0.209 0.163 -0.210 1.000
Q8 0.135 0.217 0.045 -0.121 -0.227 0.338 -0.231 1.000
Q9 -0.055 -0.224 0.077 0.191 0.309 -0.112 0.313 0.219 1.000
142
Table B.4: Sample and Attrition
Wave 6 Wave 7
Total Divorce/
Separation
Total Divorce/
Separation
(1) (2) (3) (4)
Wife’s age -0.001
(0.002)
0.000
(0.001)
-0.004*
(0.002)
-0.000
(0.003)
Husband’s age -0.000
(0.002)
-0.001
(0.001)
0.001
(0.002)
-0.001
(0.001)
Wife’s years of edu. -0.001
(0.003)
-0.001
(0.001)
-0.003
(0.004)
-0.001
(0.002)
Husband’s years of edu. 0.002
(0.003)
-0.000
(0.001)
0.001
(0.003)
-0.001
(0.001)
Number of children
under 19
-0.003
(0.008)
0.002
(0.003)
0.004
(0.010)
0.002
(0.004)
Husband works (0/1) -0.019
(0.015)
-0.010
(0.009)
-0.036*
(0.020)
-0.015
(0.011)
HH annual income
(1,000 USD)
0.000
(0.000)
0.000
(0.000)
0.000
(0.000)
-0.000
(0.000)
Region FE Yes Yes Yes Yes
Jointp-value 0.072 0.272 0.003 0.186
Num. of Attrition 287 38 486 62
Num. of Observation 5,158 4,871 5,158 4,672
Notes: Standard deviations in parenthesis. The jointp-value is the result of anF test with the null hypothesis that the variables
are jointly orthogonal to attrition.
143
Table B.5: 2SLS Estimates on Time Use in Housework/Childcare
Wife’s time spent(mins/week) Husband’s time spent (mins/week)
Household
labor
Housework Childcare Household
labor
Housework Childcare
(1) (2) (3) (4) (5) (6)
2SLS
GNP score -618.0*** -385.3*** -232.7* 78.4 77.5* 0.9
(Std.) (205.0) (123.1) (134.2) (69.8) (45.2) (48.3)
Mean Dep. 1,414.3 1,087.6 326.7 217.1 139.7 77.3
Indiv. FE Yes Yes Yes Yes Yes Yes
Controls Yes Yes Yes Yes Yes Yes
Num. obs. 13,230 13,230 13,230 13,230 13,230 13,230
Notes: ***p< 0.01, **p< 0.05, *p< 0.1. Standard errors clustered at the individual level in parentheses. Sample: 4,422 married
couples (Observations with missing time use responses are omitted). All regressions are weighted by sampling weights.
144
Table B.7: Robustness – The Results with and without Region Fixed Effects (First Stage)
GNP score (Std.)
(1) (2)
Bartik Exposure (Bartik
jt
) 0.0775*** 0.0777***
(0.0148) (0.0148)
Individual F.E. Yes Yes
Region F.E. No Yes
Controls Yes Yes
Num. of Obs. 13,230 13,230
Notes: ***p< 0.01, **p< 0.05, *p< 0.1. Standard errors clustered at the individual level in parentheses.
Table B.8: Robustness – The results with and without region fixed effects (second stage)
Without region fixed effects With region fixed effects
Husband’s
share of
time
Wife’s
time spent
(mins/
week)
Husband’s
time spent
(mins/
week)
Wife
satisfied
with
allocation
(0/1)
Husband’s
share of
time
Wife’s
time spent
(mins/
week)
Husband’s
time spent
(mins/
week)
Wife
satisfied
with
allocation
(0/1)
(1) (2) (3) (4) (5) (6) (7) (8)
2SLS
GNP score 0.060** -618.0*** 78.4 -0.261** 0.060** -627.0*** 77.5 -0.256**
(Std.) (0.030) (205.0) (69.8) (0.109) (0.030) (205.9) (69.7) (0.108)
Indiv. FE Yes Yes Yes Yes Yes Yes Yes Yes
Controls Yes Yes Yes Yes Yes Yes Yes Yes
Num. obs. 13,230 13,230 13,230 13,230 13,230 13,230 13,230 13,230
Notes: ***p< 0.01, **p< 0.05, *p< 0.1. Standard errors clustered at the individual level in parentheses.
145
Table B.6: The OLS Estimates on Other Outcome Variables
(1) (2) (3) (4) (5)
Panel A: Wife’s labor market engagement
LFP
(0/1)
Weekly
Work hours
(cond.)
GNP score (Std.) 0.011***
(0.004)
0.058
(0.194)
Panel B: (Log) Household expenditures (1,000 USD, annual)
Total Food Children Others
Edu.
GNP score (Std.) 0.005
(0.003)
0.000
(0.004)
0.008
(0.012)
0.007
(0.005)
Panel C: Frequency of shared activities with the husband (monthly)
Cultural Exercise Social Family Family
Events Activities Occasions Occasions
(Husband’s) (Wife’s)
GNP score (Std.) 0.014
(0.011)
-0.029
(0.023)
-0.001
(0.010)
0.024
(0.026)
-0.018***
(0.005)
Panel D: Wife’s marital happiness (0/1)
GNP score (Std.) -0.018***
(0.005)
Individual F.E. Yes
Controls Yes
Num. of obs. 13,230
Notes: ***p< 0.01, **p< 0.05, *p< 0.1. Standard errors clustered at the individual level in parentheses. Sample: 4,422 married
