University of Southern California Dissertations and Theses

Essays on development and health economics: social media and education policy

(USC Thesis Other)

Essays on development and health economics: social media and education policy

PDF

Download a page range

Download transcript

Copy asset link

Request this asset

Transcript (if available)

Content ESSAYS ON DEVELOPMENT AND HEALTH ECONOMICS:
SOCIAL MEDIA AND EDUCATION POLICY
by
Qin Jiang
A Dissertation Presented to the
FACULTY OF THE GRADUATE SCHOOL
UNIVERSITY OF SOUTHERN CALIFORNIA
In Partial Fulfillment of the
Requirements for the Degree
DOCTOR OF PHILOSOPHY
(ECONOMICS)
August 2021
Copyright 2021 Qin Jiang
Acknowledgements
I would like to express appreciation and warmest thanks to my advisor Professor John Strauss.
This work would not have been possible without his encouragement, friendly guidance, and in-
valuable advice. I am grateful for his wholehearted support of my research and career goals.
I would like to thank Professor Arie Kapteyn for sharing his wisdom, providing resources,
and helping me find the right directions countless times. I would like to thank Professor Jeffrey
Weaver, who always supports me in research, in career, and in life. I would like to thank Professor
Geert Ridder and Professor Gary Painter for their valuable feedback on my dissertation and de-
fense. Their suggestions have greatly improved this work. I would like to thank Professor Michael
Magill for his trust and support. I cherish the wisdom he shares with me. I would like to thank Pro-
fessor Vittorio Bassi, Professor Paulina Oliva, Professor Isabelle Brocas, Professor Juan Carrillo,
Professor Giorgio Coricelli, Professor Jeffrey Nugent, Professor Yu-Wei Hsieh, Professor Matthew
Kahn, and Professor Emily Nix for their constructive feedback and support at different stages of
my research. I would also like to thank administrative staff of the Department of Economics at
USC: Young Miller, Alexander Karnazes, Anna Emerald, Morgan Ponder, and Fatima Perez, for
their kindly administrative support. I sincerely thank all my colleagues and friends for their support
both in research and in life.
Special thanks are due to my parents for their understanding and support. They have taught me
the true meaning of independent thinking, the power of optimism, and the wisdom of happiness.
My thanks are extended to my loving boyfriend Haonan Lu, who is my strongest support. Last but
not least, thanks to my grandparents whom I will always miss.
ii
Table of Contents
Acknowledgements ii
List of Tables vi
List of Figures viii
Abstract x
Introduction 1
Chapter 1: Social Media Usage and the Level of Depressive Symptoms in the United
States 4
1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.2 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.2.1 Depression: Causes, Consequences, Measures, and Heterogeneity . . . . . 8
1.2.2 Social Media Usage in the United States . . . . . . . . . . . . . . . . . . . 9
1.3 Data: Understanding America Study . . . . . . . . . . . . . . . . . . . . . . . . . 10
1.3.1 Social Media Usage . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
1.3.2 The Level of Depressive Symptoms . . . . . . . . . . . . . . . . . . . . . 13
1.3.3 Social Media Users and Non-users . . . . . . . . . . . . . . . . . . . . . . 14
1.4 Empirical Strategy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
1.4.1 Neyman–Rubin Causal Model Framework . . . . . . . . . . . . . . . . . . 14
1.4.1.1 Challenges . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
1.4.2 Two-Way Fixed Effects Model . . . . . . . . . . . . . . . . . . . . . . . . 15
1.5 Empirical Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
1.5.1 Direct Evidence on US Population’s Feeling about Using Social Media . . 16
1.5.2 Social Media Usage Decreases the Level of Depressive Symptoms . . . . . 17
1.5.2.1 Effects on Each of Eight CESD Item . . . . . . . . . . . . . . . 18
1.5.3 Test Fixed Effect Assumptions . . . . . . . . . . . . . . . . . . . . . . . . 18
1.5.3.1 Potential Confounders . . . . . . . . . . . . . . . . . . . . . . . 18
1.5.3.2 Bias Estimation With Altonji-Oster Model . . . . . . . . . . . . 19
1.5.3.3 Test CESD Trends of Switchers . . . . . . . . . . . . . . . . . . 20
1.5.3.4 Test Potential Mediators . . . . . . . . . . . . . . . . . . . . . . 21
1.5.4 Robustness Checks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
1.5.4.1 SUTV A . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
iii
1.5.4.2 First Difference Estimation on Switchers . . . . . . . . . . . . . 22
1.5.4.3 Outliers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
1.5.4.4 Attrition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
1.5.4.5 Suggestive Evidences on “Good Social Media” . . . . . . . . . . 24
1.5.5 Heterogeneous Effect . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
1.5.6 Labor Market Benefits . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
1.5.7 Social Media Usage and Depressive Symptoms in 2020 . . . . . . . . . . . 28
1.6 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
1.7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
Chapter 2: Selection Bias Reduction by Matching and Weighting Estimators in Social
Media Data Collection 65
2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
2.2 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
2.2.1 Social Media Survey Design . . . . . . . . . . . . . . . . . . . . . . . . . 70
2.2.2 Twitter and Bias in Social Media Survey Sample . . . . . . . . . . . . . . 71
2.3 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
2.3.1 Understanding America Study (UAS) . . . . . . . . . . . . . . . . . . . . 72
2.3.2 Income, Number of Family and Close Friends, and Willingness to Pay for
COVID-19 Vaccine . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
2.3.3 Summary Statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
2.4 Empirical Strategy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
2.5 Empirical Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
2.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
Chapter 3: School Starting Age Effect and Noncompliance Behavior: Evidence from
China 92
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
3.2 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
3.2.1 School Starting Age in Compulsory Education Law of the People’s Re-
public of China, Article 5 . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
3.2.2 Relative Age Effect . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
3.3 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
3.3.1 China Family Panel Studies (CFPS) . . . . . . . . . . . . . . . . . . . . . 97
3.3.2 China Education Panel Survey (CEPS) . . . . . . . . . . . . . . . . . . . 97
3.3.3 Compliance With “Age Rule” by Month of Birth: CFPS and CEPS . . . . . 98
3.4 Empirical Strategy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
3.5 Empirical Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
3.5.1 Group Characteristics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
3.5.2 School Starting Age Effect . . . . . . . . . . . . . . . . . . . . . . . . . . 102
3.5.3 Relative Age Effect . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
3.5.4 Effect of “School Starting Age Rule” Implementation . . . . . . . . . . . . 103
3.6 Potential Explanations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
3.7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
iv
Conclusion 115
References 117
Appendix 123
A Appendix to Chapter 1: Social Media Application Updates Instrument . . . . . . . 124
A.1 Construct the Instrument . . . . . . . . . . . . . . . . . . . . . . . . . . . 124
A.2 First Stage: New Updates Change Social Media Usage . . . . . . . . . . . 125
A.3 Second Stage: Causal Effect . . . . . . . . . . . . . . . . . . . . . . . . . 125
A.4 Instrument Relevance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126
A.5 Exclusion Restriction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126
A.6 Data: Social Media Application Updates . . . . . . . . . . . . . . . . . . 126
A.7 IV Estimation With 2020 Data . . . . . . . . . . . . . . . . . . . . . . . . 127
B Appendix to Chapter 1: Bartik Instrument . . . . . . . . . . . . . . . . . . . . . . 134
B.1 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134
B.2 Data: Social Media Usage Growth . . . . . . . . . . . . . . . . . . . . . . 134
B.3 Bartik Instrument Result . . . . . . . . . . . . . . . . . . . . . . . . . . . 135
C Additional Results to Chapter 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137
D Appendix to Chapter 3: Cognitive Skill Measures in CEPS . . . . . . . . . . . . . 147
v
List of Tables
1.1 UAS Survey Information . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
1.2 Summary Statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
1.3 Descriptive Statistics - CESD & PHQ . . . . . . . . . . . . . . . . . . . . . . . . 50
1.4 Social Media Users Versus Non-users . . . . . . . . . . . . . . . . . . . . . . . . 51
1.5 Descriptive Statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
1.6 Social Media Usage Decreases the Level of Depressive Symptoms . . . . . . . . . 53
1.7 Social Media Usage Decreases the Level of Depressive Symptoms . . . . . . . . . 54
1.8 No Pre-trend in CESD . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
1.9 Switcher Groups’ Characteristics . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
1.10 Income and Marital Status Are Not Mediators . . . . . . . . . . . . . . . . . . . . 57
1.11 Social Media Usage Decreases the Level of Depressive Symptoms (Main Member
Only) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
1.12 Estimated Effect of Using Twitter on CESD . . . . . . . . . . . . . . . . . . . . . 59
1.13 Estimation for Two Switcher Groups . . . . . . . . . . . . . . . . . . . . . . . . . 60
1.14 Fixed Effect Model with CESD Winsorization . . . . . . . . . . . . . . . . . . . . 60
1.15 Attrition Statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
1.16 Effect of Social Media Usage on Depressive Symptoms (2020) . . . . . . . . . . . 62
1.17 Effect of Social Media Usage (2020) . . . . . . . . . . . . . . . . . . . . . . . . . 63
1.18 Previous CESD Does Not Change Twitter Usage . . . . . . . . . . . . . . . . . . 63
1.19 Depression Alleviated . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
1.20 Computing Labor Market Benefit . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
vi
2.1 Bias Types . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
2.2 Summary Statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
2.3 Methods Performance (T Statistic)- Facebook . . . . . . . . . . . . . . . . . . . . 90
2.4 Methods Performance (T Statistic)- Twitter . . . . . . . . . . . . . . . . . . . . . 91
3.1 Group Differences (CFPS) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
3.2 Group Differences (CEPS) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
3.3 Group Differences-Multinomial Ordered Logit . . . . . . . . . . . . . . . . . . . 110
3.4 School Starting Age Effect: CFPS . . . . . . . . . . . . . . . . . . . . . . . . . . 111
3.5 School Starting Age Effect: CEPS . . . . . . . . . . . . . . . . . . . . . . . . . . 111
3.6 School Starting Age Effect: CEPS . . . . . . . . . . . . . . . . . . . . . . . . . . 112
3.7 Regression Discontinuity Estimation: CEPS . . . . . . . . . . . . . . . . . . . . . 112
3.8 Relative Starting Age Effect: CEPS . . . . . . . . . . . . . . . . . . . . . . . . . 113
3.9 Relative Starting Age Effect: CEPS . . . . . . . . . . . . . . . . . . . . . . . . . 113
3.10 Law Implementation Effect: CFPS . . . . . . . . . . . . . . . . . . . . . . . . . . 114
A.1 Arbitrary Survey Dates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130
A.2 Effect of Social Media Usage on Depressive Symptoms (2020): IV . . . . . . . . . 131
A.3 Correlation Between 2020 Variables . . . . . . . . . . . . . . . . . . . . . . . . . 132
A.4 IV Distribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133
B.1 Summary of Bartik IV Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . 136
C.1 Depressive Measures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 144
C.2 Social Media Usage and the Level of Depressive Symptoms . . . . . . . . . . . . 145
C.3 Fixed Effect Model With Sub-samples . . . . . . . . . . . . . . . . . . . . . . . . 145
C.4 Effect of Social Media Using Frequency on CESD (2019) . . . . . . . . . . . . . . 146
vii
List of Figures
1.1 Twitter User Growth (2010-2019) . . . . . . . . . . . . . . . . . . . . . . . . . . 31
1.2 Twitter and Facebook User Percentages by Gender and Age Group . . . . . . . . . 32
1.3 CESD Distribution in 2015, 2017, and 2019 . . . . . . . . . . . . . . . . . . . . . 33
1.4 The levels of Depressive Symptoms by Gender and Age Group . . . . . . . . . . . 34
1.5 The Levels of Depressive Symptoms by Age: Users versus Non-users . . . . . . . 35
1.6 How People Feel About Using Social Media . . . . . . . . . . . . . . . . . . . . . 36
1.7 CESD Trends . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
1.8 Heterogeneous Effect: Age and Gender . . . . . . . . . . . . . . . . . . . . . . . 38
1.9 Heterogeneous Effect: Education . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
1.10 Heterogeneous Effect: Income . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
1.11 Heterogeneous Effect: Past Level of Depressive Symptoms . . . . . . . . . . . . . 41
1.12 Heterogeneous Effect: Race . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
1.13 Heterogeneous Effect: Urbanicity . . . . . . . . . . . . . . . . . . . . . . . . . . 43
1.14 Heterogeneous Effect: Region . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
1.15 Time Spent on Social Media . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
1.16 Labor Market Benefit . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
1.17 Numbers of Connections and Followers on Twitter and Facebook . . . . . . . . . . 47
2.1 Facebook Usage and Twitter Usage . . . . . . . . . . . . . . . . . . . . . . . . . . 81
2.2 Distributions: Income, Number of Family/Close Friends, and WTP for Vaccine . . 82
2.3 Distributions: Income, Number of Family/Close Friends, and WTP for Vaccine
(Facebook) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
viii
2.4 Distributions: Income, Number of Family/Close Friends, and WTP for Vaccine
(Twitter) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
2.5 Distributions: Marital Status, Education, and Race . . . . . . . . . . . . . . . . . . 85
2.6 KNN: Number of Neighbors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
2.7 Radius Caliper . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
3.1 Right Age, Early, and Late Group Proportions by Month of Birth (CFPS) . . . . . 106
3.2 Right Age, Early, and Late Group Proportions by Month of Birth (CEPS) . . . . . 107
A.1 Example: Creating Instrumental Variables . . . . . . . . . . . . . . . . . . . . . . 129
C.1 Test Overlap Assumption . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137
C.2 Heterogeneous Effect: Gender, Marital Status . . . . . . . . . . . . . . . . . . . . 138
C.3 Heterogeneous Effect: Past Level of Depressive Symptoms . . . . . . . . . . . . . 139
C.4 Heterogeneous Effect: Working Hours . . . . . . . . . . . . . . . . . . . . . . . . 140
C.5 Heterogeneous Effect: Number of Children . . . . . . . . . . . . . . . . . . . . . 141
C.6 Twitter and Facebook User Percentages by State . . . . . . . . . . . . . . . . . . . 142
C.7 The levels of Depressive Symptoms by State . . . . . . . . . . . . . . . . . . . . . 143
ix
Abstract
This dissertation consists of three essays on development economics and health economics.
The first chapter examines the impact of social media usage on depressive symptoms in the United
States. The use of social media can potentially decrease the level of depressive symptoms by
providing support or increase the level of depressive symptoms by putting social pressure on users.
This chapter leverages a fixed-effects model to estimate the effect of using social media platforms
on depressive symptoms. I find that using Twitter decreases the level of depressive symptoms
by 27%. This result explains why social media usage in the US has grown steadily even though
most studies found that more usage correlated with higher levels of depressive symptoms. There is
heterogeneity with respect to age, income, education, race, previous level of depressive symptoms,
and region. The average labor market benefit that comes from this effect is equivalent to 0.1%
GDP in the US.
In the second chapter, I examine the performances of different bias correction methods, such
as matching and weighting methods, on improving the representativeness of social media data. I
find that matching and weighting methods can effectively improve the representativeness of social
media users in most cases examined. Matching methods with smaller number of neighbors or
smaller radius produce smaller biases. Improving the representativeness of Twitter users is easier
than improving the representativeness of Facebook users.
The third chapter is a collaboration with Yinan Liu, in which we study the impact of the pri-
mary school starting age policy in China on both short-run and long-run outcomes. We examine
the household characteristics of the right age group, early group, and late group based on the
compliance. Starting school late is negatively associated with cognitive skills, test scores, highest
x
education achieved and income. We also explore the potential explanations why a large proportion
of households send children to primary school before they reach the eligible age in China.
xi
Introduction
Social media has become one of the most important inventions in our life. It is not only a great
invention, more importantly, it is also a new media channel which is changing people’s life in a
way that no traditional media has ever done. In 2019, more than 72% of US adults used social
media. Social media has also been playing a critical role during COVID-19 pandemic.
The first chapter examines the impact of social media usage on depressive symptoms in the
United States. People may have completely different experience with social media. The use of
social media can potentially decrease the level of depressive symptoms by providing support or
increase the level of depressive symptoms by putting social pressure on users. It would be inter-
esting to know if good experience dominates the bad experience or vice versa in the population.
So I compile data of US adults’ Twitter usage and depressive symptoms in 2017 and 2019. I use
this panel data and a two-way fixed-effects model and find that using Twitter decreases the level
of depressive symptoms by 27% (0.2 s.d.). This result explains why social media usage in the US
has grown steadily even though most studies found that more usage correlated with higher levels
of depressive symptoms.
In addition, I also quantify the impact of Twitter on people with different ages, income levels,
education levels, races, previous levels of depressive symptoms, and regions. Lastly, I find the
average labor market benefit that comes from this effect is equivalent to 0.1% GDP in the US.
The main contribution of this chapter is that it is the first to study the impact of social media on
depressive symptoms with a large representative sample. In the literature, it is an actively debated
question with no convincing answers yet because existing studies only used non-representative
samples and the data they used may lead their results to high biases. This chapter addresses these
1
problems. Since more than 72% US adults use social media, my results also have large policy
implications. Another exciting contribution of this chapter is the unique data. I also collected
quantitatively data on people’s social media usage, such as the number of minutes people spend on
social media everyday, number of connections and followers on different social media platforms,
and people’s general opinion about using social media. This type of data was never available
before. Hence, these data provide insights on people’s social media usage and suggestive evidence
to support the main finding, which is using social media may mitigate depressive symptoms.
Academic research is inevitably changed by social media as well. In the second chapter, I ex-
plore the potential of social media surveys and test a few methods to address the selection bias is-
sue. I examine the performances of different bias correction methods, including K-Nearest Neigh-
bors Matching (KNN), Radius Matching, Inverse Probability Weighting (IPW), and Mahalanobis
Matching. I find that matching and weighting methods can effectively improve representativeness
of social media users in most cases examined. For example, the differences of income, number of
family or close friends, and willingness to pay for COVID-19 vaccine between social media users
and those who do not use social media are not statistically significant after applying those bias
correction methods. The best bias correction method varies depending on the outcome variable
types. Hyperparameter values in matching methods also affect method performance. Improving
representativeness of Twitter users is easier than improving representativeness of Facebook users.
The third chapter is a collaboration with Yinan Liu, and it was motivated by an interesting
finding of ours. Official primary school starting age is six in China, but we find a large proportion
of children start school before age six. The opposite pattern is more prevalent in the school start-
ing age literature in which researchers find postponing school starting age (red-shirting) is more
common in developed countries. We use two data sources, China Family Panel Studies and China
Education Panel Survey, to explore potential reasons behind this interesting observation.
The Compulsory Education Law of the People’s Republic of China (Article 5) indicates the
appropriate primary school starting age should be six, and the common starting age was seven
before it went into effect in 1986. This rule also comes with an official cut-off date August 31,
2
which means children who reach six before August 31 in the current year should start primary
school in this fall, and children who reach six after August 31 in the current year should start
primary school in the next fall semester. First, we compare respondents’ household characteristics
among three groups: respondents who start primary school at the right age, respondents who start
primary school earlier than six years old, and respondents who start primary school later than six
years sold. The right age group has the best social-economic status and resources, the next is the
early group, and late group has the worst resources. We study the average effect of this age rule
and find mixed evidence.
Next, we examine the effects of school starting age on income and the highest education level
using the CFPS data. Then we also examine the effects of school starting age on test scores and
cognitive skills for children who started primary school from 2003 to 2008 with the CEPS data.
We find that starting school late is negatively associated with cognitive skills, test scores, highest
education achieved and income. We also study the effect of relative age and find evidence that
being a young member in class is positively associated with cognitive skills and test scores.
Last but not least, we discuss potential explanations behind the behavior of “counter-red-
shirting,” i.e., many parents chose to send children to school early in China. The potential ex-
planations of this particular behavior of Chinese parents include labor supply, relative age effect,
flexibility of repeating grades, and social norm.
3
Chapter 1
Social Media Usage and the Level of Depressive Symptoms in
the United States
1.1 Introduction
There is often a love-hate relationship between social media and social media users. Whether
social media makes people happier remains unclear. Most existing studies find a positive correla-
tion between using social media and the level of depressive symptoms and support a “bad social
media” opinion. At the same time, social media user population has grown steadily over the last
decades. It is puzzling that these two trends both exist. This chapter tries to solve this puzzle and
understands how social media affects users’ feelings.
Many features draw people to social media. Social media platforms (for example, Twitter and
Facebook) provide users with news, resources, entertainment, and social support. Social media
brings families and friends who live far away closer and helps strangers become friends. Some
social media platforms offer specific services. For example, LinkedIn provides business- and
employment- oriented online services, and YouTube provides video sharing. Social media of-
fers new lifestyles that many people enjoy. Popular social media platforms attract billions of active
users monthly.
Social media may also negatively impact user’s self-esteem. Spending too much time on social
media makes us feel guilty. Using social media reduces working time or studying time, which then
4
indirectly reduces income. Social media also exposes users to unrealistic views of others’ lives,
which can make users feel more anxious and disappointed about their own lives.
Understanding the effect of social media usage may help individuals make better decisions,
help the society decide whether to encourage more social media usage or not, and help policy
makers better construct policy around social media usage.
This chapter addresses the question of how the use of social media affects the level of depres-
sive symptoms. I use a two-way fixed effects model and find more social media usage decreases
the level of depressive symptoms. This finding explains the steady growth of social media industry
and social media user population. The remaining puzzle is why most existing studies find posi-
tive correlation between social media usage and the level of depressive symptoms. This chapter
explores different explanations for this gap.
This chapter also examines the impact of social media on different depressive symptoms and
in different demographic groups. Using social media consistently decreases depressive symptoms.
The impact is concentrated among people younger than 40, with at least high school education,
with annual household incomes around $20,000 to $30,000, live in small cities, live in Northeast
region of the United States, and have a medium prior level of depressive symptoms.
Most existing studies support that social media usage is positively associated with the level of
depressive symptoms. A few studies used experiments and suggested the positive causal effect. A
random experiment study assigned 143 undergraduates at the University of Pennsylvania to treat-
ment groups and limited their Facebook, Instagram, and Snapchat use to 10 minutes per platform
per day for three weeks (Hunt et al., 2018). The study found that social media usage reduction
decreased loneliness and depression but is for a small and selected group of students. Another
randomized experiment found a similar result: deactivating Facebook for the four weeks before
the 2018 US midterm election increased participants’ subjective well-being (Allcott et al., 2020).
1
However, the second experiment recruited people using a Facebook advertisement and screened
1
Their subjective well-being measures include three questions: 1. Overall, how happy do you feel right now on a
scale from 1 (not at all happy) to 10 (completely happy)? 2. What best describe how you felt over the last 10 minutes
(12 options)? 3. How lonely are you feeling right now on a scale from 1 (not at all lonely) to 10 (very lonely)?
5
samples based on a few criteria before randomization.
2
So its samples were not representative of
US adult population and their finding may be specific to the election period.
A few studies used survey data and found similar implications. For example, studies found
that social networks reduced a UK teenager’s probability of being completely satisfied with life
by approximately 14%, and that girls suffered more adverse effects than boys (McDool et al.,
2016); a study with 1,787 US young adults (ages 19 - 32) found that participants who used 7 to
11 social media sites had higher levels of both depression and anxiety than participants who used
zero to two social media sites (Primack et al., 2017); a study with 180 undergraduate students from
psychology classes from a southwestern university found a positive association between time spent
on Facebook and depressive symptoms and concluded that people would feel depressed when they
compare themselves to other people after spending a great deal of time on Facebook (Steers et al.,
2014).
Few papers found potential evidence that social media usage can be beneficial or at least not
detrimental. For example, a study found that Internet usage could decrease loneliness and depres-
sion significantly, and increased perceived social support and self-esteem significantly (Shaw and
Gant, 2004). However, the samples suffer from small sample and selection bias. They included
only 40 undergraduate students enrolled in an introductory psychology course at the University of
North Carolina at Chapel Hill, and the treatment was five chat sessions with an anonymous part-
ner. Another study used online survey to assess the level of depressive symptoms and an experience
sample method approach to assess social media usage among university students and found no as-
sociations between social media usage and either any depression or moderate to severe depression
(Jelenchick et al., 2013).
This chapter contributes to the literature in the following ways. First, this chapter studies the
impact of social media usage on the level of depressive symptoms in a large sample representing
the US adult population. Existing studies using cross-sectional survey data failed to address many
2
The criteria included that the respondents’ daily Facebook usage were between 15 to 600 minutes, respondents
confirmed willingness to deactivate Facebook for at least 48 hours, and respondents didn’t have any patterns suggesting
careless responses.
6
potential confounders. Experiments can help provide causal estimates, but lab experiments have
limited sample size and the subjects were undergraduate students. Therefore, the implications of
such studies may not be valid for other demographic groups. This chapter addresses both problems
by using a large size panel dataset and a two-way fixed effects model in a general population at a
normal time, with no special events immediately before or after the data collection.
Second, I compile a unique panel data set. Information about both social media usage and the
level of depressive symptoms is available for each individual in 2017 and 2019. This panel data set
enables the two-way fixed effects model. I also collect a rich set of social media usage information
in 2020 which provides supporting evidences to main result, and compile ancillary data sets for
two types of instruments, which are useful for future studies.
Last, this chapter finds social media usage decreases the level of depressive symptoms. This
finding explains why social media industry is growing. The finding can be surprising considering
the prevalent positive association findings from the literature. This chapter provides four explana-
tions for this apparent discrepancy.
The remainder of the chapter is organized as follows. Section 1.2 details depressive symp-
toms, depressive symptom measures, and social media usage in the United States. Section 1.3
summarizes data sources and the main variables of interest. Section 1.4 discusses the two-way
fixed effects model. Section 1.5 discusses multiple instances of empirical evidence to support that
social media has a beneficial impact on people. This section also discusses the robustness checks
of the impact, impact heterogeneity, and labor market outcome estimation. Section 1.6 discusses
this chapter’s findings and explores potential explanations of social media’s beneficial impact. Sec-
tion 1.7 provides the conclusion.
7
1.2 Background
1.2.1 Depression: Causes, Consequences, Measures, and Heterogeneity
Depression is a common mental disorder (WHO, 2020). It relates to sadness and bereavement,
and does not remit when the original causes dissipate (Belmaker and Agam, 2008). The mech-
anism of depression has not been fully established. However, evidences from psychology and
economics literature lend support to the following risk factors: genetics and biology, such as brain
chemistry imbalance, hormones, and poor nutrition; physical health problems, such as chronic ill-
ness and disability; drug and alcohol use; sleep deprivation; and stress from work and life, such
as unemployment, financial problems, and loss of a loved one (A. T. Beck and Alford, 2009; Cole
and Dendukuri, 2003; Quidt and Haushofer, 2017; Lei et al., 2014).
Depressive symptoms include cognitive-effective symptoms and somatic symptoms (Muscatell
et al., 2009). Examples of the former type are sadness, pessimism, and self-dislike. Examples of
the latter type are sleeping problem and weight loss (A. T. Beck and Alford, 2009; Muscatell et
al., 2009). These symptoms can lead to negative consequences in a patient’s personal life, such
as physical illness, social phobia, family conflict, work or school problems, and suicidal feelings.
Such consequences not only place burden on the patient, but also have impact on his or her family
and community.
Researchers can use several measures to assess a subject’s level of depressive symptoms (Radloff,
1977; Smarr and Keefer, 2011). Some examples include Center for Epidemiologic Studies Depres-
sion (CESD) Scale, Beck Depression Inventory-II (BDI-II), Geriatric Depression Scale (GDS),
Hospital Anxiety and Depression Scale (HADS), Patient Health Questionnaire (PHQ), and Com-
posite International Diagnostic Interview (CIDI).
Depressive measure consists of multiple items evaluating severity of depressive symptoms, and
the final score is the sum of all the items. For example, the original CESD Scale (Radloff, 1977) has
20 items assessing perceived mood and the level of functioning during the past week. For example,
one item is “number of days I felt that everything I did was an effort during the past week.” Each
8
item is on a four - point scale, in which 0 = rarely or none of the time (< 1 day), 1 = some or a
little of the time (1 - 2 days), 2 = occasionally or a moderate amount of time (3 - 4 days), and 3
= most or all of the time (5 - 7 days). The final score ranges from 0 to 60. The most commonly
used cutoff for clinical depressive disorder diagnosis is CESD 16 for adult population.
3
The
depressive measures also quantify mental health and/or happiness.
Depressive disorder is one of the most prevalent mental illnesses. More than 264 million peo-
ple suffer from depression worldwide (WHO, 2020). The 12-month and lifetime prevalence of
depression are 7.2% and 10.8%, respectively, between 1994 and 2014 (Lim et al., 2018). The 12-
month and lifetime prevalence of major depressive disorder (MDD) in the US adult population are
10.4% and 20.6%, respectively, between 2012 and 2013 (Hasin et al., 2018).
Prevalence of depressive disorder is higher for women than for men (Brody et al., 2018; Lei
et al., 2014; Luppa et al., 2012). The aggregate prevalence is 14.4% for women and 11.5% for men
between 1994 and 2014 (Lim et al., 2018). In particular, 13% of women experience postpartum
depression (C. T. Beck, 2001). Other high risk populations include teenagers, older adults, and
people with stressful jobs (S. K. Bhatia and S. C. Bhatia, 2007; Lei et al., 2014; Luppa et al.,
2012).
1.2.2 Social Media Usage in the United States
Spending time on social media platforms is common among the US population. In 2019,
the most popular social media platforms include YouTube (73%
4
), Facebook (69%), Instagram
(33%), LinkedIn (27%), and Twitter (22%) (Perrin and Anderson, 2019). 25% of Twitter users
report using Twitter several times in a day, and 51% of Facebook users use Facebook at similar
frequency (Perrin and Anderson, 2019).
The social media population has grown steadily over the last decade, although the growth rate
has slowed down since 2018. For example, Facebook has 2.41 billion monthly active users and a
3
This cutoff score has good sensitivity and specificity and high internal consistency. (Lewinsohn et al., 1997).
4
Numbers in parenthesis are percentages of U.S. adults who say they have ever used the platforms.
9
yearly increase of 8% (Facebook, 2019). Twitter had 330 million monthly active users in 2019,
and monetizable daily active users increased by 21%. The increase mainly has come from product
improvement (Twitter, 2020).
The demographic composition varies with social media platform types. For example, YouTube,
LinkedIn, Twitter, and Reddit attract men more, and Facebook and Instagram attract women
more (Perrin and Anderson, 2019). The 25-29 uses social media the most. Social media usage
grows with education and income but does not change much between urban and suburban areas.
Twitter has 330 million global monthly users in 2020 and 63 million users are in the US.
Figure 1.1 shows the active user population grows over time. Among the US adults, 24% of men
use Twitter and 21% of women use it. User proportion decreases with age. About 44% of US
adults who are 18–24 year olds use Twitter; the percentage for 25–30 year olds is 31%, for 30–49
year olds is 26%; for 50–64 year olds is 17%; and 7% of 65+ year olds use Twitter. High income
people are more likely to use Twitter: 20% of those who make less than $75,000 use it, and 31%
of those who make more than $75,000 use it (Chen, 2020). Users send about 500 million Tweets
everyday. 80% of users use Twitter through mobile phone.
Facebook has 190 million users in the United States. The 25-30 age group has the highest
user proportion (84%) and the US adults who are over 65 years old are least likely to use Facebook
(46%). The proportion of the US adults who use Facebook for other age groups are: 76% for 18–24
age group, 79% for 30–49 age group, and 68% for 50–64 age group. User proportion does not vary
much with income: 69% of people in income group $0-$30,000 use Facebook, 72% of people in
income group $30,000-$75,000 use Facebook, and 74% of people with income over $75,000 use
it (Chen, 2020).
1.3 Data: Understanding America Study
The Understanding America Study (UAS) is an Internet probability-based panel of approxi-
mately 8,500 respondents that is representative of the entire US adult population.
5
The coverage
5
18.8% of the samples who were reached out by UAS in the first recruitment became panel members.
10
bias is manageable, because if the chosen households from the population do not have access to
Internet, the research team will provide them with the necessary equipment and Internet access.
UAS contains almost 300 surveys covering various topics. Data for this chapter were collected
by the UAS team in 2017, 2019, and 2020. I collaborated with the UAS team in designing the
2020 survey and collected a rich set of information about social media usage and people’s general
feelings about using social media.
Surveys that this chapter uses all have response rates over 75%. Table 1.1 lists response rate
and key variables from each survey wave. Response rates to answering social media usage in 2017
and 2019 were 77% and 76%, respectively. Response rates of surveys which asked depressive
symptoms were even higher. In 2017, 99.9% sample responded and 85.1% responded in 2019.
Last, 81% sample answered questions about social media usage and depressive symptoms in 2020.
The panel dataset contains 2198 observations from 2017 and 2019 data. In this sample, 8% re-
spondents belong to one household with at least one other respondent. The panel dataset with only
main member from each household contains 1994 observations. The individual level data with
detailed social media usage and depressive symptoms in 2020 contains 3081 observations.
Tables 1.2 and 1.3 summarize weighted mean and standard deviations of key control variables
in 2017, 2019, and 2020. There is no significant change in demographic composition (gender,
income, marital status, education, working status, and disability) over time, except age increases
as we would expect.
The average age of respondents in panel dataset (2017 and 2019 data) is 48.43 (standard de-
viation = 16.39). 58.37% of the sample are women and 41.63% are men. 74.43% of the sample
are white, 9.48% are African Americans, 3.72% are American Indian or Alaska Native, 5.04% are
Asian, 1.06% are Hawaiian/Pacific Islander, and the rest 6.27% are mixed. The medium income
group is $50,000-$60,000. 8.75% of the sample have annual income less than $10,000, and 22.69%
have annual income more than $100,000. 23.48% of the sample have no college level education,
and 15.85% have at least Master’s degree. 79.33% sample live in urban type areas, including city,
suburb, and town; 30.67% live in rural area. I use US Census Bureau’s method and divide US into
11
four regions: Northeast, Midwest, South, and West. 11.76% of my sample come from Northeast,
18.82% come from Midwest, 26.04% come from South, and the rest 43.39% come from West.
Subsequently, I use weights to correct non-response problem and make this sample match the US
adult demography.
1.3.1 Social Media Usage
I focus on one social media platforms, Twitter, in this chapter. Twitter is popular and a classic
example of social media. I also use information on other social media platforms’ usage when the
data are available. I do not include platforms such as YouTube and LinkedIn, because they have
specific features (video sharing and job searching) that go beyond traditional social media.
The 2017 data include whether people use Twitter. The 2019 data include whether people use
Twitter and Facebook, and how frequently people use them from “multiples times per day” to
“never in the last 12 months”. The 2020 data contain richer information than either 2017 or 2019
data, including (1) whether people use Facebook and Twitter; (2) which year did they start using
Facebook and Twitter; (3) how many connections do they have on Facebook and Twitter; (4) how
many followers do they have on Facebook and Twitter; (5) how many minutes do they spend on
social media in a day on average; (6) how many minutes do they spend on social media last night
before sleep; (7) how many minutes do they spend on social media during working hours on a
weekday on average.
Figure 1.2 presents social media usage in my sample. Usage decreases with age, except Face-
book usage, which increases slightly before 40. Men use Twitter more than women before 40 to
50 years old, but women use Facebook more than men in any age. Preference in each state varies.
The proportion of users in some states decreases over time, but the number of users can still grow
in these states.
12
1.3.2 The Level of Depressive Symptoms
CESD-8 module is available from UAS 2017 and 2019 data. It includes eight of the original
twenty items. Acclaimed studies, such as the US Health and Retirement Study (HRS) and the
English Longitudinal Study of Ageing (ELSA), both use this eight-item module. Table C.1 details
the eight items. Each item measures how often over the past week people experienced symptoms
(e.g., restless sleep and feeling lonely) associated with depression. Response options “yes” and
“no” have values 3 and 0 respectively (two items are reverse-scored). The final CESD score is the
sum of all items. A higher CESD-8 score indicates a greater level of depressive symptoms.
PHQ-4, another similar index of depressive symptoms, is available in the 2020 data. Three
descriptive measures of feelings about using social media are also available. Table C.1 details
these measures.
Figure 1.3 presents distribution of people with different levels of CESD in 2015, 2017 and
2019. About 40% people have no CESD depressive symptoms (CESD=0) and more than 60%
people have at most one CESD depressive symptom (CESD 3). Researchers used two CESD-8
cutoff scores, CESD-83 and CESD-84, to determine if the respondent is likely to be depressed
in the literature (Kozlov et al., 2020; Mhaolain et al., 2012). Based on the first criterion, 58.10%
of sample are likely to be depressed in 2015. The proportion decreased to 56.27% in 2017 and
increased to 60.72% in 2019. Based on the second criterion, 37.01% of sample are likely to be
depressed in 2015, 35.36% of sample are likely to be depressed in 2017, and 39.98% of sample are
likely to be depressed in 2019.
The large increase of depressive population proportion in 2020 was partially because of COVID-
19 pandemic. These numbers are also much larger than official reports of US depression preva-
lence, in which under 10% of the US adults suffer from depression. The discrepancy comes from
the different purposes of depression measures. CESD-8 is a self-reported measure, and it only
captures temporary levels of depressive symptoms. CESD-8 performs well to measure extensive
margin of depressive symptoms in research. However, for individual’s diagnosis, CESD-8 cannot
13
substitute clinical diagnose procedure of depression, which is more comprehensive and based on a
longer period of individual’s health history.
Figure 1.4 shows the level of depressive symptoms decreases with age and that women have
a higher level of depressive symptoms than men on average. Figure C.7 suggests the level of
depressive symptoms in many states decreased from 2017 to 2019.
1.3.3 Social Media Users and Non-users
Table 1.4 shows differences between Twitter users and non-users in individual’s characteristics.
Twitter users have a younger age, higher income level, higher education level, better working
status, and lower disability proportion than non-users.
6
The mean difference of CESD scores is
not significant among the groups. Figure 1.5 supports this claim.
1.4 Empirical Strategy
1.4.1 Neyman–Rubin Causal Model Framework
Under the Neyman–Rubin causal model framework, the causal effect of social media usage on
the level of depressive symptoms for individual i is
t
i
= Y
i(1)
Y
i(0)
(1.1)
in which Y
i(1)
is i’s level of depressive symptoms when she uses social media, and Y
i(0)
is the level
of depressive symptoms when she does not. We cannot observe Y
i(1)
and Y
i(0)
at the same time.
Instead, we observe Y
i
= T
i
(Y
i(1)
+(1T
i
)Y
i(0)
). We also observe binary variable T
i
which indicates
if i uses the studied social media platform or not.
7
Without a randomized controlled experiment, the
6
Facebook users have a similar pattern as Twitter users in age, income, education level, and working status. Face-
book users also have a higher female proportion, and higher disability proportion than non-users.
7
For the purpose of robustness, I also extend the definition of T
i
to be a vector of dummy variables indicating
social media usage frequency, and continuous variables indicating either usage time or number of connections on the
platform. The framework follows in these cases.
14
following model (1.2) estimates average treatment effect, ˆ t, if the unconfoundedness assumption
(1.3) is true.
Y
i
=tT
i
+bX
i
+e
i
(1.2)
in whiche
i
is i.i.d. error term and X
i
is a vector of covariates that satisfy the following unconfound-
edness assumption:
(Y
(1)
;Y
(0)
)?? Tj X (1.3)
This is the empirical framework used by existing papers that study the association between social
media usage T
i
and depressive symptoms Y
i
.
1.4.1.1 Challenges
The unconfoundedness assumption faces challenges in observational study. The social media
user population and non-user population may be different after we control for available individuals’
characteristics. Section 1.4.2 details a two-way fixed effects model to address this issue.
1.4.2 Two-Way Fixed Effects Model
I use the following two-way fixed effects model:
Y
it
=t
FE
T
it
+aX
it
+d
t
+g
i
+e
it
(1.4)
in which data consist of the triple (Y
it
;T
it
;X
it
), representing depressive symptoms, social media
usage, and other characteristics for individual i at time t. The parameter of interest is t
FE
which
estimates the average effect of Twitter on depressive symptoms between 2017 and 2019. The
variation to identify t
FE
comes from the samples who change social media usage over time. d
t
is
time fixed effect andg
i
is individual fixed effect. e
it
is error term.
This model accounts for time-invariant or cohort-invariant factors. Therefore t
FE
establishes
causal effect of social media usage T
it
on the level of depressive symptoms Y
it
if no omitted variable
15
presents which are both time variant and cohort variant. I include all available time- and cohort-
variant covariates in vector X
it
except those variables that can be mediators (pathways from social
media usage to the level of depressive symptoms) or colliders (outcomes of both social media
usage and the level of depressive symptoms). I use several robustness checks to further rule out
some confounders. To address reverse causality issues, I use placebo tests to rule out pre-existing
trends of Y
it
and effect from lagged Y
i(t1)
. However, I can not fully rule out the reverse causality
from current Y
it
.
To examine if the treatment effect is asymmetric, that is if the treatment effect magnitudejt
FE
j
when T
it
changes from 0 to 1 is not equal tojt
FE
j when T
it
changes from 1 to 0, I also use the
following first difference model:
DY
it
=t
1
D
1
+t
2
D
2
+aDX
it
+De
it
(1.5)
The omitted group includes people whose treatment levels do not change: T
it
T
i(t1)
= 0.
D
1
= 1 if T
it
T
i(t1)
= 1 and D
1
= 0 otherwise. D
2
= 1 if T
it
T
i(t1)
=1 and D
2
= 0 otherwise.
t
1
andt
2
represent the asymmetric treatment effects.
1.5 Empirical Results
1.5.1 Direct Evidence on US Population’s Feeling about Using Social Media
One way to study how people feel about using social media is asking respondents the question
directly. I collect such information in the 2020 survey. Figure 1.6 shows the distribution of people’s
feeling about using social media on a scale from 0 to 100.
8
Respondents evaluate three sets of
opposite feelings: social pressure versus social support, unhappy versus happy, and anxious versus
relaxed. The higher the score, the more positive feeling they have toward using social media.
More people respond with a score over 50 than with a score below 50 in all three cases. The mean
8
This figure also shows high frequency of focal point response 50. It is a combination of people whose answers
are exactly 50 and people who express they have no idea what to answer (Fischhoff and Bruine De Bruin, 1999).
16
of social support is 50.3 (standard deviation = 27); the mean of feeling happy is 58.0 (standard
deviation = 25); and the mean of feeling relaxed is 58.7 (standard deviation = 28). The medians
of the three scores are 50, 54, and 57, respectively. These results suggest that the population has
positive feelings about using social media on average.
Table 1.5 shows mean of individual’s characteristics among these groups: people feel social
media offers zero support, 45-55 support, or 100 support; people feel using social media makes
them zero happy, 45-55 happy, or 100 happy; people feel using social media is zero relaxed, 45-
55 relaxed, or 100 relaxed. Comparing zero score and 100 score groups, people who feel more
positive about using social media are younger, with lower incomes, and more likely to be currently
working or disabled. People who have less extreme opinions about using social media (score 45-55
group) are the youngest, with the highest income, education, proportion currently working, as well
as the lowest proportion that are disabled, among the three score groups.
1.5.2 Social Media Usage Decreases the Level of Depressive Symptoms
Table 1.6 shows estimations from pooled-OLS, two-way fixed effects with no covariates, and
two-way fixed effects with covariates models. Using Twitter decreases CESD score by 1.150
points (standard error = 0.46), which is a 26.64% decrease compared to the average CESD score
(4.35) of Twitter non-users.
9
This result suggests using Twitter can decrease the level of depressive
symptoms when keeping other factors unchanged.
10
Column (1) controls for covariates which include age, income, gender, marital status, educa-
tion, and state; column (3) controls for income, working and marital status because these factors
can potentially change for the respondents over time. Standard errors are clustered to address
heteroskedasticity. I also use weights so the sample can represent the US population.
9
To better understand the magnitude of the effect, job loss due to a worker’s unexpected firm closure increased
CESD in the US by 28.2% (Riumallo-Herl et al., 2014); a phone and Internet-based behavioral intervention for over-
weight women reduces CESD-10 by 23% (Kerr et al., 2008).
10
A robustness check using quantile regressions shows no significant change in CESD spread.
17
Column (1) shows the simple correlation between Twitter usage and depressive symptoms is
positive, which is consistent with the literature.
11
Columns (2) and (3) shows the effect of Twitter
on depressive symptoms is negative. Comparing columns (2) and (3), they show the selection bias
on observables is limited because the estimation does not change much. Therefore, the remaining
selection bias on unobservables may also be limited. I also use this information to statistically
estimate potential bias from unobservables (Altonji et al., 2005; Oster, 2019) in section 1.5.3.2.
1.5.2.1 Effects on Each of Eight CESD Item
Table 1.7 decomposes outcome variable to eight CESD items so I can explore how each aspect
of depressive symptoms change after using Twitter. Twitter usage decreases all eight CESD item
scores. In particular the effects on third, fourth, and eighth items are significantly different from
zero (-0.27, -0.18, and -0.22, respectively), which suggests restless sleep (CESD3), unhappiness
(CESD4), and the feeling of not being able to get going (CESD8) are alleviated by Twitter usage
by 28% (0.27/0.93), 36% (0.18/0.51), and 37% (0.22/0.5), respectively.
12
1.5.3 Test Fixed Effect Assumptions
1.5.3.1 Potential Confounders
Table C.2 shows estimation of social media usage’s correlation with CESD score in 2017 and
2019, and with PHQ score in 2020. The results suggest such correlation of Twitter with the level
of depressive symptoms can be close to zero or positive, which is consistent with the literature.
The positive association in existing studies and negative effect in this chapter can be explained
by confounders. The two-way fixed effects model in this chapter takes time-invariate and cohort-
invariate factors into account, and gets rid of a large part of bias induced by confounders.
The remaining confounders are those factors which change for individuals over time and affect
both social media usage and the level of depressive symptoms. For example, marital status can
11
The magnitude of this correlation is not directly comparable to other studies, since other studies either did not use
CESD as measure of depressive symptoms or the target populations are different.
12
However, I cannot reject equality of different items.
18
change between 2017 and 2019, and may affect both social media usage and mental health. Sim-
ilarly, income and employment status are also potential confounders. Therefore, I include these
individual characteristics as control variables in the two-way fixed effects model estimations.
Other potential confounders include big life events and social media attribute preference. It is
hard to collect and quantify these factors. But the bias is limited if we assume positive and negative
effects from these confounders are balanced in the sample. Using other social media platforms can
also confound the estimation. Solving this problem requires more data. However, estimation using
2020 data in section 1.5.4 serves as a starting point to explore this confounder.
1.5.3.2 Bias Estimation With Altonji-Oster Model
Following the literature (Altonji et al., 2005; Oster, 2019), I use estimated parameters and R
2
from a few regression models to estimate the required minimum bias from uncontrolled factors in
order for the true impact of Twitter to be zero. We use the following three regression models. The
first one is the uncontrolled model:
Y
it
=t
0
T
it
(1.6)
I denote the R
2
from model (1.6) as R
0
.
The second model is the two-way FE model (1.4). I denote the R
2
from (1.4) as R
FE
.
The last model is the full model in which we explicitly include the potential unobservable
confounders W:
Y
it
=tT
it
+aX
it
+d
t
+g
i
+W+e
it
(1.7)
I denote the R
2
from model (1.7) as R
max
.
If the true impact from Twitter t = 0, I can estimate the ratio of impact from unobservables
versus impact from controlled variablesd with the following equation (Oster, 2019):
d =
t
FE
(R
FE
R
0
)s
2
y
t
X
+t
FE
s
2
X
t
X
(t
0
t
FE
)
2
+ 2t
2
FE
(t
X
(t
0
t
FE
)s
2
X
)+t
3
FE
t
X
(s
2
X
t
X
)
(R
max
R
FE
)s
2
y
(t
0
t
FE
)s
2
X
+t
FE
(R
max
R
FE
)s
2
y
(s
2
X
t
X
)+t
2
FE
(t
X
(t
0
t
FE
)s
2
X
)+t
3
FE
t
X
(s
2
X
t
X
)
(1.8)
19
I assume R
max
= 1:3R
FE
following the literature (Oster, 2019) and get the other values from the
regression outputs from (1.4) and (1.6). According to my estimation, remaining unobserved factors
need to have at least the same amount of effect as controlled factors for the real Twitter effectt to
be 0 (
ˆ
d = 0:97), which is unlikely to be true in this case. Therefore, I am confident that the impact
of Twitter on depressive symptoms is negative.
1.5.3.3 Test CESD Trends of Switchers
Reverse causality is a potential problem. We may capture an effect that is the combination of
social media usage’s effect on depressive symptoms and depressive symptoms’ effects on social
media usage. This chapter addresses this problem in two ways. First, I test there is no pre-existing
trend of CESD for different social media users in this section. Second, I explore two types of
instruments in Appendix A and B.
13
One case can potentially violate the fixed effects model assumptions. That is when people who
switch from not using Twitter in 2017 to using Twitter in 2019 may be people that are already on a
negative trend in terms of depressive symptoms. To rule out this case, I test if there are pre-existing
trend for this group of users. To be more specific, I define two types of switchers: Switcher group
1 includes people who did not use Twitter in 2017 but switch to being users in 2019; Switcher
group 2 includes people who use Twitter in 2017 but switch to not using Twitter in 2019; the rest
of sample belong to the third group who did not change user status. The two Switcher groups help
in identifying the fixed effect estimation, and the identification assumptions would be violated if
we see a negative pre-trend for Switcher group 1 and a positive pre-trend for Switcher group 2.
Figure 1.7 shows the CESD trends for the two switcher groups and the reference group. The
pre-trends are three points on the left. There is no pre-trend for Switcher group 1 and non-
switchers. The Switcher group 2 has a negative pre-trend which is the opposite of violation. The
t-tests in the table below also shows there is no pre-trends that violate model assumptions. Ta-
ble 1.8 is another evidence to support the validity of fixed effect model assumptions. The change
13
The first type IV uses information on social media application’s frequent updates, and the second IV uses national
social media usage growth rate as IV .
20
of CESD between time periods (t-1) and t does not correlate with change of Twitter usage between
time periods t and (t+1). These evidence supports that fixed effect model pick up the effect of
Twitter instead of pre-existing trend.
1.5.3.4 Test Potential Mediators
Column (2) and (3) from Table 1.6 details two-way fixed effects estimations without and with
time-varying covariates. Both estimated Twitter effects are negative. Without covariates, the mag-
nitude of effect is slightly larger, which suggests omitting these factors would bias estimation
upwards. However, the difference is small, which suggests the selection bias from observable
variables is limited.
I further examine if these time-varying factors are mediators between social media usage and
depressive symptoms. Table 1.10 shows that Twitter usage does not correlate with either of house-
hold income level and marital status. This result shows that these time-varying factors are not
mediators and should be included in the two-way fixed effects model.
1.5.4 Robustness Checks
1.5.4.1 SUTV A
Social media has the nature of enhancing network effect, so I test whether data violate Stable
Unit Treatment Value Assumption (SUTV A) explicitly by keeping only one individual from the
same household.
SUTV A states that one unit’s treatment status should not affect other units’ treatment status. In
this chapter, SUTV A is not valid if one respondent’s social media usage affect other respondents’
social media usage. The violation happens if some respondents have connections through social
media or in offline life. Data on which respondents have connections on social media are not
available. But it is unlikely to threaten the identification since the sample was randomly chosen by
the team from all over the United States, so I expect the case that respondents have connections
21
on social media happens rarely. Similarly data are not available on whether respondents know
each other in real life (except for family members), but I expect these cases are also rare. Some
respondents do come from the same family. In this case, one member’s social media usage can
potentially affect other family members’ usage.
Table 1.11 address this problem by keeping only main respondents from each family (the mem-
ber who has the same survey ID number as the household ID number). The households with more
than one members takes 7.93% of my original sample. The fixed effects estimation does not change
qualitatively. The two-way fixed effect model estimates Twitter’s impact to be -1.175 (standard er-
ror = 0.51), and equals to 0.19 standard deviation. The magnitude of impact increases 0.025 (from
1.150 to 1.175), and the percentage impact increases from 26.64% (1.150/4.349 = 26.64%) to
27.24% (1.175/4.313 = 27.24%).
Table 1.12 shows a comprehensive set of estimation results from pooled-OLS, individual fixed
effect, and two-way fixed effects models. The estimated impact of Twitter on CESD is stable
across different specifications.
1.5.4.2 First Difference Estimation on Switchers
I use equation (1.5), the first difference model, to examine if there is asymmetric impacts on
Switcher group 1 (who increased Twitter usage between 2017 and 2019) and Switcher group 2
(who decreased Twitter usage between 2017 and 2019). Table 1.13 presents the estimation of ˆ t
1
and ˆ t
2
in equation (1.5). Outcome variable in Column (1) is CESD score change, and outcome
variable in Column (2) is a dummy variable indicating if CESD score increased between 2017 and
2019. Column (1) shows effect on CESD extensive margin is symmetric and column (2) shows ef-
fect on CESD intensive margin is symmetric. F-test in both specifications reject effect asymmetry.
Table 1.13 also shows using Twitter decreases the probability that depressive symptoms increasing
by 11.4%.
14
14
The correlation between Twitter switching behavior and Facebook switching behavior is 0.08.
22
1.5.4.3 Outliers
CESD scores range from 0 to 24, but more than 90% of sample has CESD scores under 15,
and more than 95% of sample has CESD scores under 21 (Figure 1.3). Outliers affect estimation
accuracy if the data bear measurement error. Only a few respondents have CESD scores higher
than 21, but I do not have evidence that high CESD scores in this chapter have measurement errors.
So the main specifications include the full sample. For the purpose of robustness, Table 1.14 shows
results when I winsorize CESD score at 6, 9, 12, 15, 18, and 21 (I replace CESD values larger than
bound x with x). The estimations with winsorization are all negative and significant, which is
consistent with the main result. This finding suggests that the main result is robust to potential
outliers.
15
1.5.4.4 Attrition
524 out of 2722 respondents in 2017 do not respond in 2019. The attrition rate is 19.25%. Ta-
ble 1.15 shows the characteristics difference in 2017 for the attrited sample and panel data sample.
The Twitter usage and level of depressive symptoms are not different between attrition sample and
panel data sample. Some characteristics are different between the two groups. As a robustness
check, I use inverse probability weighting and assign weights to samples so the panel data sample
is not different from original 2017 samples in observables. The weights p
i
are based on regressing
being in both waves on observables using logistic regression: ln
p
i
1p
i
= f(X
i;2007
). X
i;2007
includes
both time-varying and time-invariant observables such as age, gender, income level, marital status,
education, and empolyment status. For each sample, the final weight is(bw
i

