University of Southern California Dissertations and Theses

Essays on work, retirement, and fostering longer working lives

(USC Thesis Other)

Essays on work, retirement, and fostering longer working lives

PDF

Download a page range

Download transcript

Copy asset link

Request this asset

Transcript (if available)

Content i

Essays on Work, Retirement, and Fostering Longer
Working Lives

by

Zeewan Lee

A Dissertation Presented to the
FACULTY OF THE USC GRADUATE SCHOOL
UNIVERSITY OF SOUTHERN CALIFORNIA

In Partial Fulfillment of the
Requirements for the Degree
DOCTOR OF PHILOSOPHY
PUBLIC POLICY AND MANAGEMENT

May 2020

Copyright 2020 Zeewan Lee
ii

Acknowledgements

I am deeply indebted to my dissertation committee—Emma Aguila, Alice Chen, Neeraj Sood, and
Julie Zissimopoulos—as well as Marco Angrisani and Dowell Myers for generous guidance. Many
thanks to my family, Kangsuk Lee, Linna Zhu, Yolanda Zhu, Frank Xu, Yi Chen, Soledad de
Gregorio, Sushant Joshi, Soyoon Choo, and Kyuri Park for all the love and support.

iii

Table of Contents

Acknowledgements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .ii
List of Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . v
List of Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .vii
Abstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . viii
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .1
1. Work in the Second Machine Age: Understanding the Impact of Automation on Retirement
Decisions
1.1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.2. Conceptual Framework . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
1.3. Study Data and Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
1.4. Empirical Approach. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
1.5. Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
1.5.1. Summary Statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
1.5.2. Estimation Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
1.6. Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
1.7. Appendix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34

2. Automation, Aging, and Unretirement Decisions
2.1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
2.2. Conceptual Framework . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
2.2.1. Theories on the Nature of Technological Growths . . . . . . . . . . . . . . . . . . . . . . . .43
2.2.2. Linking Technological Growths and Unretirement . . . . . . . . . . . . . . . . . . . . . . . 51
2.3. Study Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
2.3.1. Measuring the Extent of Automation-Resistance. . . . . . . . . . . . . . . . . . . . . . . . . 56
2.3.2. Exploring the Automation-Resistance Indices . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
2.3.3. Analysis Sample . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
2.3.4. Summary Statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
2.4. Empirical Approach. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
iv

2.4.1. Automation-Resistance and Unretirement Decisions . . . . . . . . . . . . . . . . . . . . . .68
2.4.2. Unretirement Job Choices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .72
2.4.3. Returns to Skill-to-Task Matching . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .75
2.5. Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
2.5.1. Who Returns to Work? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
2.5.2. What Types of Jobs Do the Unretirees Choose? . . . . . . . . . . . . . . . . . . . . . . . . . 80
2.5.3. Financial and Mental Health Returns to Skill-to-Task Matching . . . . . . . . . . . . .83
2.6. Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
2.7. Appendix A . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .97
2.8. Appendix B . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
2.9. Appendix C . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105

3. Lost in Complexity? The Impacts of Health Insurance Literacy on ACA-Induced Retirement
3.1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110
3.2. Conceptual Framework . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114
3.2.1 Insurance-Induced Job Lock. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114
3.2.2 Insurance Literacy, Insurance Uptake, and Retirement . . . . . . . . . . . . . . . . . . . . 116
3.2.3 Background on ACA Health Insurance Exchange and Assister Programs . . . . . 118
3.3. Study Data and Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120
3.3.1 Analysis Sample. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120
3.3.2 Measuring Health Literacy. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122
3.4. Empirical Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .124
3.5. Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126
3.5.1 Assessing Baseline Trends. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126
3.5.2 Relationship between the Insurance Literacy and Insurance Uptake . . . . . . . . . .128
3.5.3 Impact of the Exchanges on Planned Retirement . . . . . . . . . . . . . . . . . . . . . . . . .129
3.5.4 The Role of Insurance Literacy on Planned Retirement . . . . . . . . . . . . . . . . . . . .131
3.6. Discussion. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .132
3.7. Appendix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .142

Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .145

v

List of Tables

1. Work in the Second Machine Age: Understanding the Impact of Automation on Retirement
Decisions
1.1. Un-Automatable Skills . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
1.2. Examples of Skilled and Unskilled Occupations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
1.3. Summary Statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
1.4. Estimating the Impact of Automatability on Planned Retirement . . . . . . . . . . . . . . . . . 30
1.5. Impact of Automatability on Planned Retirement, by Gender . . . . . . . . . . . . . . . . . . . . 31
1.6. Impact of Automatability on Planned Retirement, by Index Components . . . . . . . . . . . 32
1.7. Automatability and Actual Retirement: A Binary-Outcome Regression . . . . . . . . . . . . 33
1.4A. Estimating the Impact of Automatability on Planned Retirement (detailed) . . . . . . . . 34
1.5A. Impact of Automatability on Planned Retirement, by Gender (detailed) . . . . . . . . . . .35
1.6A. Impact of Automatability on Planned Retirement, by Index Components (detailed) . .36
1.7A. Automatability and Actual Retirement: A Binary-Outcome Regression (detailed) . . .37

2. Automation, Aging, and Unretirement Decisions
2.1. Elements of Automation-Resistance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
2.2. Summary Statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
2.3. Unretirement Across the Analytic Period . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
2.4. Unretirement Decisions by ARIS, Linear Probability Model . . . . . . . . . . . . . . . . . . . . 93
2.5. Unretirement Decisions by ARIS, Logit Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
2.6. Unretirement Job Choices by ARIS: Task-Specific Automation Resistance of Unretirees’
Jobs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
2.7. Unretirement Job Choices by ARIS: White-Collar Status, Industry, and STEM . . . . . . 96
2.8. Monetary and Mental Health Returns to Skill-to-Task Matching. . . . . . . . . . . . . . . . . . 96
2.4A. Unretirement Decisions by Resistance to Automation Based on Skills, Linear
Probability Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
2.5A. Unretirement Decisions by ARIS, Logit Model. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
2.6A. Unretirement Job Choices by ARIS: Task-Specific Automation Resistance of
Unretirees’ Jobs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
2.7A. Unretirement Job Choices by ARIS: White-Collar Status, Industry, and STEM . . . .108
2.8A. Monetary and Mental Health Returns to Skill-to-Task Matching . . . . . . . . . . . . . . . 109

vi

3. Lost in Complexity? The Impacts of Health Insurance Literacy on ACA-Induced Retirement
3.1. Types of Exchange and Assister Grants . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137
3.2. Summary Statistics and Baseline Characteristics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 138
3.3. Evaluating Pre-Treatment Trends . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139
3.4. Impact of Introduction of Exchanges on Planned Retirement: Difference-in-Differences .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140
3.5. Impact of Introduction of Exchanges on Planned Retirement: Triple-Differences . . . .141
3.1A. ACA Assister Program Regulations by State. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142
3.2A. Impact of Introduction of Exchanges on Planned Retirement, Difference-in-
Differences, Detailed . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143
3.3A. Impact of Introduction of Exchanges on Planned Retirement, Triple-Differences,
Detailed. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 144

vii

List of Figures

1. Work in the Second Machine Age: Understanding the Impact of Automation on Retirement
Decisions

2. Automation, Aging, and Unretirement Decisions
2.1. Evolution of the AROT Index Scores by Occupational Groups . . . . . . . . . . . . . . . . . . . 90
2.2. Evolution of the ARIS Index Score by Gender. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
2.3. Evolution of the ARIS Index Score by Education. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91

3. Lost in Complexity? The Impacts of Health Insurance Literacy on ACA-Induced Retirement
3.1. Trends in Planned Retirement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .135
3.2. Correlations between the Amount of Assister Funding Available in Each State and the
Total Number of Enrollments in Exchanges . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .136
3.3. Relationship between the Assister Funding and the Exchange Enrollments . . . . . . . . .137

viii

Abstract

In a world where most nations are undergoing rapid population aging and experiencing a
shrinking tax base to provide retirees with a sufficient level of public health insurance- and
pension- plans, one of the feasible solutions to deal the lack of financial base for retirees is to
promote longer working-lives. Using data from the United States, this dissertation explores how
the automation of labor and a recent healthcare reform changing the system of health insurance
provision pose opportunities and challenges to promoting longer working-lives by influencing
workers’ timing and form of retirement.
The first chapter asks what encourages aging workers to continue working and seeks for
answers from workers’ heterogeneous experiences of the process of automation. Drawing data
from the Health and Retirement Study (HRS) and O*NET, we assess the consequence of
automation on the retirement behaviors using various panel estimations including fixed-effect,
logit, and multinomial logit models. Using the same data for years 1992-2014, the second
chapter explores retirees’ late-career job sorting behaviors as they engage in unretirement.
Specifically, we assess whether and how much individuals’ automation-resistance plays a role in
determining unretirement likelihood—and if unretiring—what types of jobs the individuals
choose, and whether the act of ‘working in old age’ can make the them financially and
psychologically better off. Lastly, in light of the strong linkage between health insurance
provision and employment in the US healthcare system, the third chapter explores whether the
healthcare policy reform alters aging individuals’ retirement timing. In doing so, we study the
effect of the ACA on workers retirement decisions while accounting for their heterogeneous
ix

levels of insurance literacy. After finding that ACA accelerates retirement by removing job-locks
induced by employer-sponsored health insurance, we close by discussing that, from the policy
perspective, the ACA-induced hastening of retirement could potentially counteract past policy
efforts to foster longer working lives.