couples (Observations with missing time use responses are omitted). All regressions are weighted by sampling weights.
146
Table B.9: Robustness - Replications of the Main Results without Gangwon and Jeju Regions
(a) First-stage results
GNP score (Std.)
(1) (2)
Exposure 0.0035***
(0.0008)
Bartik Exposure 0.0746***
(0.0159)
Individual F.E. Yes Yes
R
2
0.093 0.051
Num. of Obs. 12,262 12,262
(b) Second-stage results
Husband’s
share of
time
Wife’s
time spent
(mins/
week)
Husband’s
time spent
(mins/
week)
Wife
satisfied
with
allocation
(0/1)
(1) (2) (3) (4)
2SLS
GNP score (Std.) 0.041 -517.0** 52.1 -0.383***
(0.031) (209.9) (76.7) (0.136)
Individual F.E. Yes Yes Yes Yes
Controls Yes Yes Yes Yes
Num. of obs. 12,262 12,262 12,262 12,262
Notes: ***p< 0.01, **p< 0.05, *p< 0.1. Standard errors clustered at the individual level in parenthesis. Sample: 4,188 married
couples who lived in regions other than Gangwon and Jeju throughout the period (Observations with missing time use responses
are omitted). All regressions are weighted by sampling weights.
147
Table B.10: Robustness - Relationship between Bartik exposure and Observable Correlates
Market shares Bartik
CS JA DA MK KH HKR Exposure
(1) (2) (3) (4) (5) (6) (7)
Wife’s age -0.020
(0.020)
-0.021*
(0.009)
-0.001
(0.006)
-0.006*
(0.003)
0.000
(0.005)
0.008*
(0.004)
0.015
(0.009)
Husband’s age -0.015
(0.016)
-0.017*
(0.008)
-0.000
(0.005)
-0.006*
(0.003)
0.000
(0.004)
0.007*
(0.003)
0.012
(0.008)
Wife’s years of edu. 0.055
(0.036)
0.055***
(0.010)
-0.001
(0.011)
0.013*
(0.006)
-0.003
(0.009)
-0.015
(0.009)
-0.029
(0.018)
Husband’s years of
edu.
0.060
(0.036)
0.059***
(0.009)
-0.001
(0.012)
0.014*
(0.006)
-0.002
(0.009)
-0.015
(0.009)
-0.032
(0.019)
Number of children
under 19
0.009
(0.236)
0.159
(0.122)
-0.018
(0.063)
0.085**
(0.030)
0.021
(0.051)
-0.062
(0.053)
-0.084
(0.118)
Husband works (1/0) -1.323
(0.720)
-1.072**
(0.291)
-0.023
(0.242)
-0.181
(0.158)
-0.073
(0.193)
0.140
(0.216)
0.520
(0.419)
HH annual income
(1,000 USD)
0.013*
(0.006)
0.010***
(0.002)
0.001
(0.002)
0.001
(0.001)
-0.002
(0.002)
-0.002
(0.002)
-0.005
(0.004)
Notes: ***p< 0.01, **p< 0.05, *p< 0.1. Standard errors in parentheses. The results of single regressions of the market shares
and the Bartik exposure on the regional averages of the observable correlates in 2014. The names of newspapers are abridged
(CS: Chosun, JA: Joongang, DA: Donga, MK: Maekyung, KH: Kyunghyang, HKR: Hankyoreh) for legibility. The unit of nominal
independent variables (household income, wife’s monthly salary) is 1,000 USD (converted with the exchange rate of 1 USD = 1,200
KRW) and inflation-adjusted using Consumer Price Index (2015 = 100).
148
Table B.11: Robustness - Likelihood of Divorce/Separation (2SLS estimates)
Divorced/Separated = 1
GNP score (Std.) -0.177
(0.159)
Individual FE Yes
Controls Yes
Num. of Obs. 14,701
Notes: ***p < 0.01, **p < 0.05, *p < 0.1. Standard errors clustered at the individual level in parentheses. This table reports
the estimated results from the linear probability model with endogenous regressors (2SLS). The sample is 5,158 married women
in Wave 5. The dependent variable is a dummy which takes value 1 if a woman gets divorced or separated in yeart.
149
B.2 TheDifferencesintheNewspapers’FramingofFeminism
In this section, we illustrate the contents of the feminism-related articles analyzed in the main text. Since
Korean newspapers’ ignorance of feminism issues was the most prominent characteristic that hindered
feminism from being at the center of social discourse, it should be emphasized again that the number of
feminism-related articles is an essential measure of media influence. However, it is also worth looking into
how newspapers framed the issue differently since the qualitative gaps between conservative and liberal
newspapers might have had even more substantial impacts on women’s perceptions of gender norms and
resource allocation. Here we focus onwhere andunderwhichcontext feminism was reported. More specif-