1
p
i
) where bw
i
is the
baseline weight so that original sample is representative of the US population. The two-way fixed
effects estimation does not change qualitatively after applying inverse probability weighting.
15
Table C.3 also shows results from samples with CESD scores under 21, 18, 15, and 12.The negative effect shrinks
as high CESD score respondents are removed, but the implication stays the same that social media usage decreases
the level of depressive symptoms.
23
1.5.4.5 Suggestive Evidences on “Good Social Media”
Facebook usage information is available for 2019 and 2020, but outcome CESD is available for
2017 and 2019. Therefore the fixed effect model is not an option to study Facebook usage’s impact.
Table C.2 shows the linear regression estimation for Facebook’s impact on 2019 CESD score and
2020 PHQ score. Although the coefficients are not significant, the implication is consistent with
other results that more Facebook usage is associated with lower levels of depressive symptoms.
Table C.4 estimates the correlation between different Twitter and Facebook use frequencies
and CESD score in 2019. The coefficients represent each frequency group’s difference from the
omitted “never” frequency group. The result suggests that using Twitter a few times per day is
associated with lower levels of depressive symptoms compared to never using Twitter.
1.5.5 Heterogeneous Effect
The social media usage and the level of depressive symptoms vary among people with different
characteristics. The social media usage’s impact can also vary for different people. Figure 1.8 to
Figure 1.14 explore the possible heterogeneity.
Figure 1.8 shows that the negative impact of social media on the level of depressive symptoms
mainly comes from younger adults (age < 40). In this figure, I split sample into nine age groups.
Depressive symptoms of people with age 17 to 40 is mitigated by social media usage. Social media
may increase level of depressive symptoms for people with age over 90. The rest age groups are
not affected by social media usage much. There is also no heterogeneity with respect to gender.
Combining with Figures 1.2 and 1.4, the younger cohorts use more social media, have higher level
of depressive symptoms, and using more social media helps them mitigate depressive symptoms.
Figure 1.9 shows treatment heterogeneity in education. Based on highest education of samples,
I divide them into three groups: less than high school, equal or equivalent to high school, and
equivalent or more than college. The effect of Twitter is significantly negative for the last two
groups, which represent people with equal or more than high school education.
24
Figure 1.10 shows treatment heterogeneity in income groups. I divide the sample into 0 to 25
percentile group, 25 to 50 percentile group, 50 to 75 percentile group, and 75 to 100 percentile
group based on annual household income. Twitter effect focuses on below 25 percentile income
group, which corresponding to annual household income less than $30,000. Heterogeneity analysis
with more income categories shows Twitter mainly decreases depressive symptoms for people with
household annual income from $20,000 to $30,000.
Figure 1.11 shows heterogeneous effect by prior level of depressive symptoms. I split the
sample based on CESD in 2015. The Twitter usage decreases the level of symptoms most for those
groups whose 2015 CESD scores were larger than 3.
Figure 1.12 shows heterogeneous effect from Twitter by race groups. White, native American,
and Hispanic/Latino groups benefit from using Twitter the most. Black, Asian, and mixed groups
show no significant effect.
Figure 1.13 shows heterogeneity with respect to urbanicity. I split sample to four groups: large
city, small city, town or suburb, and rural. People from small city benefit from using Twitter the
most. The potential mechanism can be related to social interaction style, income, and occupation
types. Other groups show no significant effect.
Figure 1.14 shows little hetergeneity among different US regions. I follow US Census Bureau’s
criteria and split sample into four region groups: Northeast, Midwest, South, and West. 11.76%
sample come from Northeast, 18.82% sample come from Midwest, 26.04% sample come from
South, and 43.39% sample come from West. The Northeast people benefit from using Twitter the
most.
Other heterogeneity analysis results are in appendix. Figure C.2 shows little heterogeneity for
different marital status. Figure C.4 shows heterogeneous effect by working hours. Twitter has
a larger mitigation effect on the level of depressive symptoms for people who work more. Only
the 40 hours of work per week group has significant estimation, because it is the most common
working hours. Figure C.5 shows people with more than five children can benefit the most from
using Twitter.
25
1.5.6 Labor Market Benefits
This chapter estimates that using social media decreases the level of depressive symptoms by
around 27%. I assume that workers with depression offer 4.1 productive work hours less than
workers without depression every week (Stewart et al., 2003).
16
So social media can increase
worker’s productive work hours by improving worker’s depressive symptoms.
On the other hand, spending time on social media during working hours can decrease labor
outcome. The average time respondents spend on social media during working hours on work
days is 31.61 minutes. Figure 1.15 shows the distribution of time spent on social media. On
average each individual provide 1.47 hours more in labor market each week because some people’s
depressive symptoms are alleviated by using Twitter.
17
With an average hourly income of $35, the
monetary benefit for each person is around $51.5 per week.
In Table 1.19 I estimate the proportion of people who change from having depression to not
having depression because of using Twitter. Column (1) is based on depression cutoff CESD=3,
and Twitter alleviates about 15.4% people’s depression based on this criterion. Column (2) uses
an alternative cutoff in the literature CESD=4, and using Twitter alleviates about 6.9% people’s
depression.
18
I will show labor benefit estimation based on both criteria, but I expect the estimation
based on cutoff CESD=4 is more realistic. I estimate the total labor market benefit per year in the
following way:
Population benefit=t
d;g
US adult working population
g
saved hours
g
wage
g
(1.9)
g indicates groups by gender and age. t
d;g
is the estimated proportion of people whose depression
is alleviated by Twitter effect in each group g.
I estimate the total US labor market benefit by adding up benefits from all groups. This method
addresses heterogeneity in labor benefit. I estimate labor market benefits for male and female
16
Their definition of depression bases on two depression-screening questions.
17
4.1 hours-31.61 min/60 min*5 work days per week=1.47 hours.
18
Phone and Internet-based behavioral intervention for overweight women reduces proportion of sample with
CESD-10 over 10 by 5.7% (Kerr et al., 2008).
26
workers separately in three age groups: 20-24, 25-55, and over 55. Table 1.20 show the elements
to compute labor market benefit for each of these groups. The working population of each gender
and age groups come from US Bureau of Labor Statistics. Thet
d
is estimation of Twitter’s impact
on depression alleviation for different groups. Average number of minutes on social media per
workday and average hourly wage are also estimation using my data. Based on cutoff CESD = 4,
the proportion of people with depression decreases by about 10% in each group, except for men
older than 55 the proportion increases by 3%. Number of minutes on social media varies across
groups.
Table 1.20 and Figure 1.16 also show the saved productive work hours and labor market benefit
per week for these gender and age groups. Men save 2 more productive work hours from using
Twitter, and women save 1 more hour. The corresponding labor market benefit for men is 290
million to 1.34 billion dollars per week, and for women is about 150 million dollars per week.
Figure 1.16 also shows large heterogeneity in labor market outcome for different age groups. For
worker from 20 to 24 years old, they provide about one productive work hour less due to using
Twitter, resulting in 27 to 47 million dollar loss per week. This group spend too much time on
social media during working hours, although their depressive symptoms are also alleviated. In
the end, this age group provides less productive working hours. Worker from 25 to 54 years old
provide about one productive work hour more, which leads to about 350 to 550 million dollar labor
market benefit per week. Workers over 55 provide about two productive work hours more, which
leads to about 150 to 300 million dollar labor market benefit per week. Since some people with age
over 55 do not work and average wage is lower in this group than 25 to 54 age group, the monetary
benefit from over 55 age group is also smaller. I also split samples according to both gender and
age to six groups. Women from age 20 to 24 have negative saved productive working hours and all
the other groups have positive saved productive working hours. The labor market benefits are: 0
for men with age 20 to 24; -$25 million for women with age 20 to 24; $358 million for men with
age 25 to 54; $166 million for women with age 25 to 54; -$52 million for men with age over 55;
$110 million for women with age over 55.
27
The total labor market benefit estimation from adding up the six groups is 505 million dollars
per week or 0.12% US weekly GDP.
19
1.5.7 Social Media Usage and Depressive Symptoms in 2020
Figure 1.17 shows the distribution of number of Twitter connections and followers, and the dis-
tribution of number of Facebook connections and followers. The distribution is skewed with most
users have less than 100 connections or followers. Users have on average 174 Twitter connections,
552 Twitter followers, 469 Facebook connections, and 230 Facebook followers.
The median number of Twitter connections is 10, median number of Twitter followers is 5,
median number of Facebook connections is 200, and median number of Facebook followers is 35.
Table 1.16 shows linear regression results between social media usage and level of depressive
symptoms (four-item PHQ score).
20
Twitter usage and PHQ has positive correlation, but number
of Facebook followers significantly reduces level of depressive symptoms. Not surprisingly, social
media users bear more positive attitudes towards social media. Table 1.17 shows which exten-
sive margin bins of social media usage have effect on depressive symptoms and people’s attitudes.
Using Twitter for 300 to 400 minutes per day decreases PHQ by 2.2 points conditional on being
Twitter users. Using Facebook for 200 to 300 minutes per day decreases users positive feelings
about social media conditional on being Facebook users. This group of users feel less social sup-
port and less relaxed from using Facebook. Having 200 to 400 connection on Facebook increases
the chances users feel happy by using Facebook.
One limitation is 2020 data collection coincides with COVID pandemic. We should interpret
results with that fact in mind. We consider both social media and traditional media’s effect on PHQ
in another paper (Riehm et al., 2020).
19
Adding up gender groups suggests total benefit to be 440 million dollars per week (0.11% US weekly GDP).
Adding up age groups suggests total benefit to be 473 million dollars per week (0.12% US weekly GDP). US GDP
per week in 2018: 395 billion dollars.
20
CESD score in 2020 is not available yet.
28
1.6 Discussion
The findings of this chapter are consistent with intuition and economic theory. The following
mechanisms can explain why using social media can potentially decrease users’ levels of depres-
sive symptoms. First, social media companies are among the biggest companies worldwide and are
continually growing. It is not surprising that their business strategies include making users happy
as well as making users addicted. Figure 1.6 contains strong evidence of this argument.
Second, the content of social media changes over time. More social media platforms appear,
and more features become available. The Twitter of 2017 is not exactly the same as the Twitter
of 2019. It is possible that social media takes up a more positive role in society. New features of
social media platforms increase people’s happiness, such as customized news feed that utilize the
fast development of machine learning techniques.
Last, social media is still young. Supervising and regulating protocols in the social media
industry are in the process of becoming more sophisticated and balanced. For example, more
socially responsible social media activities and less social media abuse happen over time. Social
media may have more and more positive impact on people’s mental health over time. Overtime,
social media users are also learning the best way to utilize social media. Therefore, it is possible
that more studies find similar evidence of beneficial social media usage as in this chapter in the
future.
Some existing studies use randomized controlled trials to study the effect. This method requires
fewer assumptions and offers better estimation, but these experiments suffer from a small sample
size problem, and the samples are usually limited to a specific sub-population (such as college
students). This chapter shows heterogeneity exists in many cases. Therefore, the findings can be
different for these laboratory experiments and for this chapter when both results are valid.
Similarly, these laboratory experiments study the effect of social media usage in a shorter time
period than that discussed in this chapter. These experiments usually last about one or two weeks.
This chapter uses social media usage information and level of depressive symptoms data that are
29
around three months apart. The different time frame can be one of the reasons why the findings
are not the same.
The types of bias are different in laboratory experiment studies and this chapter. In this chapter,
two-way fixed effects model may not be able to capture all potential confounders. On the other
hand, experimental studies require participants to reduce social media usage and ask their feelings
later. Participants know the purpose of the experiment, and their answers may be biased because
of this knowledge, the presence of an interviewer or the social norm.
1.7 Conclusion
This chapter finds that social media usage decreases the level of depressive symptoms em-
ploying a two-way fixed effects model. Direct information on people’s feelings about using social
media also supports that using social media can be beneficial. This result implicates that society
may have underestimated the importance and contribution of social media. Social media platforms
may have adopted more important roles with new available features and better informed users, as
well as more responsible roles with improved regulation in the industry.
The data have limitations. There is no information on individual’s Facebook usage in 2017,
and no consistent measure of the level of depressive symptoms (CESD) in 2020 yet. Therefore, I
apply the two-way fixed effects model to a limited set of data. With richer data on social media
usage and people’s opinions about social media in the future, more accurate estimation is possible.
These limitations can be alleviated in the future when richer data become available.
The project described in this chapter relies on data from survey(s) administered by the
Understanding America Study, which is maintained by the Center for Economic and Social
Research (CESR) at the University of Southern California. The content of this chapter is
solely the responsibility of the author and does not necessarily represent the official views of
USC or UAS.
The author participated in designing 2020 survey.
30
Figure 1.1: Twitter User Growth (2010-2019)
Source: https://www.statista.com/chart/10460/twitter-user-growth/
31
0
.2
.4
.6
.8
1
Twitter
20 40 60 80
Age
male (2017) female (2017)
0
.2
.4
.6
.8
1
Twitter
20 40 60 80
Age
male (2019) female (2019)
0
.2
.4
.6
.8
Twitter
20 40 60 80
Age
male (2020) female (2020)
(a) (b) (c)
0
.2
.4
.6
.8
1
Facebook
20 40 60 80
Age
male (2019) female (2019)
.2
.4
.6
.8
1
Facebook
20 40 60 80
Age
male (2020) female (2020)
(d) (e)
Figure 1.2: Twitter and Facebook User Percentages by Gender and Age Group
This figure shows social media usage (Twitter for (a) - (c), Facebook for (d) - (e)) by gender and age. Usage decreases with age, except
Facebook usage increases slightly before 40. Men use Twitter more than women before 40 to 50 years old, but women use Facebook
more than men at any age.
32
41.9
21.09
14.08
8.525
4.431
2.636
3.421
1.739
2.187
0 10 20 30 40 50
Percent
0 3 6 9 12 15 18 21 24
CESD (2015)
(a)
43.73
20.91
13.4
5.894
4.534
3.426 3.375
2.317 2.418
0 10 20 30 40 50
Percent
0 3 6 9 12 15 18 21 24
CESD (2017)
(b)
39.28
20.74
13.03
7.814
5.116
3.644
4.31
3.434
2.628
0 10 20 30 40 50
Percent
0 3 6 9 12 15 18 21 24
CESD (2019)
(c)
Figure 1.3: CESD Distribution in 2015, 2017, and 2019
CESD-8 takes discrete values of 0, 3, 6, 9, 12, 15, 18, 21, and 24. The higher the CESD score, the
higher the level of depressive symptoms. Around 40% respondents have CESD=0.
33
0
5
10
15
CESD
20 40 60 80
Age
male (2017) female (2017)
0
.2
.4
.6
.8
1
CESD
20 40 60 80
Age
male (2019) female (2019)
0
2
4
6
8
PHQ
20 40 60 80
Age
male (2020) female (2020)
(a) (b) (c)
Figure 1.4: The levels of Depressive Symptoms by Gender and Age Group
This figure shows the level of depressive symptoms decreases with age. Women have higher level of depressive symptoms than men on
average.
34
0 5 10 15
CESD
20 40 60 80
Age
Twitter users (2017) Twitter non-users (2017)
0 1 2 3 4
CESD
20 40 60 80
Age
Twitter users (2019) Twitter non-users (2019)
0 2 4 6 8
PHQ
20 40 60 80
Age
Twitter users (2020) Twitter non-users (2020)
(a) (b) (c)
0 5 10 15
PHQ
0 100 200 300 400
Minutes/day
0 2 4 6 8
CESD
20 40 60 80
Age
Facebook users (2019) Facebook non-users (2019)
0 2 4 6 8
PHQ
20 40 60 80
Age
Facebook users (2020) Facebook non-users (2020)
(d) (e) (f)
Figure 1.5: The Levels of Depressive Symptoms by Age: Users versus Non-users
This figure shows the level of depressive symptoms by social media usage(Twitter for (a) - (d), Facebook for (e) - (f)). Users not always
have higher/lower levels of depressive symptoms than non-users. PHQ increases slightly with time spent on social media per day.
35
0 5 10 15 20 25
Percent
0 20 40 60 80 100
Social pressure (0) versus social support (100)
(a)
0 5 10 15 20 25
Percent
0 20 40 60 80 100
Unhappy (0) versus happy (100)
(b)
0 5 10 15 20
Percent
0 20 40 60 80 100
Anxious (0) versus relaxed (100)
(c)
Figure 1.6: How People Feel About Using Social Media
This figure shows the distribution of people’s feeling about using social media on a scale from 0
to 100. Respondents evaluate three sets of opposite feelings: social pressure versus social support,
unhappy versus happy, and anxious versus relaxed. The higher the score, the more positive feeling
they have towards using social media. The mean of social support is 50.338 (s.d. = 26.993); the
mean of feeling happy is 57.960 (s.d. = 24.587); and the mean of feeling relaxed is 58.720 (s.d.
= 27.501). The median of three scores are 50, 54, and 57 respectively. This result suggests the
population have positive feelings about using social media on average.This figure also shows high
frequency of focal point response 50. It is a combination of people whose answers are exactly 50
and people who express they have no idea what to answer (Fischhoff and Bruine De Bruin, 1999).
36
-4 -2 0 2 4
CESD change
2017-2015 2019-2017
Years
Twitter (1->0)
Twitter (unchanged)
Twitter (0->1)
Difference and T-test
2017-2015 2019-2017
Twitter(0!1)-Twitter(unchanged) -0.44 -1.30*
Twitter(unchanged)-Twitter(1!0) 1.76*** -1.06
Twitter(0!1)-Twitter(1!0) 1.32 -2.36**
Figure 1.7: CESD Trends
This figure shows the CESD trends for the two switcher groups and the reference group. The
pre-trends are three points on the left (2017-2015) and a visual presentation of Twitter effect es-
timation is from three points on the right (2019-2017). There is no pre-trend for Switcher group
1 and non-switchers. The Switcher group 2 has a negative pre-trend which is the opposite of as-
sumption violation. The t-tests in the table below also shows there is no pre-trends that violate
model assumptions.
37
-20 -15 -10 -5 0 5
Twitter effect
10 20 30 40 50 60 70 80 90
Age
0 Female 1 Male