1

Introduction

Population aging is a global phenomenon that has a far-reaching impact on the labor
market of affected nations. Increasing longevity and declining fertility rates have shifted the age
distribution of populations around the globe—especially in the developed nations (Anderson and
Hussey, 2000). In the United States alone, individuals aged 65 and older that constituted one-
fifth of the population in 1990s will comprise two-fifths of the population by 2040 (Costa and
McCrae, 1997). Such structural changes suggest that the number of retirees is growing while the
tax base consisting of the working-age population is shrinking—leaving the nation with
insufficient power to continue providing the current level of post-retirement benefits (e.g.,
Medicare, Social Security income) for the retirees (Gruber and Wise, 1999; Munnell and
Rutledge, 2013).
One of the strategies to remedy the financial toll of the population aging is to promote
longer working-lives, especially for workers nearing retirement (Maestas and Zissimopoulos,
2010).
1
My dissertation sheds lights on some of the lesser-known elements that alter the degree
of attractiveness of ‘working at old age’ through both financial and non-financial means, thereby
providing insights from which policy-makers can build future policies and programs. The
elements explored are technological growths that induce automation of labor and a recent US
healthcare reform. The dissertation is divided into three distinct essays, all of which are
quantitative in nature. For the purpose of this manuscript, each essay comprises a chapter.

1
Past efforts on encouraging workers to extend their working lives include terminating the Social Security earnings
test and repealing the mandatory retirement. (Gustman et al., 2019).
2

The first chapter titled, “Work in the Second Machine Age: Understanding the Impact of
Automation on Retirement Decisions,” is co-authored with Marco Angrisani from USC Center
for Economic and Social Research. The first chapter asks what encourages aging workers to
continue working, and seeks for answers from workers’ heterogeneous experiences of the
process of automation. Drawing data from the Health and Retirement Study (HRS) and
information on the skills exercised at each occupation available at Occupational Information
Network (O*NET), we assess how the pressures of automation affect retirement timing of aging
individuals, both in terms of expected and actual retirement behaviors. This study contributes to
the labor supply and retirement literature by (1) understanding the nature of the automation based
on the theoretical model of skill-biased technological change (SBTC) and (2) creating a unique
skill-index that ranks workers by the extent to which they can resist the pressures of automation
based on their skill-sets. Overall, this study lends sufficient evidence that workers who are
‘higher skilled’ in a sense that their skills are more resistant to the pressures of automation are
more likely to postpone timing of retirement: We find that a one standard deviation increase in
the skill index leads to a 4.2% increase in the probability of working full-time after age 62 from
the average of 54.7%, and a 5.6% increase in the probability of working full-time after age 65
from the average of 33.9%. Likewise, the actual retirement is also postponed with the rise in the
skill index, with a one standard deviation increase in the skill index leading to a reduction in
actual retirement in the subsequent wave by 6.4% from the average of 14.0%.
The single-authored second chapter is titled, “Automation, Aging, and Unretirement
Decisions.” In an effort to study the linkages among population aging, growing unretirement
(i.e., a phenomenon whereby the fully-retired returns to the labor force) in the labor market, and
technological changes inducing automation of labor, this chapter explores retirees’ unretirement
3