ically, we can gauge the framing strategies of the newspapers by examining (i) the main subjects reported
under the category of feminism and (ii) sections where the articles were placed.
Keysubjectsreportedunderfeminismcategorizationin2018
We compare the contents of 301 feminism-related articles published by two representative newspapers
(121 in the conservativeChosun and 180 in the liberalHankyoreh) in 2018,
28
when the conservative news-
papers also expanded their quantitative coverage of feminism-related issues. We codify the main sub-
jects of feminism-related articles into 14 categories – social movement, celebrity scandal, media review,
discrimination/diversity, violence/crime, traditional gender norms, backlash/misogyny, radical feminism,
feminism politics, feminism figures, international trend, education, abortion, and fashion. Although these
subjects are not always mutually exclusive, it is not difficult to detect apparent discrepancies between the
two newspapers.
Table B.12 shows the number of articles under each subject. In 2018, a series of pro-feminist protests
in Hyewha station and the Me Too movement were the top news of the year in both newspapers. 26%
of the articles in Chosun were about disputes on celebrities’ scandals related to their attitudes toward or
28
The total number of articles that mentioned feminism at least once was 203 in Chosun and 306 in Hankyoreh. Among 509
articles, 208 that simply mentioned the feminism without focusing on the issue are excluded from this content analysis.
150
against feminism. On the other hand, while being parsimonious about reporting those scandals,Hankyoreh
diversified its focus on feminism politics, education, and issues such as abortion rights. While Chosun
provided extensive coverage on radical feminism and its repercussions (7.4%), Hankyoreh assigned more
proportions to men’s reactionary behaviors, including anti-feminism movements (8.9%). This indicates
that conservative newspapers were more likely to frame feminism as a still-disputable ideology, while
liberal newspapers adhere to a firmer stance supporting it.
Table B.12: Subjects of Feminism-related Articles in 2018
Subjects Issues reported Number of articles
Chosun Hankyoreh
Social Movement Feminism protests, Me Too, Escape the corset 33 (27.3%) 51 (28.3%)
Celebrity Scandal Celebrities’ public attitudes toward feminism 32 (26.4%) 5 (2.8%)
Media Review Book, movie, play, TV show reviews 19 (15.7%) 55 (30.6%)
Discrimination/
Diversity
Discrimination in the workplace, Gender-biased job
positions
6 (5.0%) 9 (5.0%)
Violence/
Crime
Gangnam station homicide, Digital violence, Hidden
camera crimes
3 (2.5%) 4 (2.2%)
Traditional Norms Patriarchy, Traditional gender biases 2 (1.7%) 4 (2.2%)
Backlash/
Misogyny
Anti-feminist movement and websites 10 (8.3%) 16 (8.9%)
Radical Feminism Radical feminist websites and crimes 9 (7.4%) 4 (2.2%)
Feminism Politics Feminist candidates in the national election 1 (0.8%) 8 (4.4%)
Feminism Figures Stories and interviews of feminist leaders 2 (1.7%) 3 (1.7%)
International Trend International Me Too movement, Movement against
genital mutilation
3 (2.5%) 7 (3.9%)
Education Feminism education in schools 1 (0.8%) 8 (4.4%)
Abortion Pro-abortion movement 0 (0%) 5 (2.8%)
Fashion Feminism-associated fashion trends 0 (0%) 1 (0.6%)
121 (100%) 180 (100%)
Notes: This table shows the distribution of 301 feminism-related articles published by two newspapers in 2018 on 14 subjects
specified by the author.
151
Figure B.1: Sections Where Feminism-focusing Articles Were Located
(a) Chosun (conservative)
(b) Hankyoreh (liberal)
Notes: These figures show the proportion of feminism-related articles in each newspaper section. 421 articles – 29 in 2014, 91 in
2016, and 301 in 2018 – are analyzed.
Sectionswherethefeminism-relatedarticleswerelocated
All newspaper articles are classified into sections, such as politics, economy, international, social, culture,