Figure 1.8: Heterogeneous Effect: Age and Gender
The figure shows point estimations and 95% confidence intervals. There is little heterogeneity for different genders. Negative effect
mainly comes from young cohorts less than 40 years old.
38
-6 -4 -2 0 2
Twitter effect
< High school High school ≥ College
Education
Figure 1.9: Heterogeneous Effect: Education
The figure shows point estimations and 95% confidence intervals. The Twitter effect is significantly negative for samples with highest
education equal or more than high school.
39
-4 -3 -2 -1 0 1
Twitter effect
0-25 25-50 50-75 75-100
Income (percentile)

Figure 1.10: Heterogeneous Effect: Income
25% percentile = $30,000; median= $60,000; 75% percentile = $100,000. The figure shows point estimations and 95% confidence
intervals. All point estimations are negative. The effect is significantly negative for samples with annual household income below
$30,000.
40
-4 -3 -2 -1 0
Twitter effect
CESD2015<4 CESD2015≥4
CESD 2015
Figure 1.11: Heterogeneous Effect: Past Level of Depressive Symptoms
The figure shows point estimations and 95% confidence intervals. The Twitter usage decreases the level of symptoms for groups with
2015 CESD larger than 3.
41
-10 -5 0 5 10
Twitter effect
White African AmericanNative American Asian Hispanic/Latino Mixed
Race
Figure 1.12: Heterogeneous Effect: Race
The figure shows point estimations and 95% confidence intervals. Social media decreases depressive symptoms for these groups: white,
native American, Hispanic/Latino.
42
-6 -4 -2 0 2
Twitter effect
Large city Small city Town/suburb Rural
Urbanicity
Figure 1.13: Heterogeneous Effect: Urbanicity
The figure shows point estimations and 95% confidence intervals. Social media’s negative effect focuses on small city.
43
-6 -4 -2 0 2
Twitter effect
Northeast Midwest South West
Region