decisions in light of their heterogeneous abilities to resist the pressures of automation. Drawing
from the 1992 to 2014 HRS and the O*NET, we examine whether unretirement depends on the
automation resistance of individual skills (ARIS). For unretirees, we estimate the relationship
between ARIS and task-specific resistance to automation in the jobs that they choose, and we
assess the mental health and financial benefits of skill-to-task matching. Controlling for a rich set
of individual observables, we use a logit model to show that a one standard deviation increase in
the ARIS index leads to a 1.5% increase in unretirement from the average unretirement
likelihood of 40%. Finally, we employ a Heckman two-stage model to analyze job choices and
returns for those who unretire. We find that those with high ARIS choose jobs that have higher
task-specific automation resistance. As for returns, we find that the higher skill-to-task matching
generates an increase in wages by 82.6% when we control for within-cohort variations in the
valuation of automation-resistance and a reduction in being diagnosed with depression by 0.6-
1.8% across all cohorts.
Lastly, the third chapter, “Lost in Complexity? The Impacts of Health Insurance Literacy
on ACA-Induced Retirement,” is co-authored with Alice Chen from Sol Price School of Public
Policy, University of Southern California. The last chapter shifts to study whether there are any
policy measures in healthcare that alter aging employees’ work and retirement incentives. In
light of the strong linkages between health insurance provision and employment in the US
healthcare system, we estimate the effect of the Affordable Care Act (ACA), a comprehensive
healthcare reform, on workers retirement decisions using difference-in-differences estimations.
Next, suspecting that low levels of understanding of the system of insurance provision can
mitigate the labor market implications of the ACA, we study the effect of the ACA on retirement
timing while accounting for their heterogeneous levels of insurance literacy using triple-
4

differences estimations. We find that the ACA accelerates retirement for individuals who did not
have access to post-retirement insurance benefits from their employers by 0.31 years. Moreover,
a one-percent increase in insurance literacy further hastens retirement by 0.0041 years. In
addition to validating the job-lock removing effect of the ACA, our findings highlight the value
of promoting health insurance literacy in order to enhance the efficacy of the healthcare reform
measures. We close by discussing that—while the removal of the job-lock via the ACA can be
welfare improving—from the policy perspective, the reform-induced acceleration of retirement
can be viewed as a side-effect that counteracts past policy efforts to foster longer labor force
participation.
In sum, this dissertation provides insights on how the pressures of automation and the
ACA push and pull individuals in and out of the labor force. Such insights will assist policy
makers with promoting their policy agenda—encouraging prolonged working lives—in the face
of rapid technological changes and healthcare reforms that change individuals’ access to health
insurance. Chapter 1 and 2 on the effect of automation on retirement and unretirement decisions
show that individuals whose skills render them more automation-resistant are prone to
lengthening the overall duration of work, not only by postponing initial full retirement but also
by unretiring more than do individuals with low automation-resistance. Then, policy-makers
with the agenda to lengthen working lives ought to pay a close attention to individuals with low
automation-resistance, or those who are at risk of being replaced by the machines. For the sake
of this group of individuals, objectives of on-the-job training and re-training systems ought to be
changed to improve individuals’ soft skills (e.g., creativity and social intelligence), which are
shown in the literature and in the Chapter 1 and 2 of this dissertation to raise individuals’
automation-resistance. Examples of such efforts include providing more opportunities for the
5

trainees to work on their communication skills, presentation skills as well as to engage in team-
based projects.
2
Lastly, Chapter 3 demonstrates that, by providing accessible and affordable
health insurance plans independently of employment, via launching the Health Insurance
Exchanges, the ACA has facilitated an earlier termination of work. Specifically, the ACA
encourages aging individuals who used to continue working to maintain employer-sponsored
insurance to retire freely. While it may seem that the ACA counteracts the policy agenda to
promote longer labor force participation, it is hasty to conclude so in that the hastened timing of
the ACA-induced retirement does not preclude those retired from returning to the labor force
later. In order to reveal a linkage between the healthcare reform and the total duration of working
lives, further examinations should be conducted on whether or how job entering-, job switching-,
and unretirement- behaviors are influenced by the ACA.
In this dissertation, explored are labor supply behaviors of older adults as opposed to the
entire labor force regardless of age. The focus lies on older individuals in that the pressures of
automation is expected to influence them—as opposed to younger generations workers—more
profoundly due to their lack of flexibility to re-train themselves to be more automation-resistant.
Hence, should there be any effect of automation on the labor force, it is expected to be
particularly severe among the older individuals. Moreover, when making work, retirement, and
unretirement decisions, older individuals face different institutional considerations such as
Medicare eligibility, Social Security claiming, and private pension benefits that are less relevant
for younger workers (Blau and Gilleskie, 2006; Coile and Gruber, 2007; French and Jones, 2011;

2
While raising awareness of the value of soft skills could take place at an earlier stage, during the secondary, post-
secondary, and college education, altering objectives of the education system is more complex as they take longer to
be modified than the job training programs. To revamp the education system to better prepare students to finding
jobs in the face of rapid automation pressures, the pace of change in the education system needs to be rendered faster
(e.g., allowing for more short-term interventions) so as to match that of technological changes generating the
automation of labor.
6

Gustman and Steinmeier, 2005). Such states in which older individuals find themselves call for
research explicitly dedicated to the group of individuals within the labor force.
As all three chapters of the dissertation is written within the context of the United States,
their findings should be read while taking into consideration unique characteristics of the US
labor market. Nevertheless, insights garnered from the chapters will have implications across
national borders as population aging, changing nature and forms of retirement, technological
advancements, and healthcare reforms are structural changes shared by nations around the globe.