and sports. A newspaper locates an article in a section according to the aspect it tries to highlight about
the issue.
Figure B.1 shows several interesting features. First, the proportion of the articles in the culture sec-
tion declined over time, while that in the social section increased substantially in both newspapers. This
indicates that in the past, newspapers had treated feminism as a strand of culture usually advocated or
displayed through cultural modes, including books, plays, and TV shows. In the mid-2010s, the rise of
feminism began to be reported in the social section as a social movement that generates new gender norms
152
which are fundamentally different from traditional gender prejudices. Moreover, the word feminism began
to appear in more diverse sections, including politics, economy, and sports.
Within a similar context, we can also recognize a sharp gap in the allocation of the articles between
the two newspapers. In 2018, a much more proportion was located in the culture section for Chosun
than Hankyoreh. In addition, Hankyoreh maintained a stable portion of articles in op-eds during the steep
increase in the total number of articles. Through op-eds written by feminist authors, it kept conveying
socially influential messages supporting feminism.
153
C AppendixtoChapter3
C.1 AppendixTablesandFigures
Table C.1: Measuring Time Use: Results
Panel A: All Tasks
(i)Production (59.3%) Supervising (2.9%) Looking for workers (0.0%)
Preparation (Main) (2.7%) Training (1.3%) Managing loans (0.0%)
Processing (Main) (8.1%) Book-keeping (0.6%) Other non-prod. tasks (0.9%)
Finalizing (Main) (2.0%) Maintanence (0.4%)
Producing other prod. (46.5%) Organizing stock (4.6%) (iii)Idle (24.6%)
Procuring inputs (2.3%) Eating/Resting (13.3%)
(ii)Non-prod. Tasks (16.1%) Looking for input supp. (0.7%) Waiting for customers (10.6%)
Interacting with customers (3.4%) Looking for new mach. (0.1%) Away not for business (0.7%)
Panel B: Production Steps
(i)Carpentry (ii)Welding (iii)Grainmilling
Design (6.8%) Design (10.0%) Cob shelling (5.1%)
Drying (before prod.) (19.4%) Cutting (16.2%) Drying (9.5%)
Cutting (9.7%) Bending (11.3%) Cleaning/Destoning (13.3%)
Planing (9.5%) Grinding (13.0%) Conditioning (13.3%)
Thicknessing (6.1%) Welding (25.8%) De-hulling (22.3%)
Edging (8.2%) Polishing (12.5%) Milling (26.4%)
Sanding (11.3%) Painting (11.2%) Sealing (10.0%)
Mortising (11.3%)
Finishing (9.6%)
Drying (after painting) (7.9%)
Notes: The table reports the complete list of tasks reported in our time use data. Panel B breaks down production into the pre-
specified steps. The steps are listed in the typical order of implementation. Steps classified as “Preparation (Main product)” in
Panel A correspond to the following steps listed in Panel B: (i) Carpentry: Design∼ Drying (before production), (ii) Welding:
Design, (iii) Grain milling: Cob shelling∼ Conditioning. Steps classified as “Processing (Main product)” in Panel A correspond
to the following steps listed in Panel B: (i) Carpentry: Cutting∼ Mortising, (ii) Welding: Cutting∼ Welding, (iii) Grain milling:
De-hulling∼ Milling. Steps classified as “Finalizing (Main product)” in Panel A correspond to the following steps listed in Panel
B: (i) Carpentry: Finishing∼ Drying (after painting), (ii) Welding: Polishing∼ Painting, (iii) Grain milling: Sealing.
154
Table C.2: Comparison with IKEA Prices
Morbylanga Table
PanelA.IKEApricesaroundtheworld(USD)
USA 699
India 563
Morocco 728
Egypt 733
PanelB.PriceforsimilarproductsinUganda(USD)
Our survey data (average) 134
Quote from high-end carpenter 414
Notes: Panel A provides a comparison of prices for the IKEA Morbylanga dining table across various IKEA showrooms in the US
(Chicago), India, Morocco and Egypt. All prices are quoted in nominal USD terms. Panel B shows prices for a comparable product
using our survey data and a quote from a high-end carpenter in Uganda. Quality-adjusted prices from survey data collected in