Figure 1.14: Heterogeneous Effect: Region
The figure shows point estimations and 95% confidence intervals. Twitter significantly decreases level of depressive in Northeast region.
Northeast: New England, Connecticut, Maine, Massachusetts, New Hampshire, Rhode Island, Vermont, Mid-Atlantic, New Jersey, New
York, and Pennsylvania.
Midwest: Illinois, Indiana, Michigan, Ohio, Wisconsin, Iowa, Kansas, Minnesota, Missouri, Nebraska, North Dakota, and South Dakota.
South: Delaware, Florida, Georgia, Maryland, North Carolina, South Carolina, Virginia, District of Columbia, West Virginia, Alabama,
Kentucky, Mississippi, Tennessee, Arkansas, Louisiana, Oklahoma, and Texas.
West: Arizona, Colorado, Idaho, Montana, Nevada, New Mexico, Utah, Wyoming, Alaska, California, Hawaii, Oregon, and Washing-
ton.
44
0 5 10 15 20 25
Percent
0 100 200 300 400
Number of minutes on social media per day
0 10 20 30 40
Percent
0 50 100 150
Number of minutes on social media last night before sleep
(a) (b)
0 10 20 30 40 50
Percent
0 100 200 300
Number of minutes on social media during working hours on workdays
0 10 20 30 40 50
Percent
0 20 40 60 80 100
Less sleep (0) versus more sleep (100)
(c) (d)
Figure 1.15: Time Spent on Social Media
These histograms show the distribution of time spent on social media per day, last night before
sleep, during working hours on workdays, and whether using social media affect sleep time.
45
2.023
1.043
1.341
.157
.289
.155
0 .5 1 1.5
Benefit (billion)/week
0 1 2
Saved LPT(hour)/week
Men Women
Gender
saved LPT(hour)/wk benefit(CESD cutoff=3)
benefit(CESD cutoff=4)
-.984
.938
2.318
-.047
.546
.3
-.027
.351
.154
-.2 0 .2 .4 .6
Benefit (billion)/week
-1 0 1 2 3
Saved LPT(hour)/week
20-24 20-24 ≥55
Age category
saved LPT(hour)/wk benefit(CESD cutoff=3)
benefit(CESD cutoff=4)
1.513
-1.718
1.302
.7
2.768
1.89
0
-.029
.624
.151
.334
.045
0
-.017
.358
.106
-.052
.11
-.4 -.2 0 .2 .4 .6
Benefit (billion)/week
-2 -1 0 1 2 3
Saved LPT(hour)/week
20-24 men 20-24 women 25-54 men 25-54 women ≥55 men ≥55 women
Age-gender Group
saved LPT(hour)/wk benefit(CESD cutoff=3)
benefit(CESD cutoff=4)
Figure 1.16: Labor Market Benefit
These figures show the estimated saved productive work hours from Twitter effect in different groups (by gender and age). The total US
labor market benefit from this Twitter effect is the sum of all groups’ benefits and it is approximately about 0.1% of US weekly GDP.
46
0 10 20 30 40 50
Percent
0 100 200 300 400
Twitter connections
0 20 40 60
Percent
0 100 200 300 400 500
Twitter followers
(a) (b)
0 5 10 15 20
Percent
0 200 400 600 800 1000
Facebook connections (2020)
0 10 20 30 40 50
Percent
0 200 400 600 800 1000
Facebook followers
(c) (d)
Figure 1.17: Numbers of Connections and Followers on Twitter and Facebook
This figure shows distribution of Twitter connections/followers and Facebook connections/followers. All sample with large numbers are
winsorized. The distribution is skewed with more people having small numbers of connection/followers.
47
Table 1.1: UAS Survey Information
Surveys Field Dates Observations Response Rates Variables
UAS88 February 24, 2017 to April 10, 2017 4786 77.29% tw
UAS95 June 01, 2017 - 4636 99.85% CESD
UAS180 April 15, 2019 to May 16, 2019 5464 76.43% tw, fb, freq tw, freq fb
UAS185 June 07, 2019 - 5412 85.09% CESD
UAS230 March 10, 2020 to March 31, 2020 7145 81.06%
tw, fb, min, min last night,
min working hours, friends tw, friends fb,
followers tw, follwers fb, PHQ, support,
happy, relaxed, less sleep
This table summarize five Understanding America Study (UAS) surveys which provide data in this paper. UAS88 and UAS 95 provide
2017 data; UAS180 and UAS185 provide 2019 data; UAS230 provides 2020 data.
“tw” = using Twitter or not; “fb” = using Facebook or not; “CESD” = eight CESD items; “freq tw” = Twitter using frequency; “freq fb”
= Facebook using frequency; “min” = how many minutes do you spend on social media in a day on average; “min last night” = how
many minutes did you spend on social media last night before you went to sleep; “min working hours” = how many minutes do you
spend on social media during working hours on a weekday on average; “friends tw” = number of friends or connections on your Twitter
account; “friends fb” = number of friends or connections on your Facebook account; “followers tw” = number of people who follow
you on your Twitter account; “followers fb” = number of people who follow you on your Facebook account; “PHQ” = four PHQ items;
“support” = when using social media I usually feel social pressure versus social support; “happy” = when using social media I usually
feel unhappy versus happy; “relaxed” = when using social media I usually feel anxious versus relaxed; “less sleep” = because of social
media I get much less sleep versus get much more sleep.
48
Table 1.2: Summary Statistics
(1) (2) (3)
2017 2019 2020
Personal Characteristics
Age 49.39 51.55 51.69
(15.65) (15.68) (15.90)
Gender - Male 0.450 0.443 0.429
(0.498) (0.497) (0.495)
Household Income 11.39 11.72 11.64
(4.075) (3.899) (3.935)
Marital Status 2.775 2.707 2.725
(2.122) (2.082) (2.090)
Highest Level of Education 11.47 11.56 11.52
(2.258) (2.243) (2.241)
Currently Working 0.65 0.62 0.61
(0.48) (0.49) (0.49)
Disabled 0.08 0.08 0.08
(0.27) (0.27) (0.27)
Hours of Work per Week 39.57 39.07 39.15
(11.63) (12.74) (11.97)
Social Media Usage
Twitter 0.290 0.327 0.205
(0.454) (0.469) (0.404)
Facebook 0.777 0.700
(0.416) (0.458)
Minutes Spent on Social Media per Day 437
(29703)
Connections on Social Media
Number of Facebook Friends 514
(3307)
Number of Twitter Friends 162
(1002)
Number of Facebook Followers 228
(1024)
Number of Twitter Followers 192
(2067)
Using Social Media Makes Me...
Feel Supported (0-100) 50.24
(26.70)
Feel Happy (0-100) 57.83
(23.72)
Feel Relaxed (0-100) 58.56
(27.27)
Standard deviation in parentheses. The statistics are weighted. Household income is annual in-
come.
49
Table 1.3: Descriptive Statistics - CESD & PHQ
(1) (2) (3)
2017 2019 2020
CESD 4.515 1.620 PHQ 1.908
(6.125) (4.438) (2.750)
1. You felt depressed. 0.295 0.308 1. Feeling nervous, anxious, or on edge. 0.686
(0.893) (0.911) (0.908)
2. You felt that everything you did was an effort. 0.568 0.604 2. Not being able to stop or control worrying. 0.453
(1.175) (1.203) (0.800)
3. Your sleep was restless. 0.857 1.006 3. Not being able to stop or control worrying. 0.407
(1.355) (1.417) (0.753)
4. You were unhappy. 0.495 0.510 4. Little interest or pleasure in doing things. 0.371
(1.114) (1.127) (0.725)
5. You felt lonely. 0.576 0.475
(1.182) (1.096)
6. You do not enjoyed life. 0.463 0.500
(1.084) (1.119)
7. You felt sad. 0.675 0.630
(1.253) (1.222)
8. You could not get going. 0.603 0.592
(1.203) (1.194)
Standard deviation in parentheses.CESD include eight items, and each item has a score ranging from 0 to 3. The final CESD score is
sum of all eight items’ scores and ranges from 0 to 24. Similarly PHQ has four items and each item has a score ranging from 0 to 3. The
final PHQ score is sum of all four items’ scores and range from 0 to 12. For both CESD and PHQ, higher scores represent higher levels
of depressive symptoms. The statistics are weighted.
50
Table 1.4: Social Media Users Versus Non-users
(1) (2) (3) (4) (5) (6)
Non-user User Non-user User
(Twitter) (Twitter) Difference (Facebook) (Facebook) Difference
2-11 Variable N Mean/SE N Mean/SE (1)-(2) N Mean/SE N Mean/SE (4)-(5)
Age 2086 54.98 824 48.69 6.29*** 1166 56.61 4010 50.11 6.50***
(0.34) (0.49) (0.49) (0.24)
Gender - Male 2091 0.480 824 0.482 -0.003 1167 0.597 4012 0.399 0.198***
(0.011) (0.017) (0.014) (0.008)
Household Income 2086 11.19 823 12.55 -1.36*** 1164 11.38 4005 11.82 -0.43***
(0.09) (0.13) (0.12) (0.06)
Marital Status 2090 2.62 824 2.53 0.08 1167 2.62 4010 2.73 -0.11
(0.04) (0.07) (0.06) (0.03)
Highest Level of Education 2091 11.33 824 12.22 -0.89*** 1167 11.42 4012 11.61 -0.19**
(0.05) (0.07) (0.07) (0.04)
Currently Working 2091 0.57 824 0.72 -0.15*** 1166 0.52 4010 0.65 -0.13***
(0.01) (0.02) (0.02) (0.01)
Disabled 2091 0.096 824 0.070 0.026** 1166 0.068 4010 0.085 -0.017*
(0.006) (0.009) (0.007) (0.004)
Hours of Work per Week 1176 39.40 576 41.26 -1.86*** 600 39.13 2504 39.05 0.08
(0.32) (0.48) (0.47) (0.26)
CESD 1968 4.477 768 4.483 -0.006 1167 1.724 4012 1.590 0.133
(0.141) (0.229) (0.132) (0.070)
The value displayed for t-tests are the differences in the means across the groups. * p< 0:10, ** p< 0:05, *** p< 0:01. Column (1)-(3)
use individuals who participate in both 2017 and 2019 surveys. Column (4)-(6) use individuals who participate in 2019 survey. Mean
CESD scores are different for Column (1)-(3) and (4)-(6), but both Column (3) and (6) show there is no significant different in CESD
score between social media users and non-users.
51
Table 1.5: Descriptive Statistics
(1) (2) (3) (4) (5) (6) (7) (8) (9)
Support: 0 Support: 45-55 Support: 100 Happy: 0 Happy: 45-55 Happy: 100 Relaxed: 0 Relaxed: 45-55 Relaxed: 100
Age 54.91 46.79 48.29 53.94 48.01 48.73 55 47.98 50.16
(16.73) (15.57) (15.83) (18.14) (16.03) (16.42) (17.35) (15.91) (15.97)
Gender - Male 0.45 0.42 0.39 0.39 0.41 0.33 0.40 0.42 0.40
(0.50) (0.49) (0.49) (0.49) (0.49) (0.47) (0.49) (0.49) (0.49)
Household Income 11.06 11.51 10.61 10.21 11.44 10.50 10.80 11.44 11.06
(4.04) (4.10) (4.42) (4.51) (4.06) (4.51) (4.24) (4.08) (4.22)
Marital Status 2.77 2.92 3.23 3.08 2.91 3.07 2.89 2.90 2.98
(2.10) (2.20) (2.21) (2.16) (2.18) (2.19) (2.13) (2.18) (2.17)
Highest Level of Education 11.07 11.41 10.72 10.57 11.35 10.63 10.60 11.37 11.01
(2.30) (2.18) (2.11) (2.32) (2.22) (2.08) (2.30) (2.22) (2.15)
Currently Working 0.51 0.67 0.58 0.48 0.64 0.54 0.53 0.65 0.59
(0.50) (0.47) (0.50) (0.50) (0.48) (0.50) (0.50) (0.48) (0.49)
Disabled 0.08 0.06 0.12 0.09 0.07 0.13 0.09 0.06 0.09
(0.27) (0.24) (0.32) (0.28) (0.25) (0.33) (0.29) (0.25) (0.29)
Hours of Work per Week 40.36 39.30 40.06 38.09 39.04 38.22 38.37 39.44 39.67
(12.26) (11.69) (15.28) (12.36) (11.80) (15.41) (11.54) (11.55) (13.61)
PHQ-2020 1.69 2.01 1.84 1.91 1.92 1.75 1.79 1.84 1.29
(2.79) (2.80) (3.02) (3.28) (2.73) (2.91) (2.98) (2.66) (2.43)
Observations 489 1885 203 195 2561 274 270 2491 434
Standard deviations in parentheses. This table shows mean of individual’s characteristics among groups with extreme feelings about
using social media (extremely negative with zero scores or extremely positive with 100 scores), and mild opinions about using social
media (with 45 to 55 scores). people who feel more positive about using social media are younger, with lower income, and higher
proportion that are currently working, or disabled.
52
Table 1.6: Social Media Usage Decreases the Level of Depressive Symptoms
CESD (2017, 2019)
Pooled-OLS Two-Way FE
Twitter (2017, 2019) 0.609** -1.154** -1.150**
(0.291) (0.463) (0.463)
Covariates X 5 X
Weights X X X
Mean(outcome)
Twitter=0 4.349
(6.055)
Adjusted R-squared 0.17 0.64 0.64
F-statistics 4.95 1.22 1.25
Observations 2198 2198 2198
Standard errors in parentheses. * p < 0:10, ** p < 0:05, *** p < 0:01. The coefficients in the
table represents effect of using Twitter (Twitter = 1) on final CESD score (range from 0 to 24).
This table shows using Twitter can decrease the level of depressive symptoms when keeping other
factors unchanged. The percentage impact is around 1.150/4.349 = 26.44%. The pooled-OLS
result finds positive correlation between social media usage and depressive symptoms, which is
consistent with existing literature. Adding covariates in the two-way fixed effects model does not
change estimation much indicating the selection bias on observables is limited.
53
Table 1.7: Social Media Usage Decreases the Level of Depressive Symptoms
(1) (2) (3) (4) (5) (6) (7) (8)
FE (one member each household)
Variable CESD1 CESD2 CESD3 CESD4 CESD5 CESD6 CESD7 CESD8
Twitter (2017, 2019) -0.083 -0.107 -0.265** -0.184* -0.102 -0.054 -0.167 -0.215*
(0.085) (0.133) (0.132) (0.112) (0.095) (0.116) (0.129) (0.123)
Covariates X X X X X X X X
Weights X X X X X X X X
Mean(outcome)
Twitter=0 0.277 0.585 0.932 0.513 0.498 0.470 0.618 0.575
(0.869) (1.189) (1.389) (1.130) (1.117) (1.091) (1.213) (1.181)
Adjusted R-squared 0.47 0.42 0.50 0.47 0.50 0.42 0.45 0.48
F-statistics 1.08 0.86 2.29 1.00 1.32 0.89 1.06 1.77
Observations 1980 1980 1980 1980 1980 1980 1980 1980
Standard errors in parentheses. * p< 0:10, ** p< 0:05, *** p< 0:01. This table decomposes outcome variable to eight CESD items
(CESD1-CESD8): 1. You felt depressed; 2. You felt that everything you did was an effort; 3. Your sleep was restless; 4. You were
unhappy; 5. You felt lonely; 6. You do not enjoyed life; 7. You felt sad; 8. You could not get going. The coefficients in the table
are estimation of impact of using Twitter (Twitter = 1) on each CESD score (ranges from 0 to 3) from two-way fixed effects model.
The data include only one member from each household. Twitter usage decreases the third, fourth and eighth CESD items by 28.43%
(0.265/0.932), 35.87% (0.184/0.513), and 37.39% (0.215/0.575), respectively.
54
Table 1.8: No Pre-trend in CESD
DTwitter
20192017
DTwitter
20202019
DTwitter
LaggedDCESD 0.004 -0.000 -0.001
(0.003) (0.002) (0.002)
Observations 1024 1024 2048
Standard errors in parentheses. * p< 0:10, ** p< 0:05, *** p< 0:01.This table shows previous
CESD trend does not change Twitter usage. Column (1) regresses change of Twitter usage between
2017 and 2019 on change of CESD between 2015 and 2017. Column (2) regresses change of
Twitter usage between 2019 and 2020 on change of CESD between 2017 and 2019. Column (3)
combines column (1) and (2) and regresses change of Twitter usage between t and (t+1) on change
of CESD between (t-1) and t.
55
Table 1.9: Switcher Groups’ Characteristics
Twitter(1!0) Twitter(unchanged) Twitter(0!1) Difference
Variable (2017) Mean/SE Mean/SE Mean/SE Twitter(0!1)-Twitter(1!0)
Age 52.37 53.34 52.48 0.11
(2.06) (0.49) (1.62)
Gender - Male 0.51 0.51 0.43 -0.08
(0.06) (0.01) (0.05)
Household Income 11.97 11.52 11.59 -0.38
(0.58) (0.13) (0.44)
Married 0.75 0.59 0.60 -0.16*
(0.06) (0.02) (0.06)
Education 11.80 11.63 11.89 0.09
(0.26) (0.07) (0.24)
Number of Children (2017) 2.95 2.44 2.50 -0.5
(0.22) (0.05) (0.18)
Number of Children (2019) 2.92 2.49 2.61 0.3
(0.22) (0.05) (0.18)
Increased Number of Children (2019-2017) 0.00 0.09 0.18 0.18**
(0) (0.01) (0.08)
Observations 5.9% 86.7% 7.4%
The value displayed for t-tests are the differences in the means across the groups. * p< 0:10, ** p< 0:05, *** p< 0:01.
56
Table 1.10: Income and Marital Status Are Not Mediators
Income Marital Status
Twitter (2017, 2019) 0.038 0.026
(0.220) (0.074)
Weights X X
Adjusted R-squared 0.89 0.92
F-statistics 0.03 0.12
Observations 1994 2000
Standard errors in parentheses. * p< 0:10, ** p< 0:05, *** p< 0:01.
57
Table 1.11: Social Media Usage Decreases the Level of Depressive Symptoms
(Main Member Only)
CESD (2017, 2019)
Pooled-OLS Two-Way FE
Twitter (2017, 2019) 0.579* -1.236** -1.175**
(0.279) (0.519) (0.510)
Covariates X 5 X
Weights X X X
Mean(outcome)
Twitter=0 4.313
(6.028)
Adjusted R-squared 0.13 0.63 0.63
F-statistics 5.33 5.67 1.29
Observations 1994 1994 1994
Standard errors in parentheses. * p< 0:10, ** p< 0:05, *** p< 0:01. The coefficients in the table
represents effect of using Twitter (Twitter = 1) on final CESD score (range from 0 to 24). Column
(1)-(3) keep only the main household member in the sample, and discard other household members
from the same household to address potential violation of Stable Unit Treatment Value Assumption
(SUTV A). This table shows using Twitter can decrease the level of depressive symptoms when
keeping other factors unchanged. The percentage impact is around 1.175/4.313 = 27.24%. The
estimated effects are0:19s.
58
Table 1.12: Estimated Effect of Using Twitter on CESD
Pooled-OLS Individual-FE Two-Way FE
All Household Members
0.006 -1.181** -1.182**
(0.261) (0.468) (0.468)
Covariates 5 5 5
Weights X X X
Adjusted R-squared 0.001 0.633 0.634
F-statistics 0 6.361 6.381
Observations 2206 2206 2206
0.437 -1.154** -1.150**
(0.263) (0.463) (0.463)
Covariates X X X
Weights X X X
Mean(outcome)
Adjusted R-squared 0.133 0.638 0.639
F-statistics 5.777 1.224 1.253
Observations 2198 2198 2198
Main Household Memeber Only
0.187 -1.232** -1.236**
(0.276) (0.519) (0.519)
Covariates 5 5 5
Weights X X X
Adjusted R-squared 0.001 0.626 0.627
F-statistics 0.458 5.643 5.668
Observations 2000 2000 2000
0.579* -1.174** -1.175**
(0.279) (0.509) (0.510)
Covariates X X X
Weights X X X
Mean(outcome)
Twitter=0 4.440 4.313 4.313
(6.122) (6.028) (6.028)
Adjusted R-squared 0.134 0.632 0.633
F-statistics 5.332 1.246 1.290
Observations 1994 1994 1994
CESD Item Rescaled to [0, 1]
0.193** -0.391** -0.392**
(0.093) (0.170) (0.170)
Covariates X X X
Weights X X X
Adjusted R-squared 0.134 0.632 0.633
F-statistics 5.332 1.246 1.290
Observations 1994 1994 1994
Standard errors in parentheses. * p < 0:10, ** p < 0:05, *** p < 0:01. The coefficients in the
table represents effect of using Twitter (Twitter = 1) on final CESD score (range from 0 to 24).
The first two panels use all sample, without and with covariates. The last three panels use only
main household members. The last panel rescales CESD scores: each item in CESD now has a
score of either 0 or 1 instead of 0 or 3.
59
Table 1.13: Estimation for Two Switcher Groups
DCESD
20192017
D(DCESD
20192017
> 0)
Twitter
2017
= 0;Twitter
2019
= 1(t
1
) -1.313* -0.114**
(0.767) (0.064)
Twitter
2017
= 1;Twitter
2019
= 0(t
2
) 1.108* -0.016
(0.713) (0.070)
Covariates X X
Weights X X
Adjusted R-squared 0.01 0.01
Observations 1019 1019
F-test: t
1
=t
2
-0.18 -1.29
Standard errors in parentheses. * p < 0:10, ** p < 0:05, *** p < 0:01. This table shows the
estimation results from first difference model. In column (1) dependent variable is difference in
CESD. In column (2) dependent variable is whether the difference in CESD increases. I use F-test
to examine if effect of increasing Twitter usage and decreasing Twitter usage is asymmetric. F-test
rejects asymmetry in both specifications.
Table 1.14: Fixed Effect Model with CESD Winsorization
CESD winsorized at:
CESD6 CESD9 CESD12 CESD15 CESD18 CESD21 Original Sample
Twitter (2017, 2019) -0.668*** -0.816*** -0.911** -1.011** -1.111** -1.216** -1.175**
(0.233) (0.313) (0.372) (0.425) (0.475) (0.502) (0.510)
Covariates X X X X X X X
Weights X X X X X X X
Adjusted R-squared 0.55 0.59 0.60 0.61 0.62 0.63 0.63
F-statistics 1.90 1.97 1.71 1.45 1.34 1.33 1.29
Observations 1994 1994 1994 1994 1994 1994 1994
Standard errors in parentheses. * p< 0:10, ** p< 0:05, *** p< 0:01. The estimations are from
two-way fixed effects model with one member from each household. Dependent variable CESD is
winsorized at different values in this table. For example, in column (1) any CESD value larger than
6 will be replaced by 6 in estimation. The estimated Twitter impact after winsorization is consistent
with the main specification. This finding suggests that the main result is robust to outliers.
60
Table 1.15: Attrition Statistics
(1) (2) Difference
Attrited Panel (1)-(2)
Age 49.98 53.55 -3.57***
(0.74) (0.31)
Gender - Male 0.42 0.49 -0.08***
(0.02) (0.01)
Household Income 10.87 11.73 -0.86***
(0.19) (0.08)
Marital Status 2.95 2.53 0.42***
(0.09) (0.04)
Highest Level of Education 11.25 11.66 -0.41***
(0.10) (0.05)
Currently Working 0.57 0.62 -0.06**
(0.02) (0.01)
Disabled 0.11 0.08 0.03*
(0.01) (0.01)
Hours of Work per Week 40.60 39.99 0.60
(0.68) (0.29)
CESD (2017) 4.76 4.22 0.53
(0.49) (0.17)
Twitter (2017) 0.27 0.31 -0.04
(0.03) (0.01)
* p< 0:10, ** p< 0:05, *** p< 0:01 for t-test. Standard deviation in parentheses. The statistics
are weighted. Household income is annual income. CESD and Twitter usage in 2017 are not
different for attrition sample and panel data sample.
61
Table 1.16: Effect of Social Media Usage on Depressive Symptoms (2020)
OLS
(1) (2) (3) (4)
PHQ-2020 Social Support Feel Happy Feel Relaxed
Twitter 0.342*** 3.05* 0.52 -1.94
(0.122) (1.59) (1.36) (1.62)
Facebook -0.084 13.28*** 12.04*** 10.85***
(0.100) (1.87) (1.85) (2.06)
Covariates X X X X
Past CESD X X X X
Weights X X X X
Mean(outcome)
Twitter=0 1.885 49.32 57.56 58.80
(2.755) (27.37) (25.11) (28.08)
Facebook=0 1.815 41.87 51.69 53.36
(2.790) (28.17) (26.93) (29.33)
Adjusted R-squared 0.24 0.08 0.08 0.05
F-statistics 11.51 3.17 3.36 2.46
Observations 3081 2402 2397 2401
OLS
(5) (6) (7) (8)
PHQ-2020 Social Support Feel happy Feel relaxed
Twitter 0.902*** -0.45 3.20 -0.67
(0.329) (2.950) (2.80) (3.24)
Facebook -0.639 9.19* 0.02 -0.69
(0.596) (5.33) (5.07) (5.86)
Log(Minutes per Day) 0.116 1.78 2.21* -0.09
(0.135) (1.21) (1.15) (1.33)
Log(Facebook Connections) 0.513*** 2.02 1.66 0.99
(0.148) (1.32) (1.26) (1.46)
Log(Twitter Connections) 0.024 1.63 0.80 0.38
(0.131) (1.17) (1.11) (1.29)
Log(Facebook Followers) -0.449*** -1.40 0.64 -0.38
(0.111) (0.99) (0.94) (1.09)
Log(Twitter Followers) 0.029 -0.75 -1.63 -1.19
(0.119) (1.06) (1.01) (1.17)
Covariates X X X X
Past CESD X X X X
Weights X X X X
Adjusted R-squared 0.32 0.11 0.11 0.10
F-statistics 3.32 1.62 1.65 1.54
Observations 455 455 455 455
Standard errors in parentheses. * p< 0:10, ** p< 0:05, *** p< 0:01. Twitter usage and PHQ
have positive correlation. Both Twitter and Facebook usage positively correlate with positive feel-
ing about using social media. Number of Facebook connections and followers have opposite cor-
relations with PHQ. PHQ ranges from 0 to 12.
62
Table 1.17: Effect of Social Media Usage (2020)
OLS
PHQ-2020 Social Support Feel Happy Feel Relaxed
Twitter 0.369*** 2.26* 2.23* -1.59
(0.128) (1.34) (1.19) (1.39)
Twitter(300-400 Min/Day) -2.189*** 1.77 4.68 -6.95
(0.801) (8.16) (7.28) (8.47)
Facebook -0.150 12.71*** 11.15*** 10.12***
(0.100) (1.44) (1.28) (1.49)
Facebook(200-300 Min/Day) -0.859 -40.16** -25.63 -37.80*
(1.864) (18.97) (16.91) (19.69)
Facebook(200-400 Facebook Connections) -4.628 -26.92 66.41* 55.96
(4.694) (43.57) (38.02) (44.15)
Standard errors in parentheses. * p < 0:10, ** p < 0:05, *** p < 0:01. Only regressors with
nonzero coefficients are shown here. PHQ ranges from 0 to 12.
Table 1.18: Previous CESD Does Not Change Twitter Usage
Twitter(2017) Twitter(2019) Twitter(2020)
CESD(2015) 0.002 0.002 0.002
(0.003) (0.003) (0.002)
CESD(2017) 0.001 -0.001
(0.003) (0.002)
CESD(2019) -0.002
(0.002)
Weights X X X
Observations 1266 1078 3290
Standard errors in parentheses. * p< 0:10, ** p< 0:05, *** p< 0:01.
Table 1.19: Depression Alleviated
D(CESD< 3) D(CESD< 4)
Twitter (2017, 2019) 0.154*** 0.069*
(0.054) (0.041)
Adjusted R-squared 0.48 0.52
F 1.51 1.96
Observations 1994 1994
Standard errors in parentheses. * p< 0:10, ** p< 0:05, *** p< 0:01. Cutoff CESD =3 (Kozlov
et al., 2020) and cutoff CESD =4 (Mhaolain et al., 2012) are used in the literature.
63
Table 1.20: Computing Labor Market Benefit
Groups t
d
(cesd < 3) t
d
(cesd < 4) Pop(K) Min/Workday Wage Saved LPT(hour)/wk Benefit1(billion) Benefit2(billion)
Men 0.227 0.049 73,824 24.92 39.547 2.023 1.34 0.29
Women 0.083 0.082 56,773 36.68 31.877 1.043 0.16 0.15
20-24 0.223 0.126 9,740 61.01 21.943 -0.984 -0.05 -0.027
25-54 0.171 0.11 89,893 37.95 37.858 0.937 0.55 0.35
>55 0.132 0.068 29,439 21.39 33.268 2.317 0.30 0.15
20-24, Men 0 0 5,314 31.05 21.611 1.513 0 0
20-24, Women 0.172 0.104 4,426 69.82 22.041 -1.718 -0.03 -0.02
25-54, Men 0.228 0.131 50,806 33.58 41.381 1.302 0.62 0.36
25-54, Women 0.155 0.109 39,087 40.8 35.59 0.7 0.15 0.11
>55, Men 0.186 -0.029 16,815 15.98 38.593 2.768 0.33 -0.05
>55, Women 0.067 0.163 12,623 26.52 28.229 1.89 0.05 0.11
First two columns are estimation of D(CESD< 3) and D(CESD< 4) on Twitter usage for different groups. Pop is working population
count in each group (in thousand). Min/workday is average number of minutes on social media per work day for each group. Wage
is average hourly income for each group. Saved LPT/wk is saved productive work hours per week in each group, which equals 4.1-
(Min/workday)/60 5. The last two columns estimate labor market benefit per week for each group. Benefit1 is based on first column
and benefit 2 is based on second column.
64
Chapter 2
Selection Bias Reduction by Matching and Weighting
Estimators in Social Media Data Collection
2.1 Introduction
Traditional academic data collection approaches are costly, time-consuming and the types of
questions are limited. For example, family life surveys conducted around the world require lots of
human efforts and funds to proceed. For any organizations working on those types of surveys, they
need to raise enough money, select a sample representing target population, train local interviewers,
contact subjects, conduct interviews, and record and clean up the data. In some cases, interviewers
need to travel to remote areas in order to reach certain subjects. Even then they may still fail to find
the subjects. Therefore, traditional data collection methods are costly and subject to high attrition
rate. Survey designers may have to recruit fewer subjects who live in distant locations considering
the relatively high cost. Then researchers who use the data must either justify why their conclusions
are not affected by the unbalanced samples or find proper data adjustment procedures.
Considering the weaknesses of traditional data collection methods, more and more attention has
been drawn to Internet surveys or web surveys. Many studies discussed pros and cons of Internet
surveys over the last two decades. This chapter focuses on an extended version of Internet surveys:
social media surveys. While there are many similarities between Internet surveys and social media
surveys, the latter have unique advantages and challenges. To the best of my knowledge, there was
65
no formal discussion about quality, bias or data adjustment approaches for social media surveys.
While it would be ideal to have actual social media survey data to study these topics, the data in
this chapter can still shed light on these subjects. Propensity-score-based methods are used heavily
in Internet survey literatures. The general opinion is that these methods reduce selection bias
when proper auxiliary variables are available but they never eliminate selection bias completely.
It is interesting to discuss whether social media survey can potentially perform better than typical
Internet surveys such as access panels.
Therefore, this chapter studies how matching and weighting methods perform in correcting
selection bias in social media survey data. The matching methods include K-Nearest Neigh-
bors Matching (KNN), Radius Matching, and Mahalanobis Matching. The first two methods use
propensity score and the last method matches observations on Mahalanobis distance of covariates.
I also include a weighting method which also uses propensity score: Inverse Probability Weighting
(IPW). These methods are common in bias correction literature and easy to implement.
Literature discusses two main Internet surveys categories: non-probability-based methods and
probability-based methods (Couper, 2000; Greenacre et al., 2016). Important non-probability-
based methods include self-selected web surveys (no access restrictions, can participate multiple
times) and volunteer panels of Internet users (restrictive access). However, statistical inference
cannot be applied to non-probability-based data. Any presentation of sampling error and confi-
dence intervals is misleading (Couper, 2000). Important probability-based methods include mixed
mode surveys, pre-recruited panels of Internet users, and probability samples of full population.
Mixed mode surveys offer different completion methods at different stages, e.g. web survey only
at the first stage, web or mail survey at the second stage (Schonlau, Asch, et al., 2003). Although
mixed mode surveys can achieve higher response rate than web surveys alone, the mode effect need
to be further studied. Pre-recruited panels of Internet users are recruited by probability sampling
methods such as Random Digit Dialing (Couper, 2000). Probability samples of full population
66
recruit subjects by non-Internet channels and provide them with Internet access for later web sur-
veys (Couper, 2000). The data used in this chapter, Understanding America Study, is a probability
sample of full population.
Literature discusses several advantages of Internet surveys and social media surveys over tra-
ditional surveys. Web surveys can be completed by researchers with fast speed, low cost, and
high response rate from busy people or people who screen phone calls. Web surveys also avoid
interviewer effects and offer convenience for visual research (Duffy et al., 2005). It is easier to
convince population who are hesitant to participate in surveys in person to participate in web sur-
veys (Pedersen et al., 2015). Data collection through social media like Twitter is cheaper, faster,
and typically provides larger sample sizes than data collection through traditional methods, making
it particularly appealing for monitoring diseases that have rapid transmission rates (Dos Reis and
Culotta, 2015).
Literature also discusses several disadvantages of Internet surveys and social media surveys.
For example, selection issues, i.e., certain people are impossible or less possible to participate in
social media surveys and mode effect (Duffy et al., 2005), high drop out rate (Birnbaum, 2004),
high attrition rate for longitude studies (Pedersen et al., 2015), and difficulties in drawing statistical
inference from non-probability-based data (Biffignandi, 2011). The existence of mode effects in
web surveys has been proven in the literature but the findings are inconclusive. In some cases
online respondents prefer “midpoint” and “don’t know” options, and in other cases respondents
prefer extreme values (Duffy et al., 2005).
The bias types in Internet surveys include coverage bias, sampling bias, nonresponse bias, and
measurement error (Couper, 2000; Lee, 2004; Couper et al., 2007; Greenacre et al., 2016). Each of
the bias types occurs at different stages of data collection. Discrepancy between target population
and the population who have access to Internet creates coverage issues. Sampling among those
with Internet access incurs sampling errors. The characteristics differences between subjects who
participate and the ones who are unwilling to participate create non-response bias. Lastly, the
measurement error is the deviation of the subjects’ responses from their true values on the measure
67
in this case (Couper, 2000). These biases can be critical in some cases, such as when the target
variable and the response behavior are strongly related, when the average response probability is
low, and when the variation in the response probabilities is large (Lee and Valliant, 2009). The
remaining questions are whether some of these biases would become unacceptable in social media
surveys, and what procedures can reduce them effectively if the bias is nontrivial.
Propensity score adjustment (PSA) is one of the most common methods to correct selection bias
in surveys. PSA can also be incorporated with other adjustment methods. Studies used PSA for
Average Treatment Effect estimation with biased samples (LaLonde, 1986; Dehejia and Wahba,
2002). The data from Kentucky, Ohio and West Virginia was used by researchers to evaluate
different bias reduction methods, which used propensity of having telephone to reduce coverage
bias in telephone surveys (Duncan, 2001). They found weighing on length of disconnect made no
improvement, weighting Class Adjustment Scheme and Raking Ratio Adjustment worked better
for large population, and the newly developed propensity weighting method AUGP performed
the best. Lee (Lee, 2004; Lee and Valliant, 2009) proposed a two-step adjustment approach to
volunteer panel web surveys. With this approach, non-randomized sample selection is corrected by
propensity score adjustment in the first step, and calibration adjustment is used to correct uncertain
coverage of the sampling frames to match target population in the second step. An online self-
administered survey with previously recruited online access panel respondents and a face-to-face
survey of randomly sampled respondents of the general population showed biases were presented
in both factual question data and attitudinal question data, and weighting adjustment had only small
effect in reducing biases (Loosveldt and Sonck, 2008). Researchers used HRS data and showed
the bias of Internet surveys was 6.5 % without PSA adjustment and 3.7 % otherwise (Schonlau,
Van Soest, et al., 2009).
Literature uses a variety of covariates to estimate propensity scores. For example, social-
demographic covariates include gender, age, work status, tenure, marital status, region, urbanicity,
education, income, and race/ethnicity; behavioral and attitudinal covariates include online pur-
chasing behavior, social pressure and rules, car ownership, Internet usage(hours/week), self-rated
68
health, self-rated memory, measures of attitudes about privacy, security and risk, and measures
of physical activities and product boycotting behavior (Duffy et al., 2005; Loosveldt and Sonck,
2008; Couper et al., 2007).
There are three approaches to using propensity score as weight in the literature (Lee, 2004;
Lee and Valliant, 2009; Duffy et al., 2005; Valliant and Dever, 2011). The first approach is using
the inverse of propensity score as additional weight for each observation. The second approach is
sorting samples by propensity scores, dividing samples into subclasses, and using the inverse of
average propensity score of each subclass as the common weight for all samples within that sub-
class. The third approach is creating subclasses as in the second approach to estimate population
counts in each subclass and creating a poststratification estimator. Propensity score can also be
used in matching methods in which each individual in Internet survey is matched with a unit or
units in reference survey who has/have similar propensity score, and receives weight of the refer-
ence unit as its own weight (Valliant and Dever, 2011). For example, each Twitter participant can
be matched with an individual from the control group based on friendship, gender, location and
numbers of tweets, and followers (Dos Reis and Culotta, 2015).
Other adjustment methods include poststratification or weighting class adjustments, raking or
rim weighting, Generalized regression (GREG) modeling, etc. (Greenacre et al., 2016). How-
ever, simple stratification may not be effective enough to ensure the representativeness of online
surveys (Blasius and Brandt, 2010). Literature discussed a pseudo-randomization approach which
is a combination of participation probability and further adjustment method based on a structural
population model (Valliant and Dever, 2011). With this approach, researchers construct estimators
based on models for quota samples. Covariate balancing propensity score (CBPS) is another ad-
vanced method which emphasizes the trade-off between two roles of propensity score: conditional
probability of treatment assignment and a balancing score (Imai and Ratkovic, 2014).
Other related studies include estimation of PSA estimator variance (Lee and Valliant, 2009), a
difference in differences approach to evaluating relative trends when no ground truth data is avail-
able (Zagheni and Weber, 2015), a strategy study of data collection process to make Internet survey
69
data superior to the equivalent paper survey data (Chang and V owles, 2013), a practical study of
using social media platform (Facebook) to recruit participants for online mental health research
(Pedersen et al., 2015), a study of mixed mode web survey which showed that phone reminders
have nontrivial effect in increasing web response rate (Schonlau, Asch, et al., 2003), a discussion
of over-surveying effect when people receive too many survey requests (Couper, 2000), a discus-
sion of the potential of social media surveys in developing countries (Tijdens and Steinmetz, 2016).
These studies discussed potential social media survey designs and supported the idea that social
media can be a cost-effective channel for data collection and social science research.
Some datasets are common in a variety of studies on Internet surveys: 2002 General Social
Survey Data and 2003 Michigan Behavioral Risk Factor Surveillance Survey Data (Lee, 2004),
CAPI (MORI omnibus) data (Duffy et al., 2005)), European Social Survey (ESS) and Behavioral
Risk Factor Surveillance System (BRFSS) (Schnell et al., 2017; Valliant and Dever, 2011), Health
and Retirement Study (HRS) (Couper, 2000; Schonlau, Asch, et al., 2003).
The remainder of the chapter is organized as follows. Section 2.2 discusses the theoretical
and empirical foundation of different types of biases in social media surveys. Section 2.3 summa-
rizes data sources and the main variables of interest. Section 2.4 discusses a variety of matching
and weighting estimators. Section 2.5 discusses empirical performances of different estimators.
Section 2.6 provides the conclusion.
2.2 Background
2.2.1 Social Media Survey Design
Recruiting through social media advertisement would be more cost-effective than through tra-
ditional mail, telephone or popular website advertisement. One of the most effective advertisement
schemes is targeted advertisement. For example, almost all social media platforms provide adver-
tisement service which allows advertisement to target a specific population based on gender, age,
education, region, marital status, cellphone using behavior, etc. Future social media data collection
70
can also use this service to target specific population to recruit survey participants. Then the par-
ticipants can be led to the website of actual surveys. Participants may have the option to complete
one part of the survey at a time. This design will reduce the time cost for participants and increase
response rate.
Social media surveys create two types of incentives for participation. The first incentive is
monetary rewards, which are commonly used in all survey methods. The second incentive is
curiosity. Sharing polling results as a reward for participation can help the academic data collection
avoid being “boring and tedious.” For example, survey results summary, “fun facts” from past
surveys, and your friends’ responses summary can be shared with survey participants. Current
survey information should be shared with participants with caution in case they incur participants
bias their answers strategically. Ideally, the shared survey results are completely irrelevant to
current survey questions.
2.2.2 Twitter and Bias in Social Media Survey Sample
Twitter is a very popular social media platform both in the US and around the world. It was
created in March 2006 by Jack Dorsey, Noah Glass, Biz Stone, and Evan Williams. Until 2016,
there were more than 319 million monthly active users. One of the most attractive features of
Twitter is “tweets.” One piece of information can quickly go viral and be shared among millions
of people.
The popularity of this social media platform makes it a good channel for data collection. This
social media population is large, and its coverage over different generations is also expanding.
Therefore, it is possible to recruit many participants through Twitter fast with low costs. At the
same time, much social-demographic information is publicly available which makes stratification
and targeting possible.
We need to consider network effect when we design a probability-based survey through social
media platforms, since the spread of information is through personal network on those platforms
which can create bias. The participants pool may include clusters of similar people.
71
Measurement error is created when people do not give true answers. There are three types of
mechanisms which induce measurement error (Table 2.1).
For the strategic bias case, we can solve the problem from two directions. The first direction is
trying to design a mechanism preventing strategic bias and induce truth-telling behavior (at least in
equilibrium). The second direction is allowing strategic bias but trying to find estimation methods
to correct it. In cheap talk literature, if participant’s utility is closely aligned to researchers, a proper
partition size will result in informative data. In other words, it is important to create incentive
compatible constraints when we design a survey.
Selection bias in social media surveys comes from the coverage error, response bias, etc. For
example, if the target population is the whole American population, using social media survey will
preclude those who do not currently have social media accounts. This would not be a problem if
the variables of interest have the same distribution among those who have social media accounts
and who do not. However, that is rarely the case.
Despite the potential bias in social media survey method, it is still a valuable data collection
method. There are three reasons. First, Internet and social media access are spreading rapidly
around the world. More and more people can be reached through social media over time. In a
few years, it is possible that we can reach more people through social media than through mail or
telephone. Second, traditional surveys suffer from selection bias as well. Last but not least, when
the selection bias in social media survey is nontrivial, we can use a variety of estimation methods
to correct estimation and reduce bias to an acceptable range.
2.3 Data
2.3.1 Understanding America Study (UAS)
This chapter uses data from the Understanding America Study (UAS) to evaluate performances
of different bias correction methods. UAS is an Internet probability-based panel including approx-
imately 8,500 respondents. The panel is representative of the entire US adult population. The
72
coverage bias is addressed by providing necessary equipment and Internet access to the partici-
pants when they did not have the access. UAS covers a variety of topics. Data for this chapter were
collected by the UAS team in 2020.
1
The dataset used in this chapter includes basic demographic
information as well as information on social media usage, social connections, and willingness to
pay for COVID vaccine.
The overall response rate is 81.06%. The dataset contains 6932 observations. 68% respondents
use Facebook and 32% respondents do not use Facebook. 20% respondents use Twitter and 80%
respondents do not use Twitter. Figure 2.1 shows the distribution of social media usage and the age
distributions for each user group. Facebook users are younger than those who do not use Facebook
on average. Similarly, Twitter user population is younger than those who do not use Twitter, and
the gap in Twitter’s case is larger than the gap in Facebook’s case.
2.3.2 Income, Number of Family and Close Friends, and Willingness to Pay
for COVID-19 Vaccine
This chapter examines how different methods match those who use social media and who
do not use social media on three variables. The first variable is income, which is the average
household income of the income group which respondents belong. The second variable is number
of family and close friends. This variable represents the social connections of the respondent.
It would be helpful if researchers can derive respondents’ social connections from a simple set
of basic demographic information. The third variable is the willingness to pay for COVID-19
vaccine.
2
This variable is attitudinal and more subjective than the other two variables. Therefore,
these three variables cover objective measures as well as subjective measures, which are important
information for research but are harder for researchers to get in many cases. Figure 2.2 shows
distributions of these three variables. Distributions of the number of family and close friends
1
The data come from survey “UAS230.”
2
Willingness to pay for vaccine comes from the question “Suppose that a 100% safe and effective coronavirus
vaccine is developed today but insurance does not cover it. How much would you be willing to pay to get yourself
vaccinated?”
73
variable and willingness to pay for vaccine variable are skewed, so I use the natural logarithm of
the two variables instead. These distributions roughly follow the normal distribution.
Figure 2.3 shows the distributions of the three variables (income, natural logarithm of number
of family and close friends, natural logarithm of willingness to pay for COVID-19 vaccine) for
Facebook users (orange) and non-Facebook users (blue). Figure 2.4 shows the distributions of the
three variables (income, natural logarithm of number of family and close friends, natural logarithm
of willingness to pay for COVID-19 vaccine) for Twitter users (orange) and non-Twitter users
(blue). Social media users and non-users follow similar distributions in all the cases.
2.3.3 Summary Statistics
Figure 2.5 shows the distributions of marital status, education, and race of the sample. 55%
respondents are married, 2% respondents are separated, 14% respondents are divorced, 5% respon-
dents are widowed, and the rest 24% respondents are never married. Approximately 5% respon-
dents have education level below 12th grade, 17% respondents were high school graduate or GED,
23% respondents have some college education but no degree, 50% respondents have some un-
dergraduate level degree (education categories 10-13), and 4% respondents have master’s degree,
professional degree or doctorate degree. 77% respondents are white, 8% respondents are African
Americans, 2% respondents are American Indian or Alaska Native, 5% respondents are Asian, 1%
respondents are Hawaiian or Pacific Islander, and the rest 5% respondents are multiracial.
Tables 2.2 summarizes the mean and standard deviations of these key variables in each social
media user group. Respondents who use Facebook or Twitter are different from who do not use
these social media platforms. Facebook users have higher income, higher number of family/close
friends, lower willingness to pay for COVID-19 vaccine, higher probability of also using Twitter,
lower age, and higher education level than those respondents who do not use Facebook. The gender
and race distributions are also different. Twitter users share similar patterns, except that Twitter
users have fewer number of family/close friends, higher willingness to pay for COVID-19 vaccine,
and less likely to be married than respondents who do not use Twitter. The proportions of different
74
race categories among Facebook users and Twitter users are also different. Therefore, the different
characteristics of social media users and non-users indeed raise concerns that data collected from
social media users alone would not be representative enough without any bias correction procedure.
The next sections address this problem by examining the performances of different bias correction
methods on increasing the representativeness of social media users in three selected variables.
2.4 Empirical Strategy
This chapter examines the performances of K-Nearest Neighbors Matching (KNN), Radius
Matching, Inverse Probability Weighting (IPW), and Mahalanobis Matching. The criterion is how
well these methods can match social media users and non-users on income, number of family
or close friends, and willingness to pay for COVID-19 vaccine. Mahalanobis Matching matches
samples based on variable distance directly, and the rest of the methods use propensity score.
Logistic regression is used to compute propensity score for each observation:
Log
p
k
i
1 p
k
i
=b
k
X
i
+e
k
i
(2.1)
where p
k
i
is the propensity score for individual i in case k=fFacebook, Twitterg, i.e., the prob-
ability that individual i is a Facebook (or Twitter) user based on his or her basic information X
i
.
X
i
includes the basic demographic variables: gender, age, marital status, education, race, and their
polynomial terms.
K-Nearest Neighbors Matching method matches each social media user with k respondents
who do not use the social media platform according to the closeness of the odds ratio (Leuven and
Sianesi, 2003). k is a hyperparameter we can select. Section 3.5 also compares the performances
of KNN with different k values.
3
3
The sort order of data could affect the estimation results with Nearest Neighbor Matching on a propensity score
estimated with non-continuous variables.
75
Radius Matching method assigns each social media user all the respondents who do not use the
social media platform within the specified radius given by the caliper. The caliper c is a hyperpa-
rameter which affects the performance of Radius Matching. The choice of number of neighbors
k in KNN or radius caliper c triggers bias-variance trade-off. Increasing k or c increases bias but
decreases variance.
Inverse Probability Weighting (IPW) uses the inverse of propensity score as weight. The in-
tuition behind this weight is that we would like to assign more weights to observations which are
less likely to appear in the data as treated units (social media users), so that in the end we have a
representative group for population of interests.
Mahalanobis Matching method matches observations based on Mahalanobis Distance (MD).
MD is a scalar quantity that measures the multivariate distance between observations (Rosenbaum
and Rubin, 1985; Diamond and Sekhon, 2013). The MD of two observations i and j is
MD(X
i
;X
j
)=
q
(X
i
X
j
)
T
S
1
(X
i
X
j
) (2.2)
where S is the sample covariance matrix of X. Mahalanobis Matching tends to perform better when
X is continuous than when X is categorical. In terms of creating close pairs i and j, Mahalanobis
Matching performs better than matching based on Propensity Score because Mahalanobis distance-
paired units have close values on all of the covariates, whereas propensity score-paired units may
be close on the propensity score but not on covariates themselves (Greifer, 2020).
2.5 Empirical Results
Table 2.3 shows the performances of K-Nearest Neighbors Matching, Radius Matching, Ma-
halanobis Matching, and Inverse Probability Weighting on matching Facebook users with respon-
dents who do not use Facebook. The first column shows the t-statistic of the income difference
between Facebook users and their counterparts after matching or weighting. All the matching
methods perform well and the difference is statistically insignificant. The weighting method does
76
not perform well in this case. The second column shows the t-statistic of the social connection
(number of family or close friends) difference between Facebook users and their counterparts after
matching or weighting. K-Nearest Neighbors Matching with a small number of neighbors performs
well, whereas other methods do not perform well in this case. The third column shows the t-statistic
of the attitudinal (willingness to pay for COVID-19 vaccine) difference between Facebook users
and their counterparts after matching or weighting. Similarly, only K-Nearest Neighbors Matching
with a small number of neighbors performs well.
Table 2.4 shows the performances of K-Nearest Neighbors Matching, Radius Matching, Maha-
lanobis Matching, and Inverse Probability Weighting on matching Twitter users with respondents
who do not use Twitter. The first column shows the t-statistic of the income difference between
Twitter users and their counterparts after matching or weighting. K-Nearest Neighbors Match-
ing with small number of neighbors, Mahalanobis Matching and IPW perform well in this case,
whereas Radius Matching or KNN with a large number of neighbors do not perform well. The
second column shows the t-statistic of the social connection (number of family or close friends)
difference between Twitter users and their counterparts after matching or weighting. All matching
and weighting methods perform well in this case. The third column shows the t-statistic of the
attitudinal (willingness to pay for COVID-19 vaccine) difference between Twitter users and their
counterparts after matching or weighting. All matching and weighting methods perform well in
this case as well.
Figure 2.6 shows the performances of K-Nearest Neighbors Matching with different parameter
values. We choose the parameter k, which is how many neighbors we match to each social media
platform user. The top panel shows the performances in the case of matching Facebook users and
respondents who do not use Facebook. The number of neighbors ranges from one to twenty. When
KNN matches income, the performance is stable and good as the number of neighbors changes.
KNN with nearest one neighbor performs well when it matches the number of family or close
friends. Then the performance becomes worse as the number of neighbors grow and the t-statistic
reaches a plateau around 4. In the case of matching willingness to pay for COVID-19 vaccine,
77
KNN with less than 6 neighbors performs well. The performance is not good when the number of
neighbors exceeds 5 and t-statistic reaches a plateau around -2.5.
The bottom panel of Figure 2.6 shows the performances in the case of matching Twitter users
and respondents who do not use Twitter. The number of neighbors ranges from one to twenty. In
the case of matching income, KNN with one, two or five neighbors perform well. The performance
has an upward trend as the number of neighbors increases. KNN with any number of neighbors
performs well when it matches number of family or close friends or willingness to pay for COVID-
19 vaccine. The performances with both variables stop increasing significantly when the number
of neighbors reaches 14. In the case of number of family or close friends, the t-statistic is stable
around 0. In the case of the willingness to pay for COVID-19 vaccine, the t-statistic is stable
around 0.5.
Figure 2.7 shows the performances of Radius Matching with different parameter values. We
choose the caliper c, and we match social media platform user and non-users only within this range.
The top panel shows the performances in the case of matching Facebook users and respondents who
do not use Facebook. The radius caliper ranges from 0.02 to 1 with step size 0.02.
4
When Radius
matching matches income, the performance is good if the radius is less than or equal to 0.22. In
the cases of matching the number of family or close friends and willingness to pay for COVID-19
vaccine, Radius Matching does not perform well.
The bottom panel of Figure 2.7 shows the performances in the case of matching Twitter users
and respondents who do not use Twitter. Radius Matching does not perform well in the case of
matching income. Radius Matching with radius caliper less than or equal to 0.26 performs well
when it matches number of family or close friends. The t-statistic reaches a plateau around -2.22
after that. In the case of matching willingness to pay for COVID-19 vaccine, Radius Matching
with any radius caliper c performs well, especially when c is less than 0.2.
Based on these empirical results, matching methods perform well in the following cases: im-
proving Facebook users’ representativeness in income, Twitter users’ representativeness in income,
4
Optimal radius caliper varies across cases. For example, calipers of width equal to 0.2 of the standard deviation
of the logit of the propensity score is used in some cases (Austin, 2011).
78
social connection (number of family or close friends) and willingness to pay for COVID-19 vac-
cine. Inverse Probability Weighting (IPW) performs well in the cases of improve Twitter uses’
representativeness, but not in the cases of improving Facebook uses’ representativeness. Facebook
users’ income data combining with matching methods can be representative for the whole adult
population, whereas it is harder to use bias correction methods to make Facebook users’ social
connections or attitudes towards COVID-19 vaccine be representative than in the income case (the
difference between Facebook users and non-users is not statistically different from zero only in
KNN with one nearest neighbor case). These matching and weighting methods perform better
with Twitter users’ data than with Facebook users’ data. KNN Matching, Mahalanobis Matching,
and IPW can easily improve Twitter users’ representativeness in income, and all examined meth-
ods perform well in improving Twitter users’ representativeness in social connection and attitudes
towards COVID-19 vaccine. Therefore, these results provide evidences that with appropriate bias
correction methods and a simple set of publicly available demographic information, researcher
would be able to use data (both objective and subjective) from social media surveys and take care
of selection bias issues.
2.6 Conclusion
This chapter finds that matching and weighting methods can effectively improve the represen-
tativeness of social media users in most cases. The best bias correction method varies depending
on the types of outcome variables. Hyperparameters in matching methods also affect the per-
formance. Improving the representativeness of Facebook users’ income is easier than improving
the representativeness of social media users’ social connection and attitude towards COVID-19
vaccine. Improving the representativeness of Twitter users is easier than improving the representa-
tiveness of Facebook users. The empirical findings of this chapter provide the evidence that social
media survey is feasible when selection bias is properly addressed by bias correction methods.
79
Social media has become one of the most important inventions in our life. It is not only a great
invention, more importantly, it is a new media channel which is changing people’s life in a way
that no traditional media has ever done. Academic research is inevitably affected by social media
as well. Social media can be very helpful for data collection and survey design. Internet surveys
often recruit through websites. However, even the most popular websites other than social media
platforms cannot attract huge Internet traffic in a short amount of time or reach users around the
world like social media does.
Future studies can study the data quality of social media surveys from additional perspectives.
The selection bias is the primary concern, and other problems include network effect, strategic
bias, mode effect, measurement error, etc.
The project described in this chapter relies on data from survey(s) administered by the
Understanding America Study, which is maintained by the Center for Economic and Social
Research (CESR) at the University of Southern California. The content of this chapter is
solely the responsibility of the author and does not necessarily represent the official views of
USC or UAS.
80
0.0 1.0
Facebook User
0
1000
2000
3000
4000
Percentage
0.32
0.68
0.0 1.0
Twitter User
0
1000
2000
3000
4000
5000
Percentage
0.80
0.20
20 40 60 80 100
Age (by Facebook Usage)
0.000
0.005
0.010
0.015
0.020