7

Chapter 1

Work in the Second Machine Age: Understanding the Impact of
Automation on Retirement Decisions

1.1 Introduction

We live in a world today whereby machines are replacing human labor in a process
known as automation. Since the 1980s in and out of the US, there has been a continuous fall in
the number of jobs carried out by human labor (Rotman, 2013). The process of automation began
with machines’ replacing people performing manual jobs, but it has recently evolved to carry out
jobs that involve more complex cognitive tasks (Autor et al., 2003). While the soaring accuracy
and efficiency of machines in handling tasks formerly carried out by human labor may be a boon
to many employees, not all employees racing against the machines will welcome such rapid
technological innovation and automation encroaching upon their workplaces (Brynjolfsson and
McAffee, 2014). Thus, the process of automation is bound to have a direct impact on the
workers’ labor supply behaviors—besides the layoffs and structural employment changes
initiated from the employers.
In this study, we investigate the role played by the process of automation in shaping labor
supply decisions of workers approaching and on the verge of retirement, a question that has
received relatively little attention in the literature. In doing so, we rely on the theory of skill-
biased technological change (SBTC) (Card and DiNardo, 2002; Goldin and Katz, 1998;
8

MacCrory et al., 2014) to understand the potential effects of automation on retirement choices
and inform our empirical analysis. According to the theory of SBTC, the relationship between
technology and human beings is such that the rapid technological growths automate certain skills
performed at work while preserving the rest—thereby holding a different impact across workers
based on their skill-sets. A large literature has assessed the differential dynamics the process of
SBTC has across workers in terms of wage inequalities or productivity differentials (Bartel and
Sicherman, 1993; Card and DiNardo, 2002; Fernandez, 2001; Galor and Moav, 2000). We,
instead, link the SBTC-induced inequalities in wage and perceived job satisfaction between the
workers with high and low automation-resistance to their retirement decisions. For this purpose,
we focus on a sample of older workers, among whom labor supply choices may be relatively
more sensitive to the SBTC.
Using the Health and Retirement Study (HRS) in conjunction with the Occupational
Information Network (O*NET) database maintained by the US Department of Labor, we devise
a unique index of automatability of workers based on their main skill-sets used at their career
jobs. In that the higher index value indicates that the worker is more ‘skilled’, or more resistant
to the pressures of automation, we refer to our index simply as the skill index. Such a
categorization allows us to identify the workers with high automation-resistance, to whom the
process of automation is a positive experience as their skills are most likely complemented by
technology, from their counterparts, to whom the technology serves primarily as a substitute.
In contrast to the existing literature that tends to proxy automatability by educational
attainment, the rate of usage of technological devices, or whether workers have white-collar
occupations (Autor et al., 1998; Card and DiNardo, 2002; Goldin and Katz, 2001), our index
evaluates workers’ automatability directly by assessing how much creativity, non-routineness,
9

and social intelligence—the three major skills at which human labor has a comparative
advantage over machines and robots (Autor et al., 2003; Frey and Osborne, 2007)—the workers
exercise at work.
Using the skill index as the main explanatory variable, we examine, via multivariate
panel regressions, the extent to which retirement decisions are associated with the degree of
automation. We decompose the overall skill index into three sub-indices (i.e., creativity, non-
routineness, and social intelligence) to show the role played by each component. Our empirical
results indicate that workers’ retirement decisions are indeed influenced by the automatability of
workers’ skills used in their jobs. Compared to those who have higher automation-resistance, the
skilled workers, who can better resist the pressures of automation, plan to retire later: A one
standard deviation increase in the skill index leads to a 4.2% increase in the probability of
working full-time after age 62 from the average of 54.7%, and a 5.6% increase in the probability
of working full-time after age 65 from the average of 33.9%. Importantly, this tendency holds
true not only for workers’ retirement expectations, but also for their actual retirement behaviors,
as the workers with higher automation-resistance tend to exit the labor force at a later stage
compared to their counterparts with low resistance: A one standard deviation increase in the skill
index leads to a reduction in actual retirement in the subsequent wave by 6.4% from the average
of 14.0%. Such patterns are consistently displayed within the male- and female- subsamples of
workers.
Overall, this study lends sufficient evidence that workers who are ‘higher skilled’ in a
sense that their skills are more resistant to the pressures of automation are more likely to
postpone expected timing of retirement. Individuals’ heterogeneous levels of automation-
resistance generate inequalities in financial and nonfinancial returns to working, which is
10