2018-19 have been adjusted for inflation using Consumer Price Indices from the Bank of Uganda. For purposes of comparability,
the price quoted by the Ugandan carpentry firm was based on exact specifications from the IKEA Morbylanga product catalog.
155
Table C.3: Heterogeneity in Skill Intensity of Tasks
(Log) Salary
(1) (2) (3)
Time Share Non-prod. Tasks 0.297 0.293
(0.073) (0.081)
Supervise/Train (0/1) 0.200
(0.061)
Customer Int. (0/1) 0.081
(0.038)
Input procurement (0/1) 0.106
(0.040)
Org. Stock (0/1) -0.004
(0.036)
Other Managerial Tasks (0/1) -0.022
(0.067)
Avg. Complexity of Prod. Steps Performed 0.166
(0.031)
Firm FE Yes Yes Yes
Demographic Ctrl. Yes Yes Yes
AdjustedR
2
0.540 0.545 0.551
Observations 2,350 2,350 2,022
Notes: OLS regression coefficients, robust standard errors in parentheses. Sample: all interviewed employees with non-missing
time use responses. Column (2): A regression result with dummy variables taking 1 if an employee performs each task (baseline
= employees who do not perform non-production tasks.) Column (3): Avg. Complexity of Prod. Steps Performed is computed as
the weighted sum of each employee’s time spent on production steps with the difficulty level of each step. The difficulty level of
steps are computed as the aggregate of the inverse of each employees’ self-reported skills to achieve satisfactory level of quality.
156
Table C.4: Heterogeneity in Skill Distribution within the Firm
Yrs. schooling Age Tenure
(1) (2) (3) (4) (5) (6)
Entrepreneur (0/1) 1.035 11.961 6.630
(0.142) (0.392) (0.259)
Skilled (0/1) 0.532 3.239 1.295
(0.138) (0.383) (0.149)
Sample Ent.+Emp. Emp. Ent.+Emp. Emp. Ent.+Emp. Emp.
Firm FE Yes Yes Yes Yes Yes Yes
AdjustedR
2
0.222 0.292 0.445 0.337 0.493 0.385
Observations 3,929 2,558 3,914 2,548 3,979 2,576
Notes: OLS regression coefficients, robust standard errors in parentheses. Sample: all interviewed entrepreneurs and employees
(odd columns.) all interviewed employees (even columns.) The classification between skilled and unskilled employees are based
on whether an employee’s salary is above median among employees in each firm.
157
Figure C.1: Matching in the Labor Market
0
.05
.1
.15
Density
0 5 10 15
Worker Years of Schooling
Entrepreneur Z-Score: Below Median Entrepreneur Z-Score: Above Median
The Figure shows the density of employee years of education by the managerial ability of score of the entrepreneurs who employ
them, splitting by above or below the median. Sample: all sectors.
158
Figure C.2: Time Allocation within Production across Production Steps
(a) Carpentry
0
20
40
60
80
100
Share of time (%)
2 3 4 5 6 7 8 9 10+
Firm Size
Hours in Production
Thicknessing
Design
Edging
Planing
Mortising
Cutting
Drying (before production)
Finishing
Sanding
Drying (after painting)
(b) Welding
0
20
40
60
80
100
Share of time (%)
2 3 4 5 6 7 8 9 10+
Firm Size
Hours in Production
Design
Bending
Welding
Cutting
Polishing
Painting
Grinding
(c) Grain milling
0
20
40
60
80
100
Share of time (%)
2 3 4 5 6 7 8 9 10+
Firm Size
Hours in Production
De-hulling
Milling
Sealing
Conditioning
Cleaning/Destoning
Drying
Cob shelling
Notes: Sample: all surveyed firms. Time use reported by interviewed entrepreneurs and employees. All figures are weighted by
sampling weights within each sector and the relative number of surveyed firms for sectors. The production steps are ordered in
the level of self-reported difficulty.
159
Figure C.3: Time Allocation within Idle Time
0
20
40
60
80
100
Share of time (%)
1 2 3 4 5 6 7 8 9 10+
Idle Hours
Eat/rest
Away not for business
Waiting for customers
Notes: Sample: all surveyed firms. Time use reported by interviewed entrepreneurs and employees. All figures are weighted by
sampling weights within each sector and the relative number of surveyed firms for sectors.
160
Figure C.4: Task Allocation between Production and Non-production Tasks by Employee Skills
0 20 40 60 80 100