20 40 60 80 100
Age (by Twitter Usage)
0.000
0.005
0.010
0.015
0.020
0.025

Figure 2.1: Facebook Usage and Twitter Usage
These figures show Facebook and Twitter usage distributions and the age distributions of Face-
book users (orange), non-Facebook users (blue), Twitter users (orange), non-Twitter users (blue).
Facebook users are younger than non-Facebook users on average. Twitter users are younger than
non-Twitter users, and the gap in Twitter’s case is larger than the gap in Facebook’s case.
81
0 50000 100000 150000
Income
0.000000
0.000002
0.000004
0.000006
0.000008
0.000010
0.000012
0.000014

0 2 4 6 8 10
Log(No. of Family or Close Friends)
0.0
0.1
0.2
0.3
0.4
0.5
0.6

0 2 4 6 8
Log(Willingness to Pay for Vaccine)
0.0
0.2
0.4
0.6
0.8
1.0
1.2
1.4

Figure 2.2: Distributions: Income, Number of Family/Close Friends, and WTP for Vaccine
This paper examines how different methods match those who use social media and who do not use social media on three variables. These
variables are income (which is the average household income of the income group which respondents belong to), number of family and
close friends, and willingness to pay for COVID-19 vaccine. Willingness to Pay for Vaccine comes from the question “Suppose that a
100% safe and effective coronavirus vaccine is developed today but insurance does not cover it. How much would you be willing to
pay to get yourself vaccinated?” This figure shows distributions of these three variables. Distributions of the number of family and close
friends variable and willingness to pay for vaccine variable are skewed, so I the natural logarithm of the two variables is shown instead.
These distributions roughly follow the normal distribution.
82
0 50000 100000 150000
Income
0.000000
0.000002
0.000004
0.000006
0.000008
0.000010
0.000012
0.000014
0.000016

0 2 4 6 8 10
Log(No. of Family or Close Friends)
0.0
0.1
0.2
0.3
0.4
0.5
0.6

0 2 4 6 8
Log(Willingness to Pay for Vaccine)
0.0
0.2
0.4
0.6
0.8
1.0
1.2

Figure 2.3: Distributions: Income, Number of Family/Close Friends, and WTP for Vaccine (Facebook)
This figure shows the distributions of the three variables, income, natural logarithm of number of family and close friends, and natural
logarithm of willingness to pay for COVID-19 vaccine, for Facebook users (orange) and Facebook non-users (blue). Willingness to Pay
for Vaccine comes from the question “Suppose that a 100% safe and effective coronavirus vaccine is developed today but insurance does
not cover it. How much would you be willing to pay to get yourself vaccinated?” Facebook users and Facebook non-users follow similar
distributions in all three cases.
83
0 50000 100000 150000
Income
0.0000000
0.0000025
0.0000050
0.0000075
0.0000100
0.0000125
0.0000150
0.0000175
0.0000200