dictated by workers’ automation-resistance. Based on the life cycle model of labor supply, we
interpret that such changes in the wage trajectories and uncertainties generate distinct retirement
patterns.
In witnessing rapid population aging and a shrinking tax base, which have important
consequences for the financial sustainability of Social Security and Medicare programs,
policymakers are interested in retaining older workers in the labor force by delaying retirement
(Lee, 2001; Lee and Skinner, 1999). This study sheds light on the mechanism through which
workers may be influenced by the pressures of automation and how this translates into retirement
behaviors. Findings of this chapter can potentially help policymakers garner a better
understanding of the challenges and opportunities posed by technology upon their promoting
longer working lives.

1.2 Conceptual Framework

In this section, we summarize the theory of skill-biased technological change (SBTC)
that serves as the conceptual basis of this study applied to the context of retirement. We refer to a
simplified theoretical model similar to those proposed by Card and DiNardo (2002) and Myck
and Reed (2006). In such model, there are only two groups of workers (i.e., skilled and unskilled)
in an economy. The aggregate labor demand is generated by a production function featuring
labor inputs by the skilled and the unskilled workers, with a constant elasticity of substitution
between the two labor inputs. Formally, the production function is written as follows:
(1) Y = 𝐴 [𝛼 (𝑔 𝑆 𝑁 𝑆 )
(𝜎 −1)
𝜎 ⁄
+ (1 − 𝑎 )(𝑔 𝑈 𝑁 𝑈 )
(𝜎 −1)
𝜎 ⁄
]
𝜎 (𝜎 −1)
⁄
.
11

Here, Y is the value of output, NS and NU are the labor inputs—in terms of employment or hours
worked—of skilled workers and unskilled workers respectively, and 𝜎 is the elasticity of
substitution between NS and NU such that 𝜎 ≥ 0. Lastly, A, 𝑎 , 𝑔 𝑆 and 𝑔 𝑈 are time-variant
technological parameters. In this context, the relative demand for the skilled workers is
determined by making the ratio of the marginal product of the skilled and the unskilled equal to
the ratio of their wages.
The evolution of relative wages between the two groups can be written by taking
logarithms and first-differencing over time:

(2) ∆ log (
𝑤 𝑆 𝑤 𝑈 ) = ∆ log [
𝛼 (1−𝑎 )
] + [
𝜎 −1
𝜎 ∆ log (
𝑔 𝑆 𝑔 𝑈 )] − [(
1
𝜎 ) ∆ log (
𝑁 𝑆 𝑁 𝑈 )].

Eq. (2) shows that changes in the relative wage must stem from either the (1) changes in the
relative supply of skilled workers or (2) changes in technology.
3
In equation (2), it should be
noted that the relative wage of skilled workers varies directly with their relative supply—in the
absence of technological change. Regardless of what the estimated value of 𝜎 (i.e., the elasticity
of substitution between NS and NU) is, an increase in the relative proportion of skilled workers
would correspond to a decrease in their relative wage holding constant the technology. If the
relative supply of skilled-workers were to be held constant, then the technological growths would
generate a wage differential—with the skilled workers’ relative wage being higher than that of
the unskilled counterpart.
Another key feature of this model is that it is the skill-biased nature of technology that
generates the wage differential between the skilled and the unskilled. A simple technological

3
Other features that influence relative wages such as premiums for efficiency or productivity have been ignored in
the model.
12

growth that is not skill-biased is understood to hold the same effect across all skill-levels,
resulting in a shift in the parameter α or a proportional shift in 𝑔 𝐻 and 𝑔 𝐿 and thus leaving the
relative productivity of the two skill groups unchanged. On the contrary, technological growths
that are ‘skill-biased’ differ from the skill-neutral technology in that the SBTC entails a relative
increase in the technological parameter of the skilled to that of the unskilled.
The literature provides an ample empirical evidence that the wage differentials between
the skilled and the unskilled were facilitated by the absence of growth in the relative size of the
skilled population. With the onset of technology and computer revolution, the demand for skilled
labor experienced a sharp rise in the 1980s. Despite the increased demand, the relative supply of
skilled workers stopped growing in the 1980s due to the slowing educational attainment of
cohorts born after 1949 as well as of fewer labor force entrants. Thus, lacking growth in the
relative supply of skilled-workers, the technology-induced wage gap kept widening (Goldin and
Katz, 2001). In that such a process persisted from 1990 to 2005 (Acemoglu and Autor, 2011;
Autor et al., 2008), it can be said that the wage differential between the skilled and the unskilled
generated by the SBTC is not a one-time event and instead a persistent phenomenon in the past
three or four decades.
Experiencing the higher wage growth compared to the unskilled, the skilled workers face
both income and substitution effects that alter their timing of retirement. The income effect
incentivizes the skilled to hasten the timing of retirement by increasing their demand of leisure
(i.e., time spent in retirement). In contrary, the substitution effect encourages the skilled to
postpone retirement as they, in seeing the rise of the opportunity cost of leisure, demand less of
it. In addition to the effects generated by income growths, it is plausible that additional non-
monetary factors stemming from SBTC may play a role in individuals’ retirement decisions. For
13