Share of time (%)
managerial tasks
look for new machines
maintanence
others
look for input suppliers
train other workers
book-keeping
organize stock
procure inputs
interact with customers
supervise other workers
wait for customers
production
away not for business
eating or resting
Skilled Employee Unskilled Employee
Notes: The replication of Figure 3 in the main text. Dark bars: skilled employees. Light bars: unskilled employees. Sample: all
surveyed firms. Time use reported by interviewed employees. The classification between skilled and unskilled employees are
based on whether an employee’s salary is above median among employees in each firm.
161
Figure C.5: Employee Contribution to Demand Generation by Firm Size
first-time customers from owner more than emp
owner in charge of bargaining
owner follows up with customers
customers complain to owner
index
-.2 -.15 -.1 -.05 0 .05
discuss with person producing imp. reason for orders
customer has phone number of person producing
worker performs independent orders
index
-.2 -.1 0 .1 .2 .3
Notes: The figure coefficients of the OLS regressions of the survey responses on log firm size with region and sector fixed effects.
Regressions are weighted by sampling weights. Indices in the final rows represent the average of the above responses, respectively.
162
Figure C.6: Heterogeneity in Share of Firms Performing a Production Step by Sector
0
.2
.4
.6
.8
1
Average Share of Firms Doing a Step
1 2 3 4 5 6 7 8 9 10+
Firm Size
Carpentry
0
.2
.4
.6
.8
1
1 2 3 4 5 6 7 8 9 10+
Firm Size
Welding
0
.2
.4
.6
.8
1
1 2 3 4 5 6 7 8 9 10+
Firm Size
Grain Milling
Notes: Replication of Figure 2, Panel (a), in the main text, separately for each sector.
Figure C.7: Time Allocation to Non-production tasks (Heterogeneity by Sector)
0
.2
.4
.6
.8
1
Share of Time in Non-Production Tasks
1 2 3 4 5 6 7 8 9 10+
Firm Size
Carpentry
Welding
Grain milling
Employees
0
.2
.4
.6
.8
1
Share of Time in Non-Production Tasks
1 2 3 4 5 6 7 8 9 10+
Firm Size
Carpentry
Welding
Grain milling
Entrepreneurs
Notes: We replicate panels (a) and (b) of Figure 3.4, but separately for each sector. All surveyed firms. Shaded areas: 95%
confidence intervals. The size of dots and squares represent the number of firms in each size group. Time use reported by
interviewed entrepreneurs and employees.
163
Figure C.8: Heterogeneity in Allocation within Production between Steps
0
.2
.4
.6
.8
1
Average Share of Employees Performing a Production Step
2 3 4 5 6 7 8 9 10+
Firm Size
Full Specialization: Carpentry Full Specialization: Welding Full Specialization: Grain Milling
Employees
0
.2
.4
.6
.8
1
Share of Entrepreneurs Working on a Production Step
2 3 4 5 6 7 8 9 10+
Firm Size
Carpentry Welding Grain Milling
Entrepreneurs
Notes: These figures depict the share of employees (left panel) and entrepreneurs (right panel) working on a production step by
sector. The navy, green, and red lines correspond to the carpentry, welding and grain milling sectors, respectively. In each firm
we calculate the share of persons working on each production step performed, and construct an average measure over steps,
weighing by their time intensity in the production process. The navy, red and green diamond markers in the employee panel
represent the share of employees that would work on a production step under full specialization (reported for firms of size 6
and 10 only). To compute this, we simply reassign employees across production steps to minimize the overlap in steps between
employees, while keeping the total amount of time worked by employees in production within the firm fixed.
164
C.2 AttritionfromtheFollow-upSurveys
Table C.5: Attrition Table – Owners
Dependent variable: Attrited (0/1)
(1) (2) (3) (4) (5) (6)
CW (0/1) -0.221 -0.247 -0.240 -0.215 -0.280 -0.254
(0.052) (0.051) (0.051) (0.056) (0.090) (0.095)
Manager Ability (Std.) 0.027 0.033 0.008 0.019
(0.017) (0.017) (0.043) (0.043)
Firm Size -0.008 -0.010 -0.012 -0.012