0 2 4 6 8 10
Log(No. of Family or Close Friends)
0.0
0.1
0.2
0.3
0.4
0.5
0.6

0 2 4 6 8
Log(Willingness to Pay for Vaccine)
0.0
0.2
0.4
0.6
0.8
1.0
1.2
1.4

Figure 2.4: Distributions: Income, Number of Family/Close Friends, and WTP for Vaccine (Twitter)
This figure shows the distributions of the three variables, income, natural logarithm of number of family and close friends, and natural
logarithm of willingness to pay for COVID-19 vaccine, for Twitter users (orange) and Twitter non-users (blue). Willingness to Pay for
Vaccine comes from the question “Suppose that a 100% safe and effective coronavirus vaccine is developed today but insurance does
not cover it. How much would you be willing to pay to get yourself vaccinated?” Twitter users and Twitter non-users follow similar
distributions in all three cases.
84
1 Married (spouse lives with me) 2 Married (spouse lives elsewhere) 3 Separated 4 Divorced 5 Widowed 6 Never married
Marital Status
0
500
1000
1500
2000
2500
3000
3500
Percentage
0.54
0.01
0.02
0.14
0.05
0.24
1 Less than 1st grade
2 Up to 4th grade
3 5th or 6th grade
4 7th or 8th grade
5 9th grade
6 10th grade
7 11th grade
8 12th grade-no diploma
9 High school graduate or GED
10 Some college-no degree
11 Assoc. college degree-occ/voc prog
12 Assoc. college degree-academic prog
13 Bachelor's degree
14 Master's degree
15 Professional school degree
16 Doctorate degree
Education
0
200
400
600
800
1000
1200
1400
1600
Percentage
0.00 0.00 0.00
0.00
0.01 0.01
0.01
0.02
0.17
0.23
0.07
0.07
0.24
0.12
0.02 0.02
1 White Only 2 Black Only 3 American Indian or Alaska Native Only 4 Asian Only 5 Hawaiian/Pacific Islander Only 6 Mixed
Race
0
1000
2000
3000
4000
5000
Percentage
0.77
0.08
0.02
0.05
0.01
0.05
Figure 2.5: Distributions: Marital Status, Education, and Race
These figures show the distributions of marital status, education, and race of the sample. 55%
respondents are married and 45% respondents are single. 46% respondents do not have a college
level degrees (education categories 1-9), 50% respondents have some undergraduate level degree
(education categories 10-13), and 4% respondents have higher degrees (education categories 14-
16).
85
0.0 0.2 0.4 0.6 0.8 1.0
0.0
0.2
0.4
0.6
0.8
1.0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
No. of Neighbors
2
1
0
1
2
3
4
T-statistic
FB-Income-KNN
FB-F/F-KNN
FB-WTP-KNN
0.0 0.2 0.4 0.6 0.8 1.0
0.0
0.2
0.4
0.6
0.8
1.0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
No. of Neighbors
2
1
0
1
2
3
T-statistic
TW-Income-KNN
TW-F/F-KNN
TW-WTP-KNN
Figure 2.6: KNN: Number of Neighbors
These figures show the performances of K-Nearest Neighbors Matching with different number of
neighbors. The number of neighbors ranges from one to twenty. The space between red lines
indicates good performance (t-statistic between -1.96 and 1.96). The top panel shows the perfor-
mances in the case of matching Facebook users and respondents who do not use Facebook. In the
case of income, the performance is stable and good as the number of neighbors changes. KNN
with nearest one neighbor perform well when it matches number of family or close friends. KNN
with less than 6 neighbors perform well when matching willingness to pay for COVID-19 vaccine.
The bottom panel shows the performances in the case of matching Twitter users and respondents
who do not use Twitter. KNN with one, two or five neighbors perform well when matching in-
come. KNN with any number of neighbors perform well when it matches number of family or
close friends or willingness to pay for COVID-19 vaccine.
86
0.0 0.2 0.4 0.6 0.8 1.0
0.0
0.2
0.4
0.6
0.8
1.0
0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0
Radius Caliper
4
2
0
2
4
6
T-statistic
FB-Income-Radius
FB-F/F-Radius
FB-WTP-Radius
0.0 0.2 0.4 0.6 0.8 1.0
0.0
0.2
0.4
0.6
0.8
1.0
0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0
Radius Caliper
2
0
2
4
6
8
T-statistic
TW-Income-Radius
TW-F/F-Radius
TW-WTP-Radius
Figure 2.7: Radius Caliper
These figures show the performances of Radius Matching with different caliper c. The radius
caliper ranges from 0.02 to 1 with step size 0.02. The space between red lines indicates good
performance (t-statistic between -1.96 and 1.96). The top panel shows the performances in the case
of matching Facebook users and respondents who do not use Facebook. When Radius matching
matches income, the performance is good if radius c 0:22. In the cases of matching number of
family or close friends and willingness to pay for COVID-19 vaccine, Radius Matching does not
perform well. The bottom panel shows the performances in the case of matching Twitter users
and respondents who do not use Twitter. Radius Matching does not perform well in the case of
matching income. Radius Matching with radius caliper c 0:26 performs well when it matches
number of family or close friends. In the case of matching willingness to pay for COVID-19
vaccine, Radius Matching with any radius caliper c performs well.
87
Table 2.1: Bias Types
Policy Related Not Policy Related
Pivotal Strategic Bias ø
Not Pivotal Cheap Talk Standard Bias
88
Table 2.2: Summary Statistics
(1) (2) (3) (4) (5) (6)
Non-User User Non-User User
(Facebook) (Facebook) T-Stat (Twitter) (Twitter) T-Stat
Variable Mean/SE Mean/SE (2)-(1) Mean/SE Mean/SE (5)-(4)
Income 66748 70791 3.40*** 67086 79378 8.82***
(997) (668) (609) (1306)
Log(Number of Family/Close Friends) 2.72 2.89 5.95*** 2.85 2.79 -2.02**
(0.02) (0.01) (0.01) (0.03)
Log(WTP Vaccine) 4.78 4.62 -3.80*** 4.66 4.74 1.74*
(0.03) (0.02) (0.02) (0.04)
Use Facebook 0 1 N/A 0.63 0.83 14.51***
0 (0) (0.01) (0.01)
Use Twitter 0.09 0.24 14.51*** 0 1 N/A
(0.01) (0.01) (0) (0)
Gender - Male 0.52 0.36 -12.77*** 0.41 0.43 1.32
(0.01) (0.01) (0.01) (0.01)
Age 52.51 48.71 -9.15*** 51.53 43.36 -16.94***
(0.36) (0.22) (0.21) (0.40)
College 0.74 0.79 4.73*** 0.76 0.84 6.79***
(0.01) (0.01) (0.01) (0.01)
Married 0.54 0.55 0.68 0.55 0.52 -2.25**
(0.01) (0.01) (0.01) (0.01)
White 0.78 0.84 6.17*** 0.82 0.82 0.3
(0.01) (0.01) (0.01) (0.01)
Black 0.12 0.08 -4.33*** 0.09 0.10 0.21
(0.01) (0.00) (0.00) (0.01)
American Indian or Alaska Native 0.06 0.05 -1.94* 0.05 0.05 -0.01
(0.01) (0.00) (0.00) (0.01)
Asian 0.07 0.06 -2.52** 0.06 0.09 3.88***
(0.01) (0.00) (0.00) (0.01)
Observations 2244 4688 6932 5580 1352 6932
* p< 0:10, ** p< 0:05, *** p< 0:01. This table summarizes mean and standard deviations of key variables in each social media user
group. WTP vaccine (Willingness to Pay for Vaccine) comes from the question “Suppose that a 100% safe and effective coronavirus
vaccine is developed today but insurance does not cover it. How much would you be willing to pay to get yourself vaccinated?”
89
Table 2.3: Methods Performance (T Statistic)- Facebook
Methods Income Log(No. of Family or Close Friends) Log(Willingness to Pay for Vaccine)
KNN(1) 0.75 1.32 -1
KNN(5) 0.93 3.70*** -1.92*
Radius(0.03) 1 4.46*** -2.57***
Radius(0.1) 1.33 4.83*** -3.04***
Mahalanobis 1.24 3.11*** -1.77*
IPW 3.42*** 6.15*** 7.98***
* p< 0:10, ** p< 0:05, *** p< 0:01. This table shows the performances of K-Nearest Neighbors
Matching (KNN), Radius Matching, Mahalanobis Matching, and Inverse Probability Weighting
(IPW) on matching Facebook users with respondents who do not use Facebook. All the matching
methods perform well when matching on household income (the first column) and the difference
is not different from zero. The weighting method does not perform well in this case. K-Nearest
Neighbors Matching with small number of neighbors perform well when matching on number of
family or close friends (the second column), whereas other methods do not perform well in this
case. Similarly, only K-Nearest Neighbors Matching with small number of neighbors perform well
when matching on willingness to pay for COVID-19 vaccine (the last column). Willingness to Pay
for Vaccine comes from the question “Suppose that a 100% safe and effective coronavirus vaccine
is developed today but insurance does not cover it. How much would you be willing to pay to get
yourself vaccinated?”
90
Table 2.4: Methods Performance (T Statistic)- Twitter
Methods Income Log(No. of Family or Close Friends) Log(Willingness to Pay for Vaccine)
KNN(1) 0.69 -0.01 -0.29
KNN(5) 1.92* -0.92 0.83
Radius(0.03) 4.02*** 0.34 0.70
Radius(0.1) 4.48*** -0.29 0.97
Mahalanobis 1.61 -0.73 1.38
IPW 1.04 1.19 1.17
* p< 0:10, ** p< 0:05, *** p< 0:01. This table shows the performances of K-Nearest Neigh-
bors Matching (KNN), Radius Matching, Mahalanobis Matching, and Inverse Probability Weight-
ing (IPW) on matching Twitter users with respondents who do not use Twitter. The first column
shows the t-statistic of the income difference between Twitter users and their counterparts after
matching or weighting. K-Nearest Neighbors Matching with small number of neighbors, Maha-
lanobis Matching and IPW perform well in this case, whereas Radius Matching or KNN with a
large number of neighbors do not perform well. The second and third column show the t-statistic
of the number of family or close friends difference and willingness to pay for COVID-19 vaccine
difference between Twitter users and their counterparts after matching or weighting, respectively.
All matching and weighting methods perform well in these two cases. Willingness to Pay for
Vaccine comes from the question “Suppose that a 100% safe and effective coronavirus vaccine is
developed today but insurance does not cover it. How much would you be willing to pay to get
yourself vaccinated?”
91
Chapter 3
School Starting Age Effect and Noncompliance Behavior:
Evidence from China
with Yinan Liu
3.1 Introduction
Primary school education is usually the first time children get formal curriculum education.
The starting age, graduating age, length of primary education, curriculum design, and many other
factors will affect the return to primary school education. Most countries have formal laws or
policies to indicate the school starting age. The prevalent starting ages are from five to seven.
For example, school starting age is five in England and Netherlands; school starting age is six in
Denmark, France, Germany, Hungary, Norway, Poland, Portugal, Spain, Turkey, etc (Whitbread,
2014). In the case of China, Compulsory Education Law of the People’s Republic of China (Ar-
ticle 5) requires children who have reached the age of six to enroll in primary school and receive
compulsory education for the prescribed number of years.
1
Studies about the effect of (primary) school starting age on short-run or long-run outcomes in
developed countries show mixed findings. The existing evidence that starting school later is better
for children is more compelling because of the following reasons. First, there is strong evidence
1
The law allows the schooling to be postponed to the age of seven in areas where starting at six is not possible.
92
which supports this claim. The evidence does not only come from economic literature, but also
from psychological literature. The age effect may not be very large in later stage of children’s
development, but at earlier ages like around age five or six, children’s cognitive and noncognitive
skills develop rapidly. Therefore, holding one more year before exposing children to primary
school studying materials can make a huge difference.
Second, “later is better” is a popular opinion among parents. The relative age effect (RAE)
is a well-known effect in education and sports, which is a bias where participation is higher or
performance is better amongst those born early in the cohort. Not only does relative age effect exist
in literature, it is also consistent with people’s belief. The Red-shirting behavior lends support on
this point. Red-shirting or “the gift of time” originally describes the practice of holding college
athletes out of play until they have grown stronger. Later these terms also describe the practice of
allowing children to wait for another year before starting schooling so they can benefit more from
the schooling. Red-shirting behavior is commonly implemented by parents or coaches. In the US,
upper-income, white, highly-educated parents red-shirt their children at the highest rate (Cook and
Kang, 2018). Boys are more commonly red-shirted than girls.
Despite the empirical evidence and the common belief “later is better,” a nontrivial proportion
of children went to school earlier than they are officially eligible for primary schooling. Starting
primary school earlier than six years old is common in China in particular. This noncompliance
behavior is largely driven by children who were born right after the cut-off date, August 31, which
indicates children who reach six after this date in the current year should start primary school
the next year (Zhang and Xie, 2017). These observations raise the question why noncompliance
behaviors in China and in developed countries are opposite.
In this chapter, we focus on this particular noncompliance behavior of Chinese households and
cover the following topics. First, we compare respondents’ household characteristics among three
groups: respondents who start primary school at the suggested age, respondents who start primary
school earlier than six years old, and respondents who start primary school later than six years
old. Second, we examine the effects of school starting age on short-run and long-run outcomes
93
when the school starting age rule just changed from seven to six. Third, we examine the effects of
school starting age on test scores and cognitive skills for children who started primary school from
2003 to 2008, which were more than 17 years after the school starting age rule changed. Fourth,
we study the effect of relative age. Fifth, we study the average effect of this age rule. Finally, we
discuss potential explanations behind the behavior of “counter-red-shirting.”
The first related literature discusses the effects of school starting age. Effects can be short-run
such as test scores in primary school, drop-off rate, grade repeating, skipping grade, etc. Effects
can also be long-run such as earnings and health in adulthood. Evidence supports young mem-
bers in the cohort have worse outcomes than old members. For example, youngest members had
lower scores in both grade four and eight across OECD countries, and the probability of attending
university was also lower for younger members in the US and Canada (Bedard and Dhuey, 2006);
younger members in the same cohort are less likely to enter good middle schools, boys are affected
more than girls, and the effect cannot be mitigated by family social-economic status in China (Li
et al., 2012); postponing one year entering school will reduce the probability of juvenile crime for
young black females, and age-related human capital accumulation can be a potential explanation
(Depew and Eren, 2016); one additional month of relative age decreases the likelihood of receiv-
ing special education services (especially for learning disability) by 2–5% (Dhuey and Lipscomb,
2010). Literature also provides evidence that late school starting age does not improve children’s
life later. For example, a natural experiment of Norweigian Army found that school starting age did
not matter for IQ and late school starting age decreased income in adulthood up to age 30 (Black
et al., 2011); primary school starting age in the US was increasing gradually and this increase also
increases inequality in human capital and social welfare (Deming and Dynarski, 2008).
A few studies which focus on the effect of kindergarten starting age provide similar evidence
and potential explanations for relative age effect. For example, NICHD Early Child Care Research
Network found that children entering later perform better than children entering earlier in most tests
examined except for one Woodcock-Johnson-Letter-Word recognition sub-test (Network et al.,
2007); children entering kindergarten later will have higher test scores and this positive relationship
94
is mainly due to skills accumulated before kindergarten instead of learning from older classmates,
and having older classmates will increase test scores but also increase the probability of grade
repetition and diagnosis of learning disability such as ADHD (Elder and Lubotsky, 2009).
Another related literature studies the effects of changes to compulsory education policy. For
example, low compliance rate of the cut-off date and the relative age effect cannot be fully ex-
plained by school starting age (Zhang and Xie, 2017); the change in Chinese education policy that
extended the primary school length from five to six years (with curriculum unchanged) had little
effect on long-run educational outcomes and a small effect on average monthly income (Eble and
F. Hu, 2019); reducing the cost of education improved education access without damaging school
performances during a Kenya Free Primary Education Program (Lucas and Mbiti, 2012).
Literature on the cost of child care and labor supply of parents is helpful to explain an early
school starting age. For example, introducing of a new child care program in Quebec will shift
maternal labor supply and affect children’s welfare measured by motor skill, social skills, and
illness (Baker et al., 2008).
The last literature takes a dynamic perspective and study household choices of children’s educa-
tion taking into account the later life outcomes of parents and/or children. For example, reduction
in the cost of children and reduction in wage-gender gap account for the observed increasing labor
participation in early life-cycle with American women’s life-cycle labor supply data (Attanasio et
al., 2008); upward spillovers of human capital from children to parents exist in China that children
with more years of schooling will lead to better cognitive functions, lung functions, and higher
subjective survival expectation for parents through social support, access to resources, parental
labor supply and psychological channels (Ma, 2017).
The remainder of the chapter is organized as follows. Section 3.2 discusses the background of
Compulsory Education Law of the People’s Republic of China and relative age effect. Section 3.3
summarizes data sources and the main variables of interest. Section 3.4 discusses the empirical
models. Section 3.5 discusses the empirical results, including noncompliance households’ charac-
teristics, estimations of school starting age effect, relative age effect, and the effect of this school
95
starting age rule implementation. 3.6 discusses five potential explanations of early school starting
age. Section 3.7 provides the conclusion.
3.2 Background
3.2.1 School Starting Age in Compulsory Education Law of the People’s
Republic of China, Article 5
The main official law guiding the education system in China is Compulsory Education Law of
the People’s Republic of China, which went into effect on July 1, 1986. Article 5 indicates the ap-
propriate primary school starting age should be six, and the common starting age was seven before
1986. However, the “school starting age rule” was not implemented immediately in all places in
China. Education Administration allowed provinces to gradually adapt to this rule, especially in
the areas which had limited resources. It is not until 2000 that all provinces in China adopted this
rule completely. The official cut-off date for age calculation is August 31, which means children
who reach six before August 31 in the current year should start primary school in this fall, and
children who reach six after August 31 in the current year should start primary school in the next
fall semester.
2
3.2.2 Relative Age Effect
Relative Age Effect (RAE) is also known as birth date effect, month of birth bias, and season
of birth bias. It is a bias where participation is higher or performance is better amongst those born
early in the relevant selection period than would be expected from the normalised distribution of
live births. RAE exists in age related systems which establish the eligibility of inclusion based on
specific cut-off dates. For example, in academic cohort structuring, August or September cut-off
dates are common in the Northern Hemisphere; February or March cut-off dates are common in
2
Starting from 2017, different provinces are allowed to set up their own cut-off dates.
96
the Southern Hemisphere based on autumn start dates. The use of autumn start dates as school
starting dates reflects the historical need for children to be involved in summer-time agricultural
work with school starting after harvesting. Nowadays, cut-off dates together with other academic
structuring rules are written and determined by national education authorities. In some countries,
states or provinces may also be allowed to make their own rules.
3.3 Data
3.3.1 China Family Panel Studies (CFPS)
The first data source in this chapter is China Family Panel Studies or CFPS (Xie and J. Hu,
2014), which is a nearly nationwide, comprehensive, longitudinal social survey.
3
The samples
come from 25 provinces (out of 34 provinces in total) and represent 94.5% of the total population
in Mainland China. The baseline survey in 2010 includes 14960 households (42590 individuals).
2010 survey receives responses from 81.3% households (84.1% individuals). Three follow-up
surveys are in 2012, 2014, and 2016. China Family Panel Studies includes four types of ques-
tionnaires designed specifically for adults, children, families, and communities. We use data from
adult questionnaire in this chapter.
The dataset we compiled from CFPS includes the following variables: school starting age;
demographic information such as gender, birth weight (in 500g), month of birth, number of sib-
lings, highest education achieved, income, and Hukou; household information such as education
expenditure in last year, father’s education, and mother’s education.
4
3.3.2 China Education Panel Survey (CEPS)
The second data source in this chapter is China Education Panel Survey or CEPS (NSRC,
2013), which is a large-scale, nationally representative, longitudinal survey.
5
CEPS started with
3
https://www.isss.pku.edu.cn/cfps/index.htm?CSRFT=23E8-750Q-UXSX-8GRB-47D2-V5GX-XAVB-7MVI
4
Education year is education variable 3 approximately.
5
http://ceps.ruc.edu.cn/index.htm
97
two cohorts – the 7th and 9th grade students in the 2013-2014 academic year. It includes approx-
imately 20,000 students in 438 classrooms of 112 schools in 28 county-level units in mainland
China. The follow-up survey is in the 2014-2015 academic year.
CEPS provides the following variables: school starting age; demographic information such as
gender, birthweight (in 500g), month of birth, number of siblings, and Hukou; household informa-
tion such as financial conditions, father’s education, and mother’s education. Since CEPS samples
clustered in classes, we also have data on relative age of each respondent in his or her own class.
3.3.3 Compliance With “Age Rule” by Month of Birth: CFPS and CEPS
Figure 3.1 shows the proportions of respondents who started primary school before six years
old (early group), at six years old (right age group), and after six years old (late group) by their
month of birth in CFPS. CFPS dataset includes many respondents who started primary school both
right before and right after the “school starting age rule” in Article 5 was implemented, and thus a
large proportion of respondents did not follow this rule immediately and drove the proportions of
late group up. The late group contains about 60% to 70% of people who were born from January to
August, and rapidly decreases to 38% for people who were born in September and increases slowly
in later months. The trends for early group and right age group are the opposite of late group.
Right age group proportion decreases slowly from 28% to 19% between January and August, and
increases to 40% for September group and does not change much in later months. Early group
proportion is around 10% between January and August, and increases to 21% in September and
slowly decreases in later months.
Figure 3.2 shows the proportions of respondents who started primary school before six years
old (early group), at six years old (right age group), and after six years old (late group) by their
month of birth in CEPS. Respondents in CEPS started primary school mostly from 2003 to 2008,
and thus the age six is the prevalent primary school starting age for them. Right age group pro-
portion is around 70% from January to July, decreases to 63% in August and 47% in September.
It increases rapidly to 65% from October to December. Similarly, late group proportion increases
98
slowly from 19% to 34% between January and August, and decreases rapidly to less than 10% in
September and does not change in later months. The trend of early group is different from the
other two groups. Early group proportion decreases slowly from 13% to 3% between January and
August, and increases rapidly to 44% in September. It then decreases to 25% in later months.
Figure 3.1 and Figure 3.2 both show sudden changes from August to September in all groups,
where late group proportion always decreases and early group proportion always increases. More
CFPS respondents experienced the change of common school starting age from seven to six, when
the age six was already the most common school starting age for respondents in CEPS, which
can also be observed from these two figures. In Figure 3.1, the proportions ranked from highest
to lowest are late group, right age group, and early group. Figure 3.2 shows the right age group
has the largest proportion and early group and late group proportions are approximately the same.
6
Such large proportions of early group from September to December raise the question why sending
children to primary school earlier than six years old is popular in China.
3.4 Empirical Strategy
This chapter covers (1) comparing respondents’ household characteristics among three groups:
right age group, early group, and late group; (2) examining the effects of school starting age on
highest education and income when the school starting age rule just changed from seven to six with
CFPS data; (3) examining the effects of school starting age on test scores and cognitive skills with
CEPS data, which includes children who started primary school from 2003 to 2008; (4) studying
the effect of relative age; (5) studying the average effect of the “school starting age rule.”
To examine which characteristics are different for right age group, early group, and late group,
we compare the means of individual’s characteristics and household’s characteristics among these
three groups. We also use Multinomial Ordered Logit Model to examine how covariates are corre-
lated with group choices.
6
More respondents were born in later months of the year.
99
To examine the effects of school starting age on a set of outcome variables (income, highest
education, test scores, and cognitive skills), we use the following two regression models:
Y
i
=t
SSA
SSA
i
+b
1
X
i
+d
1 j
+e
1i
(3.1)
Y
i
=t
Early
Early
i
+t
Late
Late
i
+b
2
X
i
+d
2 j
+e
2i
(3.2)
in which Y
i
is individual i’s outcome variable from CFPS (income and highest education) or CEPS
(test scores and cognitive skills); SSA
i
is primary school starting age of i; Early
i
and Late
i
are
dummy variables indicating if i is in early group, late group, or right age group; X
i
is a vector of
covariates, d
i
is region fixed-effects, and e
i
is i.i.d. error term. The parameters of interest include
t
SSA
,t
Early
, andt
Late
.
To examine the effect of relative age of individual i in class, we use CEPS data and follow these
two regression models:
Y
i
=t
RA
RA
i
+b
3
X
i
+d
3 j
+e
3i
(3.3)
Y
i
=t
Older
Older
i
+b
4
X
i
+d
4 j
+e
4i
(3.4)
in which Y
i
is individual i’s outcome variable from CEPS (test scores and cognitive skills); RA
i
is
relative age of i in class;
7
Older
i
is a dummy variable indicating if i’s age is above the median age
of the class; X
i
is a vector of covariates, d
i
is region fixed-effects, and e
i
is i.i.d. error term. The
parameters of interest includet
RA
andt
Older
.
We use CFPS data to examine the average effect of “school starting age rule” change of age
seven to age six. Different province implemented the School Starting Age in Compulsory Educa-
tion Law of the People’s Republic of China in different years (Ma, 2017). We impute a dummy
variable indicating if the respondent started primary school before or after the law implemented in
7
Relative age is the difference between median birth date and i’s birth date in months.
100
his or her province. We then estimate the association between this variable and outcome variables
(income and highest education).
3.5 Empirical Results
3.5.1 Group Characteristics
Table 3.1 shows the differences in group characteristics among right age group, early group,
and late group from CFPS data. Right age group has the best social economic status and re-
sources: this group has the largest birth weight, education expenditure, highest education level,
highest mother’s education level, highest father’s education level, and highest urban Hukou per-
centage. Early group has the second best social economic status and resources and late group has
the worst social economic status and resources. The first column in Table 3.3 shows the result
from Multinomial Ordered Logit model with the same data.
8
This column shows school starting
age has positive association with the number of siblings and negative association with the highest
education level and mother’s education level.
Table 3.2 shows the differences in group characteristics among right age group, early group,
and late group from CEPS data. Samples from CEPS shared similar patterns with samples from
CFPS that right age group, early group, and late group are ranked from best social economic status
and resources (birth weight, number of siblings, financial conditions, father’s education, mother’s
education, and urban Hukou) to the worst. The second column in Table 3.3 shows the result from
Multinomial Ordered Logit model with the CEPS data. This column shows school starting age
has positive association with being male and negative association with mother’s education and
financial conditions.
8
Early group is coded -1, right age group is coded 0, and late group is coded 1 in the Multinomial Ordered Logit
model.
101
3.5.2 School Starting Age Effect
Table 3.4 shows the association between primary school starting age and two outcome variables
(income and highest education level) in CFPS data. The first two columns show the respondents
who started primary school late have lower income and education level. They are the empirical
estimations of Model 3.1. The coefficients are negative and significant in both cases. The third
and fourth columns show the empirical estimations of Model 3.2 in which being in late group is
negatively associated with the education level.
Table 3.5 shows the association between primary school starting age and five outcome variables
in CEPS data. The outcome variables are raw cognitive skills index, standardized cognitive skills
index (see Appendix D), Chinese test score, Mathematics test score, and English test score.
9
This
table is the empirical estimation of Model 3.1 and it shows a late school starting age is negatively
associated with cognitive skills, Chinese test score, and Mathematics test score. Table 3.6 shows
the empirical estimations of Model 3.2 with CEPS data. Being in late group is negatively associated
with cognitive skills and Mathematics test score in this case. Table 3.7 shows no significant local
effect of being born between September and December versus being born in early months with a
regression discontinuity model.
These results provide empirical evidence that starting primary school early is positively asso-
ciated with both short-run outcomes (such as test scores, cognitive skills, education) and long-run
outcomes (such as income), which implies that either households with more resources prefer start-
ing school early or starting school early improves these outcomes. In either case, early school
starting age is an rational choice from parents’ perspective in China.
3.5.3 Relative Age Effect
Table 3.8 shows the effect of relative age in class on five outcomes variables: raw cognitive
skills index, standardized cognitive skills index (see Appendix D), Chinese test score, Mathematics
test score, and English test score. It is the empirical estimation of Model 3.3. Assuming classmates’
9
Chinese, Mathematics, and English are the three main courses for primary school education in China.
102
age distribution does not change, being one month older would decrease standardized cognitive
skill score by 0.01, Chinese test score by 0.08, Mathematics test score by 0.11, and English test
score by 0.1. Table 3.9 shows the empirical estimation of Model 3.4, in which we study the impact
of being in the older half of the class cohort. Results in this table are consistent with the results in
Table 3.8: being an old member in class cohort is negatively associated with cognitive skills and
test scores. This set of empirical results also provide a potential explanation why Chinese parents
may choose to send children to school before eligible age.
3.5.4 Effect of “School Starting Age Rule” Implementation
Table 3.10 shows the average effect estimation of “school starting age rule” change from age
seven to age six. Individuals who were affected by this change had higher education level but
lower income, and males were more negatively affected than females. Decreasing school starting
age increased the highest education level as expected. The effect of implementation of “age rule”
on income is more complicated and requires further research.
3.6 Potential Explanations
There are five potential explanations why Chinese parents have such “counter-red-shirting”
behavior. First, parents try to go back to labor market early. This mechanism relates to the child
care cost and female labor supply. Parents may need to quit labor market and stay at home with
the young child before sending him or her to school. They have the choice of sending the child to
primary school one year before the eligible age so they can return to labor market early.
10
The second explanation also relates to labor market supply. School starting age younger than
six can be the result of hoping the child enter labor market early. the child would start working one
year earlier if he or she enters primary school one year before being eligible.
10
If parents send the child to kindergarten one year earlier in order to go back to labor market, the child’s school
starting age also decreases one year.
103
Third, the age of classmates or relative age in cohort may be an important factor both in par-
ents’ decision process and in determining schooling returns. On one hand, being with younger
classmates may give the child advantages in competition, which may then lead to better resources,
more confidence and better mental health. In addition, older children in the same cohort are more
likely to have experiences as team leaders, which can be beneficial for social skills development.
On the other hand, the young child receives positive spillover from older members of the cohort,
and the spillover can throw the child into a faster development trajectory.
The fourth potential explanation is the flexibility in repeating grade. If the child enters school
one year before the eligible age and later needs to repeat a grade, he or she can still graduate with
other children of the same age.
The last explanation is social norm. Sending children to school before they reach age seven is
very common in Chinese societies, then sending children at the right age or later may be seen as
a signal that my children are not smart. In addition to working through the “signaling” channel,
social pressure can also affect parents’ decisions because parents are typically following what other
parents are doing. In some places in China, it results in starting school one year before the eligible
age, especially for children who were born after September. Therefore, although official eligible
age is six, the social norm is age five.
We need future works to explore other potential explanations and empirically test which ones
are valid among these potential explanations.
3.7 Conclusion
This chapter is motivated by an interesting observation in China: a large proportion of children
enter primary school before age six (even though Compulsory Education Law of the People’s
Republic of China requires children to reach age six). We examine household characteristics of
households who sent children to primary school before age six (early group), at age six (right age
group), and after age six (late group), and find right age group has the best social-economic status
104
and resources, followed by early group, and lastly late group. We also find large school starting
age is negatively associated with outcomes such as test scores, cognitive skills, highest education
level, and income. The estimation of relative effect provides similar evidence that being a young
member in class is positively associated with cognitive skills and test scores. These empirical
findings provide evidence that Chinese parents who choose to send children to primary school
before age six may have made the rational decision. Following the empirical findings, we discuss
five potential explanations of this interesting decision, which include labor supply, relative age
effect, flexibility, and social norm.
105
1 2 3 4 5 6 7 8 9 10 11 12
Month of Birth
0.1
0.2
0.3
0.4
0.5
0.6
0.7
Proportions
0.13
0.09
0.1
0.09
0.07
0.1
0.09
0.08
0.21
0.22
0.19
0.17
0.28
0.27
0.23
0.27
0.22 0.22
0.19 0.19
0.4
0.39
0.37 0.37
0.59
0.63
0.66
0.65
0.72
0.68
0.71
0.73
0.38
0.39
0.44
0.46
CFPS-Early
CFPS-Right Age
CFPS-Late
Figure 3.1: Right Age, Early, and Late Group Proportions by Month of Birth (CFPS)
This figure shows the proportions of respondents who started primary school before six years old
(early group), at six years old (right age group), and after six years old (late group) by their month
of birth in CFPS. CFPS dataset includes many respondents who started primary school both right
before and right after the “school starting age rule” in Article 5 was implemented. Therefore, a
large proportion of respondents did not follow this rule immediately and drove the proportions of
late group up. Rapid changes are present in all three groups from August to September where the
cut-off date is.
106
1 2 3 4 5 6 7 8 9 10 11 12
Month of Birth
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
Proportions
0.13
0.09
0.07
0.06
0.03
0.04
0.03 0.03
0.44
0.35
0.3
0.25
0.68
0.71
0.72
0.71
0.72
0.68
0.67
0.63
0.47
0.56
0.61
0.65
0.19
0.2
0.21
0.23
0.25
0.28
0.3
0.34
0.088
0.093
0.089
0.094
CEPS-Early
CEPS-Right Age
CEPS-Late
Figure 3.2: Right Age, Early, and Late Group Proportions by Month of Birth (CEPS)
This figure shows the proportions of respondents who started primary school before six years old
(early group), at six years old (right age group), and after six years old (late group) by their month
of birth in CEPS. Respondents in CEPS started primary school mostly from 2003 to 2008, and thus
the age six is the prevalent primary school starting age for them. Rapid changes are present in all
three groups from August to September where the cut-off date is. Early group proportion increases
from 3% to 44% in September.
107
Table 3.1: Group Differences (CFPS)
Right Age Early Late
Gender - Male 0.50 0.52 0.48
(0.50) (0.50) (0.50)
Birthweight (500g) 6.37 6.17 6.23
(1.05) (1.05) (1.15)
Siblings 0.91 1.05 1.46
(0.94) (1.01) (1.17)
Highest Education 3.65 3.58 3.25
(1.11) (1.02) (1.02)
Education Expenditure Last Year 2029 1670 1114
(4319) (4477) (3821)
Father’s Education 2.89 2.87 2.57
(1.13) (1.10) (1.09)
Mother’s Education 2.53 2.37 2.01
(1.16) (1.11) (1.04)
Income 11001 11732 11500
(23992) (22064) (33697)
Hukou - Urban 0.41 0.34 0.21
(0.49) (0.47) (0.41)
Observations 1796 828 14376
* p< 0:10, ** p< 0:05, *** p< 0:01. This table shows means and standard deviations of house-
hold characteristics for three groups: right age group, early group, and late group. The respon-
dents in CFPS started primary school after 1984. Education year is education variable 3 ap-
proximately. Right age group has the best social economic status and resources: this group has
the largest birthweight, education expenditure, highest education level, highest mother’s educa-
tion level, highest father’s education level, and highest urban Hukou percentage. The early group
has the second best social economic status and resources and the late group has the worst social
economic status and resources.
108
Table 3.2: Group Differences (CEPS)
Right Age Early Late
Gender - Male 0.51 0.46 0.59
(0.50) (0.50) (0.49)
Birthweight (500g) 6.60 6.53 6.55
(1.26) (1.29) (1.34)
Siblings 1.28 1.32 1.47
(0.62) (0.62) (0.78)
Financial Conditions 3.16 3.08 2.94
(0.89) (0.95) (1.05)
Father’s Education 4.42 4.02 3.53
(2.04) (1.90) (1.70)
Mother’s Education 4.06 3.65 3.10
(2.01) (1.86) (1.69)
Hukou - Urban 0.51 0.38 0.34
(0.50) (0.49) (0.47)
Observations 12426 3109 3951
* p< 0:10, ** p< 0:05, *** p< 0:01. This table shows means and standard deviations of house-
hold characteristics for three groups: right age group, early group, and late group. The CEPS
respondents started primary school between 2003 and 2008. Samples from CEPS shared simi-
lar pattern with the sample from CFPS: right age group, early group, and late group are ranked
from the best social economic status and resources (birth weight, number of siblings, financial
conditions, father’s education, mother’s education, and urban Hukou) to the worst.
109
Table 3.3: Group Differences-Multinomial Ordered Logit
(CFPS) (CEPS)
Gender - Male -0.10 0.39***
(0.86) (0.10)
Birthweight (500g) 0.04 -0.01
(0.04) (0.03)
Siblings 0.34*** 0.05
(0.05) (0.10)
Highest Education -0.11*
(0.05)
Education Expenditure Last Year -0.000014
(0.000088)
Father’s Education -0.03 -0.01
(0.04) (0.03)
Mother’s Education -0.14** -0.07*
(0.05) (0.03)
Income -0.000002
(0.000003)
Hukou - Urban 0.13 0.05
(0.10) (0.11)
Financial Conditions -0.22*
(0.09)
* p < 0:10, ** p < 0:05, *** p < 0:01. This table shows the result from Multinomial Ordered
Logit model. The dependent variable is the group category when the early group is coded -1, the