instance, it has been documented that skill-obsolescence and psychological distress affect
workers’ attachment to the labor force (Calvo, Haverstick, and Sass, 2009; Graebner, 1980;
Petkoska and Earl, 2009). The SBTC may induce lower levels of distress and increase the sense
of self-worth for the skilled individuals compared to their unskilled counterpart, thereby
generating a psychological ‘pull’ into the workforce away from retirement.
In the end, skilled workers’ hastening or postponing of retirement relative to the unskilled
in the face of the SBTC is the result of the combined wage- (i.e., income and substitution effects)
and psychological- effects, as well as of other effects unascertained in the current discussion.
This study provides empirical evidence of the overall, net effect of experienced SBTC on
retirement decisions across workers with different skills. The data at our disposal does not allow
us to disentangle the role of the various underlying mechanisms that generate different degrees of
labor force attachments across workers.

1.3 Study Data and Methods

Data is drawn from the HRS, a biennial survey representative of the American population
over the age of 50. From the HRS, we make use of information on respondents’ retirement plans
and expectations, basic demographics, family structure, employment history, health, financial
and housing wealth, and income. Also extracted are information on the institutional factors
known to influence individuals’ retirement decisions, including access to employer-sponsored
health insurance, Medicare eligibility, private pension, and Social Security (Aguila, 2014; Sood
et al., 2009). For this study, we use Version O of the RAND HRS for the waves 2004-2014.
14

We restrict our sample to individuals between 50 and 64 years of age, all of whom are
still working.
4
We investigate the relationship between skills and retirement decisions using
various outcome variables. Specifically, we consider three hypothetical retirement indicators,
namely planned retirement year, self-reported probabilities of working after age 62 and after age
65. We also rely on actual retirement decisions as observed across waves. In incorporating both
planned and actual retirement variables, we estimate the impact of automation not only for
individuals who either retire or change their labor supply behaviors during the observation
period, but also for aging workers approaching retirement.
In devising the skill index, we diverge from the previous studies evaluating the theory of
SBTC in the way we define workers’ skill levels. We devise the skill index from the
respondents’ career job information, thus calculating the automatability of the most
representative skill-sets of the HRS respondents. The higher the index of an individual is, the
more ‘skilled’ an individual is by the skills exercised at work. By linking the restricted HRS’s
occupation with the O*NET, we are able to assess the intensity of skills used by the HRS
respondents at work. The restricted HRS show occupations in the 4-digit Census Occupational
Classification System—ranging from 0010 to 9750—thereby providing a comprehensive
coverage of most jobs available in the US economy. The O*NET contains the task composition
of each occupation by weighing the ‘intensity’ of 42 tasks performed at the job. The assessment,
carried out by experts endorsed by the US Department of Labor, is based on how much such
skills are used, varying from zero (i.e., least used) to five (i.e., most used).
5
As an example, for

4
We also estimated our model on the sample of individuals aged 50-61, therefore excluding those subject to
retirement incentives provided by early Social Security claiming age. Overall, the results are similar to those
presented in this study.
5
Experts and researchers working for and with the O*NET Data Collection Program in the US Department of Labor
create the intensity-weights based on the following sources of information: Statistics available through the Bureau of
15