(0.005) (0.005) (0.009) (0.009)
CW× Man. Ability 0.022 0.016
(0.046) (0.047)
CW× Firm Size 0.005 0.003
(0.010) (0.010)
Mean of dep. var. 0.322 0.322 0.322 0.322 0.322 0.322
Subcounty FE Yes Yes Yes Yes Yes Yes
AdjustedR
2
0.050 0.049 0.054 0.049 0.049 0.052
Observations 1,101 1,101 1,101 1,101 1,101 1,101
Notes: OLS regression coefficients, robust standard errors in parentheses.
Ownersurvey In the follow-up phone survey conducted in Spring 2022, the owners of 1,101 firms with
phone numbers from the baseline survey were targeted. The attrition rate was 32%, which is in line with
other firm surveys in developing countries.
29
In general, owners in larger firms and those with higher
managerial ability showed less attrition. In addition, attrition was higher in the grain milling sector than
in other two sectors. However, there was no differential attrition in terms of size and managerial ability
across sectors, limiting sample selection bias in evaluating dynamics by comparing the three sectors.
29
For example, the attrition rate is 38% in McKenzie and Woodruff (2008), 36% in Dupas and Robinson (2013) and 45% in
McKenzie, Assaf, and Cusolito (2017). The business training experiments summarized in McKenzie and Woodruff (2014) show
attrition rates ranging from 5% to 46%.
165
Table C.6: Attrition Table – Employees
(1) (2) (3) (4) (5) (6)
VARIABLES attrited attrited attrited attrited attrited attrited
CW -0.098 -0.104 -0.090 -0.094 -0.065
(0.042) (0.043) (0.043) (0.042) (0.039)
Manager Ability (Std.) -0.009 -0.007 -0.009 -0.012
(0.015) (0.015) (0.015) (0.014)
Years of Schooling -0.005 -0.006 -0.008 -0.009
(0.004) (0.004) (0.004) (0.007)
Age 0.003 0.002 -0.003 -0.007
(0.008) (0.008) (0.007) (0.014)
Age Squared -0.000 -0.000 0.000 0.000
(0.000) (0.000) (0.000) (0.000)
Tenure at the Firm (Yrs) -0.012 -0.012 -0.009 -0.006
(0.004) (0.004) (0.003) (0.007)
Vocational Training (0/1) 0.037 0.045 0.037 0.029
(0.038) (0.037) (0.036) (0.060)
Cognitive Score (0-4) -0.018 -0.009 0.008
(0.011) (0.010) (0.018)
Agreeableness (1-5) 0.027 0.020 0.007
(0.017) (0.016) (0.031)
Conscientiousness (1-5) 0.021 0.030 0.040
(0.019) (0.018) (0.028)
Extraversion (1-5) 0.009 0.011 0.040
(0.015) (0.015) (0.024)
Openness (1-5) 0.016 0.009 -0.015
(0.023) (0.022) (0.034)
Neuroticism (1-5) (reversed scale) -0.010 -0.013 0.002
(0.017) (0.016) (0.031)
Log Firm Size -0.019
(0.027)
Owner attrited 0.050
(0.032)
Observations 2,177 2,177 2,177 2,177 2,177 2,177
AdjustedR
2
0.038 0.039 0.048 0.051 0.137 0.345
Subcounty FE Yes Yes Yes Yes Yes No
Firm FE No No No No No Yes
F-test Employee Char 0 0 0.0100 0.750
Notes: OLS regression coefficients. Standard errors (in parentheses) are clustered at the firm level.
166
Employeesurvey In the follow-up phone survey wave, we leveraged the dyadic nature of our baseline
survey data and attempted to interview 2,177 employees who originally worked for 899 of our sample firms
in 2018-19. The resulting attrition rate was 41%, which is reasonable considering the survey mode (phone
interviews) and over a three-year gap since we originally interviewed these employees. To try and lower
attrition from the panel survey, our field team reached out to successfully interviewed firm owners for
assistance to contact their unreachable employees (and vice-versa). Table C.6 highlights whether employee
and firm characteristics are predictive of attrition from the employees’ sample. On average, we were 9-10
percentage points less likely to interview employees in the grain milling sector compared to carpentry
and welding. This pattern is linked to differential attrition among both firm owners and employees in the
milling sector (column 4). The managerial ability of the firm owner and the size of the firm (in 2018-19) are
not correlated with higher attrition among employees. Moreover, employee skills do not jointly predict
higher attrition after controlling for firm fixed effects (p-value of the F-test: 0.75).
167