right age group is coded 0, and the late group is coded 1. The first column shows the result from
CFPS data. This column shows school starting age has positive association with the number of
siblings and negative association with the highest education level and mother’s education level.
The second column shows the result from the CEPS data. This column shows school starting age
has positive association with being male and negative association with mother’s education and
financial conditions.
110
Table 3.4: School Starting Age Effect: CFPS
Log(Income) Highest Education Log(Income) Highest Education
School Starting Age -0.06** -0.07***
(0.02) (0.01)
Early -0.12 -0.08
(0.18) (0.08)
Late -0.11 -0.12*
(0.12) (0.06)
Covariates X X X X
County FE X X X X
Observations 1614 2634 1614 2634
Standard errors in parentheses. * p < 0:10, ** p < 0:05, *** p < 0:01. This table shows the
association between primary school starting age and two outcome variables in the CFPS data. The
first two columns show that respondents who started primary school late have lower income and
lower education level. They are the empirical estimations of Model 3.1. The third and fourth
columns show the empirical estimations of Model 3.2 in which being in late group is negatively
associated with the education level. Interactions of school starting age gender, early gender,
and late gender are also included in these models, but the coefficients of interaction terms are
not significant.
Table 3.5: School Starting Age Effect: CEPS
(1) (2) (3) (4) (5)
Cognitive(Raw) Cognitive(Std.) Chinese Maths English
School Starting Age -0.28*** -0.06** -0.59* -0.80*** -0.38
(0.08) (0.02) (0.24) (0.22) (0.24)
Covariates X X X X X
County FE X X X X X
Observations 4850 4850 4760 4756 4751
Adjusted R
2
0.22 0.18 0.11 0.03 0.102
Standard errors in parentheses. * p < 0:10, ** p < 0:05, *** p < 0:01. This table shows the
association between primary school starting age and five outcome variables in the CEPS data. The
outcome variables are raw cognitive skills index, standardized cognitive skills index (see Appendix
D), Chinese test score, Mathematics test score, and English test score. This table is the empirical
estimation of Model 3.1 and it shows a late school starting age is negatively associated with cog-
nitive skills, Chinese test score, and Mathematics test score. Interaction of school starting age
gender is also included in these models, but the coefficients of interaction terms are not significant.
111
Table 3.6: School Starting Age Effect: CEPS
(1) (2) (3) (4) (5)
Cognitive(Raw) Cognitive(Std.) Chinese Maths English
Early 0.12 0.04 0.19 1.02* 0.47
(0.20) (0.04) (0.43) (0.46) (0.44)
Late -0.69*** -0.13** -0.76 -1.81** -0.99
(0.20) (0.05) (0.53) (0.57) (0.56)
Gender -0.02 -0.01 -5.86*** -1.13*** -5.31***
GenderEarly -0.18 -0.06 0.66 -0.01 -0.40
GenderLate 0.06 -0.01 0.01 1.35 0.28
Covariates X X X X X
County FE X X X X X
Observations 4850 4850 4760 4756 4751
Standard errors in parentheses. * p < 0:10, ** p < 0:05, *** p < 0:01. This table shows the
empirical estimations of Model 3.2 with CEPS data. Being in late group is negatively associated
with cognitive skills and Mathematics test score in this case. Interactions of early gender and
late gender are also included in these models, but the coefficients of interaction terms are not
significant.
Table 3.7: Regression Discontinuity Estimation: CEPS
(1) (2) (3) (4) (5)
Cognitive(Raw) Cognitive(Std.) Chinese Maths English
D(Month of Birth 9) 0.20 0.04 -0.50 -1.83 -0.53
(0.53) (0.11) (1.39) (1.40) (1.39)
Covariates X X X X X
Observations 4850 4850 4760 4756 4751
Standard errors in parentheses. * p < 0:10, ** p < 0:05, *** p < 0:01. This table shows no
significant local effect of being born between September and December versus being born in early
months with a regression discontinuity model.
112
Table 3.8: Relative Starting Age Effect: CEPS
(1) (2) (3) (4) (5)
Cognitive(Raw) Cognitive(Std.) Chinese Maths English
Relative Age -0.04*** -0.01*** -0.08*** -0.11*** -0.10***
(0.01) (0.00) (0.02) (0.02) (0.02)
Covariates X X X X X
County FE X X X X X
Observations 4850 4850 4760 4756 4751
Adjusted R
2
0.22 0.18 0.11 0.03 0.11
Standard errors in parentheses. * p < 0:10, ** p < 0:05, *** p < 0:01. This table shows the
effect of relative age in class on five outcomes variables: raw cognitive skills index, standardized
cognitive skills index (see Appendix D), Chinese test score, Mathematics test score, and English
test score. It is the empirical estimation of Model 3.3. Assuming classmates’ age distribution
does not change, being one month older would decrease raw cognitive skill score, standardized
cognitive skill score and three test scores.
Table 3.9: Relative Starting Age Effect: CEPS
(1) (2) (3) (4) (5)
Cognitive(Raw) Cognitive(Std.) Chinese Maths English
D(Starting Age Above the Class -0.44*** -0.10*** -0.67* -1.01*** -0.66*
Median) (0.09) (0.02) (0.26) (0.27) (0.27)
Covariates X X X X X
County FE X X X X X
Observations 4850 4850 4760 4756 4751
Adjusted R
2
0.22 0.18 0.11 0.03 0.10
Standard errors in parentheses. * p < 0:10, ** p < 0:05, *** p < 0:01. This table shows the
empirical estimation of Model 3.4, in which we study the impact of being in the older half of the
class cohort. Results in this table are consistent with the results in Table 3.8, i.e., being an old
member in class cohort is negatively associated with cognitive skills and test scores.
113
Table 3.10: Law Implementation Effect: CFPS
Highest Education Log(Income)
D(Entering School After Law Implemented) 0.26*** -0.48***
(0.03) (0.05)
D(Entering School After Law Implemented) -0.44*** -0.23***
Gender (0.03) (0.05)
Gender 0.48*** 0.80***
Province FE X X
Observations 3360 2505
Standard errors in parentheses. * p < 0:10, ** p < 0:05, *** p < 0:01. This table shows the
average effect estimation of the change of the “school starting age rule” from age seven to age six.
Individuals who were affected by this change have higher education level but lower income, and
males were more negatively affected than females.
114
Conclusion
The social media industry is growing rapidly for good reasons. In the first chapter, I find that
social media usage decreases the level of depressive symptoms by 27% or 0.2 standard deviation.
The distributions of people’s feeling indices about using social media also support the idea that
using social media can be beneficial.
The second chapter finds that matching and weighting methods can effectively improve the rep-
resentativeness of social media users in most cases. Improving the representativeness of Facebook
users’ income is easier than improving the representativeness of social media users’ social connec-
tion and attitude towards COVID-19 vaccine. Improving the representativeness of Twitter users is
easier than improving the representativeness of Facebook users. This chapter provides evidence
that social media survey is feasible and reliable when the selection bias is properly addressed by
bias correction methods. Academic research can benefit from social media as well.
The first two chapters show that society may have underestimated the importance and contri-
bution of social media. Social media platforms may have been playing more important roles with
new available features as well as more responsible roles with improved regulation in the industry.
I use different surveys from Understanding America Study (UAS) for the first two chapters.
The data has limitations. For example, in the first chapter, there is not enough information on indi-
vidual’s Facebook usage and no consistent measure of the level of depressive symptoms (CESD)
in 2020. With richer data on social media usage and people’s opinions about social media in the
future, more accurate estimation is possible. These limitations can be alleviated in the future when
richer data becomes available.
115
Future studies can also discuss the data quality of social media surveys from additional per-
spectives, such as network effect, strategic bias, mode effect, measurement error, etc.
The third chapter tries to explore why there is a large proportion of children entered primary
school before they reach age six, when the Compulsory Education Law of the People’s Republic
of China requires children to start schooling at age six. We examine household characteristics for
different groups and find children who follow the “age rule” have the best social-economic status
and resources, followed by the children group who started schooling early. We also find large
school starting age is negatively associated with outcomes such as test scores, cognitive skills,
highest education level, and income. The estimation of relative effect provides similar evidence
that being a young member is beneficial for cognitive skills and test scores. School starting age
effect and relative age effect both provide evidence that starting school early is a valid option for
Chinese parents.
Following the empirical findings, we discuss five potential explanations why a large proportion
of households are in the early group, which include labor supply, relative age effect, flexibility,
and social norm. Future studies include exploring other potential explanations, empirically testing
which explanations are valid, estimating the latent cost of starting school early and starting school
late, and conducting the social welfare and inequality analysis of the “school starting age rule.”
116
References
Ahrens, Achim, Christian B Hansen, and Mark E Schaffer (2018). “pdslasso and ivlasso: Programs
for post-selection and post-regularization OLS or IV estimation and inference”. In: Statistical
Software Components, Boston College Department of Economics.
Allcott, Hunt, Luca Braghieri, Sarah Eichmeyer, and Matthew Gentzkow (2020). “The welfare
effects of social media”. In: American Economic Review 110.3, pp. 629–76.
Altonji, Joseph G, Todd E Elder, and Christopher R Taber (2005). “Selection on observed and
unobserved variables: Assessing the effectiveness of Catholic schools”. In: Journal of political
economy 113.1, pp. 151–184.
Attanasio, Orazio, Hamish Low, and Virginia S´ anchez-Marcos (2008). “Explaining changes in
female labor supply in a life-cycle model”. In: American Economic Review 98.4, pp. 1517–52.
Austin, Peter C (2011). “Optimal caliper widths for propensity-score matching when estimating
differences in means and differences in proportions in observational studies”. In: Pharmaceu-
tical statistics 10.2, pp. 150–161.
Baker, Michael, Jonathan Gruber, and Kevin Milligan (2008). “Universal child care, maternal labor
supply, and family well-being”. In: Journal of political Economy 116.4, pp. 709–745.
Bartik, Timothy J (1991). “Who benefits from state and local economic development policies?” In:
Beck, Aaron T and Brad A Alford (2009). Depression: Causes and treatment. University of Penn-
sylvania Press.
Beck, Cheryl Tatano (2001). “Predictors of postpartum depression: an update”. In: Nursing re-
search 50.5, pp. 275–285.
Bedard, Kelly and Elizabeth Dhuey (2006). “The persistence of early childhood maturity: Inter-
national evidence of long-run age effects”. In: The Quarterly Journal of Economics 121.4,
pp. 1437–1472.
Belmaker, RH and Galila Agam (2008). “Major depressive disorder”. In: New England Journal of
Medicine 358.1, pp. 55–68.
Bhatia, Shashi K and Subhash C Bhatia (2007). “Childhood and adolescent depression”. In: Amer-
ican family physician 75.1, pp. 73–80.
117
Biffignandi, Silvia (2011). “Internet Survey Methodology-recent trends and developments”. In:
Birnbaum, Michael H (2004). “Human research and data collection via the Internet”. In: Annu.
Rev. Psychol. 55, pp. 803–832.
Black, Sandra E, Paul J Devereux, and Kjell G Salvanes (2011). “Too young to leave the nest? The
effects of school starting age”. In: The review of economics and statistics 93.2, pp. 455–467.
Blasius, J¨ org and Maurice Brandt (2010). “Representativeness in online surveys through strati-
fied samples”. In: Bulletin of Sociological Methodology/Bulletin de M´ ethodologie Sociologique
107.1, pp. 5–21.
Brody, Debra J, Laura A Pratt, and Jeffery P Hughes (2018). Prevalence of depression among
adults aged 20 and over: United States, 2013-2016. US Department of Health and Human
Services, Centers for Disease Control and . . .
Chang, Tung-Zong Donald and Nicole V owles (2013). “Strategies for improving data reliability for
online surveys: A case study”. In: ” International Journal of Electronic Commerce Studies”
4.1, pp. 121–130.
Chen, Jenn (2020). Social media demographics to inform your brand’s strategy in 2020. https:
//sproutsocial.com/insights/new-social-media-demographics/#TW-demos.
Cole, Martin G and Nandini Dendukuri (2003). “Risk factors for depression among elderly com-
munity subjects: a systematic review and meta-analysis”. In: American journal of psychiatry
160.6, pp. 1147–1156.
Cook, Philip J and Songman Kang (2018). The school-entry-age rule affects redshirting patterns
and resulting disparities in achievement. Tech. rep. National Bureau of Economic Research.
Couper, Mick P (2000). “Web surveys: A review of issues and approaches”. In: The Public Opinion
Quarterly 64.4, pp. 464–494.
Couper, Mick P, Arie Kapteyn, Matthias Schonlau, and Joachim Winter (2007). “Noncoverage and
nonresponse in an Internet survey”. In: Social Science Research 36.1, pp. 131–148.
Dehejia, Rajeev H and Sadek Wahba (2002). “Propensity score-matching methods for nonexperi-
mental causal studies”. In: Review of Economics and statistics 84.1, pp. 151–161.
Deming, David and Susan Dynarski (2008). “The lengthening of childhood”. In: Journal of eco-
nomic perspectives 22.3, pp. 71–92.
Depew, Briggs and Ozkan Eren (2016). “Born on the wrong day? School entry age and juvenile
crime”. In: Journal of Urban Economics 96, pp. 73–90.
Dhuey, Elizabeth and Stephen Lipscomb (2010). “Disabled or young? Relative age and special
education diagnoses in schools”. In: Economics of Education Review 29.5, pp. 857–872.
118
Diamond, Alexis and Jasjeet S Sekhon (2013). “Genetic matching for estimating causal effects:
A general multivariate matching method for achieving balance in observational studies”. In:
Review of Economics and Statistics 95.3, pp. 932–945.
Dos Reis, Virgile Landeiro and Aron Culotta (2015). “Using matched samples to estimate the
effects of exercise on mental health via twitter”. In: Proceedings of the AAAI Conference on
Artificial Intelligence. V ol. 29. 1.
Duffy, Bobby, Kate Smith, George Terhanian, and John Bremer (2005). “Comparing data from
online and face-to-face surveys”. In: International Journal of Market Research 47.6, pp. 615–
639.
Duncan, Kristin B (2001). “Using propensity scores to control coverage bias in telephone surveys”.
In: Survey methodology 27, pp. 121–130.
Eble, Alex and Feng Hu (2019). “Does primary school duration matter? Evaluating the conse-
quences of a large Chinese policy experiment”. In: Economics of Education Review 70, pp. 61–
74.
Elder, Todd E and Darren H Lubotsky (2009). “Kindergarten entrance age and children’s achieve-
ment impacts of state policies, family background, and peers”. In: Journal of human Resources
44.3, pp. 641–683.
Facebook (2019). Facebook Reports Second Quarter 2019 Results. https://s21.q4cdn.com/
399680738/files/doc_financials/2019/Q2/FB-Q2-2019-Earnings-Release.pdf/.
Fischhoff, Baruch and W¨ andi Bruine De Bruin (1999). “Fifty–fifty= 50%?” In: Journal of Behav-
ioral Decision Making 12.2, pp. 149–163.
Goldsmith-Pinkham, Paul, Isaac Sorkin, and Henry Swift (2018). Bartik instruments: What, when,
why, and how. Tech. rep. National Bureau of Economic Research.
Greenacre, Zerrin Asan et al. (2016). “The importance of selection bias in internet surveys”. In:
Open Journal of Statistics 6.03, p. 397.
Greifer, Noah (2020). Matching Methods. https://cran.r-project.org/web/packages/
MatchIt/vignettes/matching-methods.html#ref-diamond2013/.
Hasin, Deborah S., Aaron L. Sarvet, Jacquelyn L. Meyers, Tulshi D. Saha, W. June Ruan, Malka
Stohl, and Bridget F. Grant (2018). “Epidemiology of Adult DSM-5 Major Depressive Disorder
and Its Specifiers in the United States”. In: JAMA Psychiatry 75.4, pp. 336–346. DOI:10.1001/
jamapsychiatry.2017.4602.
Hunt, Melissa G, Rachel Marx, Courtney Lipson, and Jordyn Young (2018). “No more FOMO:
Limiting social media decreases loneliness and depression”. In: Journal of Social and Clinical
Psychology 37.10, pp. 751–768.
119
Imai, Kosuke and Marc Ratkovic (2014). “Covariate balancing propensity score”. In: Journal of
the Royal Statistical Society: Series B: Statistical Methodology, pp. 243–263.
Jelenchick, Lauren A, Jens C Eickhoff, and Megan A Moreno (2013). ““Facebook depression?”
Social networking site use and depression in older adolescents”. In: Journal of Adolescent
Health 52.1, pp. 128–130.
Kerr, Jacqueline, Kevin Patrick, Greg Norman, Murray B Stein, Karen Calfas, Marion Zabinski,
and Athena Robinson (2008). “Randomized control trial of a behavioral intervention for over-
weight women: impact on depressive symptoms”. In: Depression and anxiety 25.7, pp. 555–
558.
Kozlov, Elissa, XinQi Dong, Amy S Kelley, and Claire K Ankuda (2020). “The epidemiology of
depressive symptoms in the last year of life”. In: Journal of the American Geriatrics Society
68.2, pp. 321–328.
LaLonde, Robert J (1986). “Evaluating the econometric evaluations of training programs with
experimental data”. In: The American economic review, pp. 604–620.
Lee, Sunghee (2004). “Statistical estimation methods in volunteer panel web surveys”. PhD thesis.
Lee, Sunghee and Richard Valliant (2009). “Estimation for volunteer panel web surveys using
propensity score adjustment and calibration adjustment”. In: Sociological Methods & Research
37.3, pp. 319–343.
Lei, Xiaoyan, Xiaoting Sun, John Strauss, Peng Zhang, and Yaohui Zhao (2014). “Depressive
symptoms and SES among the mid-aged and elderly in China: Evidence from the China Health
and Retirement Longitudinal Study national baseline”. In: Social Science & Medicine 120,
pp. 224–232.
Leuven, Edwin and Barbara Sianesi (2003). “PSMATCH2: Stata module to perform full Maha-
lanobis and propensity score matching, common support graphing, and covariate imbalance
testing”. In:
Lewinsohn, Peter M, John R Seeley, Robert E Roberts, and Nicholas B Allen (1997). “Center
for Epidemiologic Studies Depression Scale (CES-D) as a screening instrument for depression
among community-residing older adults.” In: Psychology and aging 12.2, p. 277.
Li, Hongbin, Pak Wai Liu, and Junsen Zhang (2012). “Estimating returns to education using twins
in urban China”. In: Journal of Development Economics 97.2, pp. 494–504.
Lim, Grace Y ., Wilson W. Tam, Yanxia Lu, Cyrus S. Ho, Melvyn W. Zhang, and Roger C. Ho
(2018). “Prevalence of Depression in the Community from 30 Countries between 1994 and
2014”. In: Scientific Reports 8.2861.
Loosveldt, Geert and Nathalie Sonck (2008). “An evaluation of the weighting procedures for an
online access panel survey”. In: Survey Research Methods. V ol. 2. 2, pp. 93–105.
120
Lucas, Adrienne M and Isaac M Mbiti (2012). “Access, sorting, and achievement: The short-
run effects of free primary education in Kenya”. In: American Economic Journal: Applied
Economics 4.4, pp. 226–53.
Luppa, Melanie, Claudia Sikorski, Tobias Luck, L Ehreke, Alexander Konnopka, Birgitt Wiese,
Siegfried Weyerer, H-H K¨ onig, and Steffi G Riedel-Heller (2012). “Age-and gender-specific
prevalence of depression in latest-life–systematic review and meta-analysis”. In: Journal of
affective disorders 136.3, pp. 212–221.
Ma, Mingming (2017). “Does Children’s Education Matter for Parents’ Health and Cognition in
Old Age? Evidence from China”. In: Evidence from China (November 7, 2017).
McDool, Emily, Philip Powell, Jennifer Roberts, and Karl Taylor (2016). “Social media use and
children’s wellbeing”. In:
Mhaolain, Aine M Ni, Chie Wei Fan, Roman Romero-Ortuno, Lisa Cogan, Clodagh Cunningham,
Rose-Anne Kenny, and Brian Lawlor (2012). “Frailty, depression, and anxiety in later life”. In:
International Psychogeriatrics 24.8, p. 1265.
Muscatell, Keely A, George M Slavich, Scott M Monroe, and Ian H Gotlib (2009). “Stressful
life events, chronic difficulties, and the symptoms of clinical depression”. In: The Journal of
nervous and mental disease 197.3, p. 154.
Network, NICHD Early Child Care Research et al. (2007). “Age of entry to kindergarten and
children’s academic achievement and socioemotional development”. In: Early education and
development 18.2, pp. 337–368.
NSRC, National Survey Research Center (2013). China Education Panel Survey. http://ceps.
ruc.edu.cn/English/Home.htm.
Oster, Emily (2019). “Unobservable selection and coefficient stability: Theory and evidence”. In:
Journal of Business & Economic Statistics 37.2, pp. 187–204.
Pedersen, Eric R, Eric D Helmuth, Grant N Marshall, Terry L Schell, Marc PunKay, and Jeremy
Kurz (2015). “Using facebook to recruit young adult veterans: online mental health research”.
In: JMIR research protocols 4.2, e3996.
Perrin, Andrew and Monica Anderson (2019). “Share of US adults using social media, including
Facebook, is mostly unchanged since 2018”. In: Pew Research Center 10.
Primack, Brian A, Ariel Shensa, C´ esar G Escobar-Viera, Erica L Barrett, Jaime E Sidani, Jason
B Colditz, and A Everette James (2017). “Use of multiple social media platforms and symp-
toms of depression and anxiety: A nationally-representative study among US young adults”.
In: Computers in human behavior 69, pp. 1–9.
121
Quidt, Jonathan de and Johannes Haushofer (2017). “Depression through the lens of economics: A
research agenda”. In: The Economics of Poverty Traps. University of Chicago Press, pp. 127–
152.
Radloff, Lenore Sawyer (1977). “The CES-D scale: A self-report depression scale for research in
the general population”. In: Applied psychological measurement 1.3, pp. 385–401.
Riehm, Kira E, Calliope Holingue, Luther G Kalb, Daniel Bennett, Arie Kapteyn, Qin Jiang, Cindy
Veldhuis, Renee M Johnson, M Daniele Fallin, Frauke Kreuter, et al. (2020). “Associations
between media exposure and mental distress among US adults at the beginning of the COVID-
19 pandemic”. In: American Journal of Preventive Medicine.
Riumallo-Herl, Carlos, Sanjay Basu, David Stuckler, Emilie Courtin, and Mauricio Avendano
(2014). “Job loss, wealth and depression during the Great Recession in the USA and Europe”.
In: International journal of epidemiology 43.5, pp. 1508–1517.
Rosenbaum, Paul R and Donald B Rubin (1985). “Constructing a control group using multivariate
matched sampling methods that incorporate the propensity score”. In: The American Statisti-
cian 39.1, pp. 33–38.
Schnell, Rainer, Marcel Noack, and Sabrina Torregroza (2017). “Differences in general health of
internet users and non-users and implications for the use of web surveys”. In: Survey Research
Methods. V ol. 11. 2, pp. 105–123.
Schonlau, Matthias, Beth J Asch, and Can Du (2003). “Web surveys as part of a mixed-mode strat-
egy for populations that cannot be contacted by e-mail”. In: Social Science Computer Review
21.2, pp. 218–222.
Schonlau, Matthias, Arthur Van Soest, Arie Kapteyn, and Mick Couper (2009). “Selection bias
in web surveys and the use of propensity scores”. In: Sociological Methods & Research 37.3,
pp. 291–318.
Shaw, Lindsay H and Larry M Gant (2004). “In defense of the Internet: The relationship between
Internet communication and depression, loneliness, self-esteem, and perceived social support”.
In: Journal of Obstetrics and Gynaecology Canada 41.10.
Smarr, Karen L and Autumn L Keefer (2011). “Measures of depression and depressive symptoms:
beck depression inventory-II (BDI-II), Center for Epidemiologic Studies Depression Scale
(CES-D), geriatric depression scale (GDS), hospital anxiety and depression scale (HADS), and
patient health Questionnaire-9 (PHQ-9)”. In: Arthritis care & research 63.S11, S454–S466.
Steers, Mai-Ly N, Robert E Wickham, and Linda K Acitelli (2014). “Seeing everyone else’s high-
light reels: How Facebook usage is linked to depressive symptoms”. In: Journal of Social and
Clinical Psychology 33.8, pp. 701–731.
122
Stewart, Walter F, Judith A Ricci, Elsbeth Chee, Steven R Hahn, and David Morganstein (2003).
“Cost of lost productive work time among US workers with depression”. In: Jama 289.23,
pp. 3135–3144.
Tijdens, Kea and Stephanie Steinmetz (2016). “Is the web a promising tool for data collection in
developing countries? An analysis of the sample bias of 10 web and face-to-face surveys from
Africa, Asia, and South America”. In: International Journal of Social Research Methodology
19.4, pp. 461–479.
Twitter (2020). Q4 and Fiscal Year 2019 Letter to Shareholders. https://s22.q4cdn.com/
826641620/files/doc_financials/2019/q4/Q4-2019-Shareholder-Letter.pdf/.
Valliant, Richard and Jill A Dever (2011). “Estimating propensity adjustments for volunteer web
surveys”. In: Sociological Methods & Research 40.1, pp. 105–137.
Whitbread, D (2014). “Hard evidence: at what age are children ready for school”. In: The Conver-
sation.
WHO (2020). Depression. http://www.who.int/news-room/fact-sheets/detail/
depression/.
Xie, Yu and Jingwei Hu (2014). “An introduction to the China family panel studies (CFPS)”. In:
Chinese sociological review 47.1, pp. 3–29.
Zagheni, Emilio and Ingmar Weber (2015). “Demographic research with non-representative inter-
net data”. In: International Journal of Manpower.
Zhang, Chunni and Yu Xie (2017). “Does school starting age restriction make “July/August Trap””.
In: Sociological Studies 1, pp. 54–77.
123
Appendix
A Appendix to Chapter 1: Social Media Application Updates
Instrument
This section details a new type of instrumental variable to address the reverse causality issue.
I use users’ exposures to social media application updates as instruments for social media usage.
This instrument can only affect outcome through social media usage, and outcome cannot affect the
instrument. As a robustness check, Section B employs Bartik Instrument. Two types of instruments
use different sources of variation.
A.1 Construct the Instrument
I create a new instrument (IV), exposure to social media application updates, specifically for
this data setting. Individuals in the data finish survey on different dates, therefore some individuals
experience more updates than others. For example, if five Facebook updatesfz
1
;z
2
;z
3
;z
4
;z
5
g
were available during survey period (in time order), some individuals experienced only z
1
update,
some experienced(z
1
;z
2
;z
3
), and some experienced all five updates. In addition, iPhone users and
Android users face two different sets of updates. I use this variation in exposure as an instrument
for potentially endogenous social media usage.
Figure A.1 is an example of how to construct this type of instrument. If five updates happen dur-
ing survey period, I create five binary instrument variables corresponding to each update/version.
For two respondents i and j, i participates the survey in early February and j participates in late
124
March. Therefore i experiences only first one among the five updates, and only first IV has value
one and the rest of IVs have value zero. Respondent j experiences first four updates when she
participates the survey, so first four IVs have value one and only the fifth IV has value zero. The
instrument variation comes from the different updates dates of iPhone and Android users and their
different survey dates. Respondents have experienced different versions of social media applica-
tion, so the probabilities they use social media when they participate in the survey will be different.
The instrumental variable model consists of two linear stages. This relationship establishes the first
stage correlation between instrument and endogenous social media usage variables.
A.2 First Stage: New Updates Change Social Media Usage
I assume individual i’s exposure to social media updates, vector Z
i
, affect her social media
usage T
i
in the following way:
T
i
=q
1
X
i
+q
2
Z
i
+ u
i
(A.1)
in which X
i
is a vector of covariates and u
i
is i.i.d. error term.
In the simplest case, Z
i
consists of one binary variable and T
i
is also binary.
A.3 Second Stage: Causal Effect
I use the prediction of first stage,
ˆ
T
i
, as the main regressor in the second stage:
Y
i
=t
IV;app
ˆ
T
i
+q
3
X
i
+ v
i
(A.2)
in which parameter of interestt
IV;app
estimates the local average treatment effect (LATE) for peo-
ple who start or stop using social media because of application updates and v
i
is i.i.d. error term.
t
IV;app
estimates the effect for the population whose usage is easily under the influence of new fea-
tures. The instrumental variable model requires relevance and exclusion restriction assumptions.
125
A.4 Instrument Relevance
The primary goals of application updates are to improve the product and attract more users.
New version usually offers new features or fixes existing problems. Based on the fact that social
media usage in population grows over time, I assume platform updates have impact on individual’s
usage (q
2
6= 0). I use LASSO-IV (Ahrens et al., 2018) alternative model if the overall instrument
vector is weak but some instruments are not.
A.5 Exclusion Restriction
Instrument should not affect the level of depressive symptoms in channel other than social
media usage. The variation of instrument is a combination of exogenous decision of social media
company and arbitrary survey date. I argue the first part is valid in this setting, and empirically
test the second part. I regress survey date on outcome measures in the past (including CESD
score in 2015) and individual’s other characteristics. Table A.1 presents the test results. Column
(1) to (3) have 2017, 2019, and 2020 survey dates as outcome variables, respectively. Prior level
of depressive symptoms does not affect respondent’s choice about when to take the survey. F-
statistics are low for all three years’ estimations. Age, a few education level indicators, ethnicity
indicators, and employment indicators correlate with survey date, the subsequent IV estimation
includes these characteristics as covariates. I assume that survey dates are random conditioning on
these characteristics. This result supports that using survey date variation to construct instrument
satisfies exclusion restriction.
A.6 Data: Social Media Application Updates
Social media application updates provide the first type of instrument. iPhone users and Android
users experience different updates at different times. I collect data on iPhone Twitter/Facebook
updates from https://www.ipa4fun.com/, and the data on Android Twitter/Facebook updates
126
from https://www.apk4fun.com/. The data includes application versions’ IDs and published
dates. For some versions, brief descriptions are also available.
A.7 IV Estimation With 2020 Data
I use social media application updates as instrumental variables for 2020 social media usage.
There are four Twitter updates for iPhone users and three for Android users. For Facebook updates,
there are four updates for iPhone users and four updates for Android users. The final instrument
set includes 13 variables excluding updates having the same publish dates. Table A.4 shows these
IVs means and standard deviations.
Table A.2 shows the IV estimation results (Table 1.16 are corresponding linear regression re-
sults). The top panel of Table A.2 includes two main endogenous social media usage variables:
using Twitter and using Facebook. The result suggests using Twitter increases the support respon-
dents feel from social media.
The bottom panel of Table A.2 includes seven social media usage variables. Using Facebook
indicator is omitted because of its collinearity with the other two Facebook variables. Using Twitter
decreases PHQ and increases positive feelings about social media, although the estimated effects
are not significant. The longer respondents spend on social media everyday, the higher the the
level of depressive symptoms. This result suggests that although social media may have positive
impact on people’s happiness, on average people tend to spend more than optimal time on social
media. Facebook connections decrease the level of depressive symptoms and Facebook followers
increase the level of depressive symptoms. The Twitter connections and followers have similar
impact even though the Twitter effects are not significantly different from zero. The interpretation
is that more friends/connections offer more support, and more followers place extra stress on re-
spondents. These two effects may cancel each other out. Number of connections and followers
are positively correlated with whether they use the two social media platforms (Table A.3). The
followers and connections can be two attributes of social media that affect people’s feelings in
opposite directions.
127
This set of instruments suffer from the weak IV problem, so I also use LASSO to select a strong
set of IVs in each specification. The chosen IVs still have the weak IV problem, and thus Table A.2
only offers suggestive evidence. A limitation is that the 2020 data coincides with COVID-19
pandemic time, so we should interpret the results with caution. Another paper studies how social
media and traditional media associate with mental distress at the beginning of the COVID-19
pandemic by using the same survey data (Riehm et al., 2020).
128
Figure A.1: Example: Creating Instrumental Variables
This figure shows an example of how to construct the first type of instrument. If five updates
happen during the survey period, there will be five binary instrument variables corresponding to
each update. For example we can compare two respondents i (who participates in survey earlier)
and j (who participates in survey later). i experiences only the first one among the five updates, so
only first IV z1 has value one. Respondent j experiences the first four updates, so z1 to z4 have
value one and only z5 has value zero. An intensity version of the instruments are Z1 to Z5. I use
exposure days of each update instead of an indicator as IV .
129
Table A.1: Arbitrary Survey Dates
(1) (2) (3)
2017 Date 2019 Date 2020 Date
CESD (2015) 0.018 0.207 -0.019
(0.023) (0.254) (0.015)
CESD (2017) -0.224 0.008
(0.244) (0.014)
CESD (2019) 0.021
(0.017)
Covariates X X X
F-statistic 2.79 2.26 2.08
Standard Errors are in parenthesis. * p < 0:10, ** p < 0:05, *** p < 0:01. This table tests IV
exclusion restriction. Column (1) to (3) have 2017, 2019, and 2020 survey dates as outcome
variables, respectively. Prior level of depressive symptoms does not affect respondent’s choice
about when to take the survey. Age, part of education level indicators, ethnicity indicators, and
employment indicators correlate with survey date, the subsequent IV estimation includes these
characteristics as covariates.
130
Table A.2: Effect of Social Media Usage on Depressive Symptoms (2020): IV
IV-LASSO
(1) (2) (3) (4)
PHQ-2020 Social Support Feel Happy Feel Relaxed
Twitter 1.648 51.017** 14.728 -18.499
(1.757) (26.068) (28.890) (20.807)
Facebook 14.211 -78.059 19.252
(88.776) (73.783) (65.664)
Covariates X X X X
Past CESD X X X X
Weights X X X X
Mean(outcome)
Twitter=0 1.885 49.324 57.560 58.801
(2.755) (27.372) (25.108) (28.083)
Facebook=0 1.815 41.867 51.691 53.357
(2.790) (28.166) (26.933) (29.333)
1st stage F-stats 0.59 5.75 1.08 0.53
Observations 3081 2402 2397 2401
IV-LASSO
(5) (6) (7) (8)
PHQ-2020 Social Support Feel Happy Feel Relaxed
Twitter -0.878 18.630 14.126 26.126
(1.902) (26.571) (22.598) (27.498)
log(Minutes per Day) 1.049 -14.642 -5.327 -12.340
(0.873) (10.056) (8.734) (10.006)
log(Facebook Connections) -1.467* -0.745 3.994 -8.016
(0.855) ( 8.054) (6.354) (8.276)
log(Twitter Connections) -0.227 17.876 15.377 11.409
(0.993) (11.345) (9.783) (11.839)
log(Facebook Followers) 1.171* -1.896 0.698 4.794
(0.614) (7.721) (6.176) (7.791)
log(Twitter Followers) 0.333 -16.186 -16.931 -18.657
(1.075) (15.159) (12.775) (15.703)
Covariates X X X X
Past CESD X X X X
Weights X X X X
1st stage F-stats 0.13 0.68 0.28 0.45
Observations 3081 2402 2397 2401
Standard errors in parentheses. * p< 0:10, ** p< 0:05, *** p< 0:01. This table shows the IV
estimation results. The top panel includes two social media usage variables and the bottom panel
includes seven ones. The results suggest using Twitter increases positive feelings. Using Facebook
indicator is omitted in bottom panel because of its collinearity with the other two Facebook vari-
ables. We need to interpret this set of results with caution for two reasons: the weak IV problem
and the fact that the 2020 data coincides with COVID-19.
131
Table A.3: Correlation Between 2020 Variables
PHQ-2020 Social Support Feel Happy Feel Relaxed Twitter Facebook
log(Minutes
per Day)
log(Facebook
Connections)
log(Twitter
Connections)
log(Facebook
Followers)
PHQ-2020 1
Social Support -0.0451 1
Feel Happy -0.102*** 0.560*** 1
Feel Relaxed -0.185*** 0.477*** 0.669*** 1
Twitter 0.0187 0.0204 0.0400 0.000762 1
Facebook -0.0449 0.131*** 0.0805** 0.0645* 0.0864** 1
log(Minutes per Day) 0.108*** 0.165*** 0.157*** 0.0641* 0.0638* 0.0718* 1
log(Facebook Connections) 0.0189 0.0961*** 0.0459 0.0232 -0.00442 0.194*** 0.152*** 1
log(Twitter Connections) 0.0351 0.0328 0.00894 -0.0470 0.333*** -0.0212 0.127*** 0.337*** 1
log(Facebook Followers) -0.00381 0.0267 0.00664 -0.0320 0.0558* 0.0917** 0.129*** 0.641*** 0.338*** 1
log(Twitter Followers) 0.0624* 0.0140 -0.00304 -0.0546 0.302*** -0.0785** 0.114*** 0.284*** 0.796*** 0.389***
p< 0:05 p< 0:01p< 0:001
132
Table A.4: IV Distribution
Instruments Mean Standard Deviation
Twitter-iPhone-IV1 .212 .408
Twitter-iPhone-IV2 .079 .269
Twitter-iPhone-IV3 .032 .177
Twitter-iPhone-IV4 .010 .099
Twitter-Android-IV1 .185 .388
Twitter-Android-IV2 .034 .183
Twitter-Android-IV3 .009 .097
Facebook-iPhone-IV1 .212 .408
Facebook-iPhone-IV2 .101 .301
Facebook-iPhone-IV3 .023 .151
Facebook-iPhone-IV4 .006 .077
Facebook-Android-IV1 .009 .097
Facebook-Android-IV2 .001 .037
133
B Appendix to Chapter 1: Bartik Instrument
B.1 Background
I exploit another source of variation as instrument to check the robustness of my findings.
Bartik instrument (Bartik, 1991; Goldsmith-Pinkham et al., 2018) is in the form of interaction
between local industry shares and national industry growth rates. Following the same methodology,
I construct Bartik-type instrument by interacting state age shares and average age-specific social
media usage growth in the US.
In the model of how social media usage growth affects state level of depressive symptoms, the
main endogenous regressor T
s
and Bartik instrument B
s
for state s are:
T
s
=
å
k
a
sk
g
sk
and B
s
=
å
k
a
(0)sk
g
k
(B.1)
k is age groups. a
sk
is state s’s current age composition and a
(0)sk
is the composition in the past.
g
sk
is state level social media usage growth rate for each category k and g
k
is average usage growth
rate of category k in the US. The validity and issues of Bartik-type instrument are established in
the literature (Goldsmith-Pinkham et al., 2018).
B.2 Data: Social Media Usage Growth
I use US age-specific social media usage growth to construct the instrument. I collect this
information from https://www.pewresearch.org/. Table B.1 summarizes the usage growth
rates by age. I use age group categories from Pew Research Center’s Social Media Use reports.
Twitter usage grows for 18-24 and 30-49 age groups, and decreases for the other three groups. The
changes are all under five percentage points.
134
B.3 Bartik Instrument Result
Identification with Bartik Instrument is based on exogeneity of the demographic shares. This
share measures the states’ differential exposures to common national growth of social media us-
age. The national growth serves as instrument because it correlates with state social media usage
growth, but is exogenous to other state-specific shocks.
Table B.1 (a) shows the national growth rate of Twitter usage for five age groups, and (b) shows
the Bartik Estimation of Twitter usage growth on CESD score at state level. The estimated effect
is -0.067. The estimation is not accurate due to the small sample size (51 state level observations),
and the magnitude is small. Nevertheless the estimation suggests negative correlation between
social media usage and the level of depressive symptoms. With more data points available in the
future, this model can offer more accurate insights.
Other demographic groups’ differential exposures to national social media growth are potential
Bartik-type instruments. I choose age groups because there are enough variations in differential
exposure (for example, gender differential exposures cannot offer enough variations), and the re-
quired age group level data is available. For example, I do not use income as demographic groups
because the national growth data and state level data categorize income groups differently.
135
Table B.1: Summary of Bartik IV Estimation
(a) US Age-specific Twitter Usage Growth Rate g
k
Age Group k 18-24 25-29 30-49 50-64 > 65
2016 0.36 0.23 0.21 0.1
2018 0.45 0.33 0.27 0.19 0.08
2019 0.44 0.31 0.26 0.17 0.07
2019-2018 -0.01 -0.02 -0.01 -0.02 -0.01
2017 (Imputed) 0.405 0.345 0.25 0.2 0.09
2019-2017 0.035 -0.035 0.01 -0.03 -0.02
(b) Bartik IV Estimation
Effect of Twitter Usage Growth on CESD -0.067
(0.388)
Observations 51
F-statistics 3.10
R-squared 0.194
Standard errors in parentheses. * p< 0:10, ** p< 0:05, *** p< 0:01. Panel (a) shows the national
growth rate of Twitter usage for five age groups. Data on Twitter usage in 2016, 2018, and 2019
is available. I use 2016 and 2018 data to impute 2017 data (mean of the two years), then estimate
the growth rate between 2017 and 2019 by comparing the social media usage in the two years.
The national growth rate for five age groups are 0.035, -0.035, 0.01, -0.03, and -0.02 respectively.
Panel (b) shows the Bartik Estimation of Twitter usage growth on CESD score at state level. The
estimated effect is -0.067. The estimation is not accurate due to the small sample size (51 state
level observations).
136
C Additional Results to Chapter 1
0 1 2 3
density
0 .2 .4 .6 .8
Propensity score, waves=1
waves=1 waves=2
Figure C.1: Test Overlap Assumption
This figure shows estimated densities of the probability of being in each wave of panel data (2017
and 2019). The overlap assumption to use inverse probability weighting is not violated.
137
-3 -2 -1 0
Twitter effect
female male married married* separated divorced widowed never married