Abstract (if available)

Abstract In this dissertation, I adopt advanced microeconometric tools to analyze three important topics in modern labor economics. In the first chapter, I discuss the role of incomplete information during a job search in occupational segregation by gender. I administered a randomized experiment with female Korean undergraduates and provided half of them with detailed statistics for representative male- and female-dominated industrial sectors. The treatment group reported substantially less biased beliefs on sector characteristics and showed relatively higher aspirations to enter jobs in male-dominated sectors than the control group. In the second chapter, I examine what happened to household resource allocation when gender norms began to change. By collecting newspaper articles about feminism in the past ten years in Korea and exploiting their region-year variations, I provided evidence that increased newspaper coverage of feminism contributed to egalitarian attitudes among women. Then, I constructed a shift-share instrumental variable with the growth of the feminism-related articles to show that the change in women's perceptions of gender norms induced by those articles affected both spouses' time use in household labor and women's welfare. Finally, in the third chapter, I zoom in on the internal organization of manufacturing firms in Uganda to analyze the relationship between limited specialization and scalability. Based on detailed time-use data, I estimated the structural model of labor specialization within firms and demonstrated that the internal barrier to specialization – delegation cost – was the key to understanding low returns to scale, resulting in small firm size in developing countries.

Linked assets

University of Southern California Dissertations and Theses

Conceptually similar

PDF

Three essays on macro and labor finance

PDF

Essays in labor economics: demographic determinants of labor supply

PDF

Essays on applied microeconomics

PDF

Essays on development economics and adolescent behavior

PDF

Essays in development economics

PDF

Essays on political economy and corruption

PDF

Essays on development and health economics

PDF

The impact of minimum wage on labor market dynamics in Germany

PDF

Essays in macroeconomics

PDF

Essays on the new economic history of the Middle East

PDF

Three essays on health & aging

PDF

Essays in the economics of education and conflict

PDF

Essays on effects of reputation and judicial transparency

PDF

Two essays on financial crisis and banking

PDF

Three essays on cooperation, social interactions, and religion

PDF

Against the wind: labor force participation of women in Iran

PDF

Essays on family and labor economics

PDF

Essays on macroeconomics and income distribution

PDF

Three essays on heterogeneous responses to macroeconomic shocks

PDF

Reproducing inequity in organizations: gendered and racialized emotional labor in pubic organizations

Asset Metadata

Creator Lee, Jung Hyuk (author)

Core Title Three essays on the microeconometric analysis of the labor market

School College of Letters, Arts and Sciences

Degree Doctor of Philosophy

Degree Program Economics

Degree Conferral Date 2023-05

Publication Date 03/29/2023

Defense Date 03/09/2023

Publisher University of Southern California (original), University of Southern California. Libraries (digital)

Tag feminism,firm growth,firm organization,firm productivity,gender norms,household labor,information frictions,job search,labor specialization,OAI-PMH Harvest,occupational segregation,time use

Format theses (aat)

Language English

Contributor Electronically uploaded by the author (provenance)

Advisor Bassi, Vittorio (committee chair), Duquette, Nicolas J. (committee member), Nugent, Jeffrey B. (committee member), Weaver, Jeffrey B. (committee member)

Creator Email jay.junghyuk.lee@gmail.com,leejungh@usc.edu

Permanent Link (DOI) https://doi.org/10.25549/usctheses-oUC112850436

Unique identifier UC112850436

Identifier etd-LeeJungHyu-11531.pdf (filename)

Legacy Identifier etd-LeeJungHyu-11531

Document Type Dissertation

Format theses (aat)

Rights Lee, Jung Hyuk

Internet Media Type application/pdf

Type texts

Source 20230330-usctheses-batch-1013 (batch), University of Southern California (contributing entity), University of Southern California Dissertations and Theses (collection)

Access Conditions The author retains rights to his/her dissertation, thesis or other graduate work according to U.S. copyright law. Electronic access is being provided by the USC Libraries in agreement with the author, as the original true and official version of the work, but does not grant the reader permission to use the work if the desired use is covered by copyright. It is the author, as rights holder, who must provide use permission if such use is covered by copyright. The original signature page accompanying the original submission of the work to the USC Libraries is retained by the USC Libraries and a copy of it may be obtained by authorized requesters contacting the repository e-mail address given.

Repository Name University of Southern California Digital Library

Repository Location USC Digital Library, University of Southern California, University Park Campus MC 2810, 3434 South Grand Avenue, 2nd Floor, Los Angeles, California 90089-2810, USA

Repository Email cisadmin@lib.usc.edu