Figure C.2: Heterogeneous Effect: Gender, Marital Status
Married group means “married and spouse lives with me.” Married* means “married and spouse lives elsewhere.” There is little hetero-
geneity for different genders and marital status.
138
-20 -10 0 10 20
Twitter effect
0 3 6 9 12 15 18 21 24
CESD 2015
Figure C.3: Heterogeneous Effect: Past Level of Depressive Symptoms
The figure shows point estimations and 95% confidence intervals. The Twitter usage decreases the level of symptoms for groups where
2015 CESD equals 0 and 9.
139
-10 -5 0 5
Twitter effect
0 20 40 60 80 100 120
Hours of work per week

Figure C.4: Heterogeneous Effect: Working Hours
Social media has a larger mitigation effect on the level of depressive symptoms for people who work more. The estimation for people
who work 40 hours per week is the most accurate mainly because most samples belong to this group.
140
-8 -6 -4 -2 0 2
Twitter effect
1 2 3 4 5 6
Number of children

Figure C.5: Heterogeneous Effect: Number of Children
Samples with more than six children are grouped in six-children group. People with more than five children benefit from using Twitter
the most in terms of depression.
141
(a) (b) (c)
(d) (e)
Figure C.6: Twitter and Facebook User Percentages by State
This figure shows social media usage in each state. Top panels (a) - (c) show Twitter usage in 2017, 2019, and 2020. Bottom panels
(d) and (e) show Facebook usage in 2019 and 2020. Some states’ user proportions grow over time, and some change in the opposite
direction. Twitter user proportion in 2020 grows on average (see panel (c) versus panel (b)). Facebook user proportion in 2020 decreases
on average (see panel (e) versus panel (d)) in my sample, but the user population can still grow. Another explanation of lower Facebook
user proportion in 2020 is that the survey coincides with COVID-19, and it can change people’s preference over social media. For
example, more people use Twitter instead of Facebook during COVID-19.
142
(a) (b) (c)
Figure C.7: The levels of Depressive Symptoms by State
This figure shows the levels of depressive symptoms in different states over time. Panel (a) and (b) shows CESD scores (range from
0 to 24), and panel (c) shows PHQ scores (range from 0 to 12). The figure suggests the levels of depressive symptoms in many states
decrease from 2017 to 2019 (see panel (a) versus (b)).
143
Table C.1: Depressive Measures
(a) CESD-8
Items Coded Answer
1. You felt depressed. 3 (YES) 0 (NO)
2. You felt that everything you did was an effort. 3 (YES) 0 (NO)
3. Your sleep was restless. 3 (YES) 0 (NO)
4. You were happy. 0 (YES) 3 (NO)
5. You felt lonely. 3 (YES) 0 (NO)
6. You enjoyed life. 0 (YES) 3 (NO)
7. You felt sad. 3 (YES) 0 (NO)
8. You could not get going. 3 (YES) 0 (NO)
(b) PHQ-4
Items Coded Answer for Each Item
1. Feeling nervous, anxious, or on edge.
(0) not at all, (1) several days,
2. Not being able to stop or control worrying.
3. Feeling down, depressed, or hopeless.
(2) more than half the days, (3) nearly every day.
4. Little interest or pleasure in doing things.
(c) Descriptive Measures
Items Coded Answer
When using social media I usually feel social pressure/support. 0 (Pressure) - 100 (Support)
When using social media I usually feel unhappy/happy. 0 (Unhappy) - 100 (Happy)
When using social media I usually feel anxious/relaxed. 0 (Anxious) - 100 (Relaxed)
CESD includes eight items. Each item has a score from zero to three. The fourth and sixth items
are coded by the author in the opposite direction because unlike the rest of the items, these two
questions ask about positive feelings. PHQ includes four items and all items ask about negative
feelings. Lastly, Table C.1 (c) summarizes three questions which directly ask about people’s feel-
ings about using social media.
144
Table C.2: Social Media Usage and the Level of Depressive Symptoms
(1) (2) (3)
OLS OLS OLS
Variable CESD (2017) CESD (2019) PHQ (2020)
Twitter (2017) -0.219
(0.279)
Twitter (2019) 0.233
(0.229)
Twitter (2020) 0.342**
(0.152)
Facebook (2019) -0.356
(0.261)
Facebook (2020) -0.084
(0.119)
Covariates X X X
Past CESD X X X
Weights X X X
Mean(outcome)
Twitter=0 4.617 1.305 1.885
(6.284) (3.959) (2.755)
Facebook=0 3.845 1.815
(6.027) (2.790)
Adjusted R-squared 0.322 0.167 0.237
F-statistics 10.965 2.715 11.511
Observations 3757 3622 3081
Standard errors in parentheses. * p < 0:10, ** p < 0:05, *** p < 0:01. This table shows the
estimation of social media usage’s correlation with CESD score in 2017, 2019, and on PHQ score
in 2020. The results suggest correlation of Twitter usage with the level of depressive symptoms
can be close to zero or positive. It also suggests more Facebook usage is associated with lower
level of depressive symptoms, even though the correlation is not significantly different from zero.
Table C.3: Fixed Effect Model With Sub-samples
(1) (2) (3) (4) (5)
Two-Way Fixed Effects
CESD (2017, 2019)
Full Sample CESD<21 CESD<18 CESD<15 CESD<12
Twitter (2017, 2019) -1.175** -1.271** -0.690 -0.395 -0.326
(0.510) (0.507) (0.444) (0.376) (0.339)
Covariates X X X X X
Weights X X X X X
Mean(outcome)
Twitter=0 4.313 3.809 3.339 2.774 2.400
(6.028) (5.283) (4.597) (3.761) (3.192)
Adjusted R-squared 0.633 0.576 0.502 0.472 0.489
F-statistics 1.290 1.503 1.215 1.663 1.614
Observations 1994 1914 1842 1738 1652
Standard errors in parentheses. * p< 0:10, ** p< 0:05, *** p< 0:01. The estimations are from
two-way fixed effects model with one member from each household. The negative effect shrinks
as sample becomes smaller, but the implication stays the same, i.e., social media usage decreases
the level of depressive symptoms.
145
Table C.4: Effect of Social Media Using Frequency on CESD (2019)
Frequency Twitter Facebook
(Omitted Group: Never in Last 12 Months)
A Few/Year
-0.195 0.088
(-0.57) (0.20)
A Few/Month
-0.374 0.42
(-0.95) (0.89)
Once/Week
-0.445 0.171
(-0.98) (0.36)
A Few/Week
-0.153 -0.0146
(-0.40) (-0.04)
Once/Day
-0.235 -0.189
(-0.61) (-0.48)
A Few/Day
-0.530* 0.104
(-1.65) (0.27)
Covariates X X
Observations 1647 4232
This table estimates the correlation between different Twitter and Facebook use frequencies and
CESD score in 2019. The coefficients represent each frequency group’s difference from omitted
“never” frequency group. * p< 0:10, ** p< 0:05, *** p< 0:01.
146
D Appendix to Chapter 3: Cognitive Skill Measures in CEPS
China Education Panel Survey (CEPS) includes 20 cognitive skill questions for the 7th grade
and 22 cognitive skill questions for the 9th grade. These questions cover topics such as language,
graphics, calculation, and logic. The raw score is the number of correct answers, and the stan-
dardized score is based on three-parameter logistic model. Here is a brief description of the three-
parameter logistic model (3PL) (NSRC, 2013):
Let X
i j
= 0 indicate an incorrect answer and X
i j
= 1 indicate a correct answer, the 3PL model
posits the probability of a correct response as
P
j
(q
i
)= P(X
i j
= 1)= c
j
+
1 c
j
1+ exp[a
j
(q
i
b
j
)]
(D.1)
in which q
i
is the latent “amount” of cognitive skill for individual i, a
j
measures how influential
changes in q
i
are on changes in P(X
i j
= 1), b
j
represents the “item difficulty,” and c
j
represents a
“guessing parameter.”
147

Abstract (if available)

Abstract This dissertation consists of three essays on development economics and health economics. The first chapter examines the impact of social media usage on depressive symptoms in the United States. The use of social media can potentially decrease the level of depressive symptoms by providing support or increase the level of depressive symptoms by putting social pressure on users. This chapter leverages a fixed-effects model to estimate the effect of using social media platforms on depressive symptoms. I find that using Twitter decreases the level of depressive symptoms by 27%. This result explains why social media usage in the US has grown steadily even though most studies found that more usage correlated with higher levels of depressive symptoms. There is heterogeneity with respect to age, income, education, race, previous level of depressive symptoms, and region. The average labor market benefit that comes from this effect is equivalent to 0.1% GDP in the US.
In the second chapter, I examine the performances of different bias correction methods, such as matching and weighting methods, on improving the representativeness of social media data. I find that matching and weighting methods can effectively improve the representativeness of social media users in most cases examined. Matching methods with smaller number of neighbors or smaller radius produce smaller biases. Improving the representativeness of Twitter users is easier than improving the representativeness of Facebook users.
The third chapter is a collaboration with Yinan Liu, in which we study the impact of the primary school starting age policy in China on both short-run and long-run outcomes. We examine the household characteristics of the right age group, early group, and late group based on the compliance. Starting school late is negatively associated with cognitive skills, test scores, highest education achieved and income. We also explore the potential explanations why a large proportion of households send children to primary school before they reach the eligible age in China.

Linked assets

University of Southern California Dissertations and Theses

Conceptually similar

PDF

Essays on development economics

PDF

Essays on family and labor economics

PDF

Essays in the economics of education and conflict

PDF

Essays on development and health economics

PDF

Essays on health economics

PDF

Three essays on econometrics

PDF

Essays on economics of education

PDF

Three essays on cooperation, social interactions, and religion

PDF

Essays on family planning policies

PDF

Essays on labor and development economics

PDF

Essays in health economics: evidence from Medicare

PDF

Essays on economics of education and private tutoring

PDF

Essays on treatment effect and policy learning

PDF

The EITC, labor supply, and child development

PDF

Essays on econometrics

PDF

Essays in development economics

PDF

Essays on consumer conversations in social media

PDF

Essays on health and aging with focus on the spillover of human capital

PDF

Essays on education and institutions in developing countries

PDF

Essays on health insurance programs and policies

Asset Metadata

Creator Jiang, Qin (author)

Core Title Essays on development and health economics: social media and education policy

School College of Letters, Arts and Sciences

Degree Doctor of Philosophy

Degree Program Economics

Degree Conferral Date 2021-08

Publication Date 06/28/2021

Defense Date 06/01/2021

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

Tag depressive symptoms,development economics,education,facebook,fixed effects model,Health,matching,OAI-PMH Harvest,school starting age,selection bias,social media,Twitter,weighting

Language English

Contributor Electronically uploaded by the author (provenance)

Advisor Strauss, John (committee chair), Kapteyn, Arie (committee member), Painter, Gary (committee member), Ridder, Geert (committee member), Weaver, Jeffrey (committee member)

Creator Email qinjiang@usc.edu,qinjiangnat@gmail.com

Permanent Link (DOI) http://doi.org/10.25549/usctheses-c89-472054

Unique identifier UC13012640

Identifier etd-JiangQin-9673.pdf (filename), usctheses-c89-472054 (legacy record id)

Legacy Identifier etd-JiangQin-9673

Dmrecord 472054

Document Type Dissertation

Rights Jiang, Qin

Internet Media Type application/pdf

Type texts

Source 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 a...

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