an average chief executive, the skill of ‘judging the qualities of things, services, and people’ is
identified as intensely used (4.35) while ‘repairing and maintaining mechanical equipment’
comprises one of the least intensely used (1.21) skills.
To identify the skills that help workers resist the pressures of SBTC from the rest that are
easily automatable, we refer to the literature and professional reports. According to the
McKinsey Global Institute’s Report (2017), skills that cannot be automated are identified as
those that demand workers’ exercise of (1) creativity, (2) nonroutineness, and (3) social
intelligence. On the contrary, automatable skills include data processing, data collection, and
predictable/routine tasks both in manual and non-manual jobs. Autor, Levy, and Murnane (2003)
as well as Frey and Osborne (2017) add further evidence that abstract and nonroutine skills are
immune to computerization. In this chapter, we select and group skills at which humans have a
comparative advantage over machines into three categories: (1) creativity, (2) non-routineness,
and (3) social intelligence. All tasks that comprise each category are shown in Table 1.1.
We sum the intensity-weights of all automation-resistant skills used in each job and
standardize it to create the main skill index. The index is designed so that it can be disaggregated
into three sub-indices based on the three categories:

(3) 𝐺 𝑡𝑜𝑡 𝑎𝑙 ,𝑖𝑡
= ∑ ( 𝐺 𝑐𝑟𝑒𝑎𝑡𝑖𝑣𝑒 ,𝑖𝑡
+ 𝐺 𝑛𝑜𝑛𝑟𝑜𝑢𝑡𝑖𝑛𝑒𝑛𝑒𝑠𝑠 ,𝑖𝑡
+ 𝐺 𝑠𝑜𝑐𝑖𝑎𝑙 ,𝑖𝑡
)
𝑛 𝑖 =1

Because the number of skills associated with each sub-category varies, we normalize each to
have a mean of zero and a standard deviation of one, with a higher value corresponding to higher
skill—and hence lower automatability, for easier comparisons.

Labor Statistics, CareerOneStop and Office of Apprenticeship within the Dept. of Labor and the Classification of
Instructional Programs within the US Department of Education.
16

Unlike the previous studies that have equated the ‘skilled’ workers as those with higher
education or in white-collar occupations (Autor et al., 1998; Card and DiNardo, 2002; Goldin
and Katz, 2001), our skill index is neutral to both. If a white-collar occupation requires the job
holders to utilize mostly predictable or repetitive skills, then the job is marked as an unskilled
occupation. In Table 1.2, we list some of the occupations with the lowest and the highest skill-
index scores. While the ‘skilled’ occupations that can best resist the threats of automation are
still mostly carried out by white-collar workers, the white- and blue-collar neutrality is more
apparent in the ‘unskilled’ category: it can be seen that what are traditionally considered as
white-collar professions (e.g., file clerks or payroll clerks) that perform office administrative
duties are highly automatable. It can also be inferred that the occupations in either category
require various extents of formal educational training.

1.4 Empirical Approach

Our estimation model for the expected retirement outcomes is as follows,

(4) 𝑌 𝑖𝑡𝑠 = 𝛾 𝐺 𝑖
+ 𝛽 𝑋 𝑖𝑡𝑠
+ 𝑐 𝑡 + 𝜔 𝑠 + 𝑢 𝑖𝑡𝑠 ,

where the subscripts i, t, and s, indicate individual, time and Census division, respectively. 𝑌 is
either distance from planned retirement year, or the self-reported probability of working full-time
after age 62 and after 65. The main explanatory variable of interest is 𝐺 𝑖
, a standardized total
skill index. We also run separate regressions using each of the total skill index sub-components,
that is, creativity, non-routineness, and social intelligence. 𝑋 𝑖𝑡
is a vector of individual
17

demographics and determinants of retirement, described in details below. The elements 𝑐 𝑡 and
𝜔 𝑠 represent year and Census division fixed effects, respectively. Lastly, 𝑢 𝑖𝑡𝑠 is an idiosyncratic
error term. We refrain from including individual fixed-effects in that our main explanatory
variable, the skill index, lacks within-individual variations over time.
6
Since individuals may be
represented in multiple observations from different waves, standard errors are clustered at the
individual level.
We adopt three different specifications depending on the set of explanatory variables
included in the vector X. In our baseline specification, we control for a wide array of
demographics and potential determinants of retirement. Specifically, the vector X features
gender, race, age and its quadratic polynomial, an indicator for age older than 62 (i.e., early
Social Security claiming age), marital status, educational attainment, and household size.
7
Also
included are individual earnings, household income less the respondent’s earnings, and
household wealth—all of which are converted into real dollars indexed to 2012 dollars and taken
natural logs.
8
We also include indicators for whether respondents currently have access to Social
Security benefits, are eligible for early Social Security benefits claiming by age (i.e. an identifier
for age 62 and above), and are contributing to employer-sponsored private pension plans by type:
defined-benefit (DB) plans, defined-contribution (DC) plans, both, or no access in each wave.
9

We capture health status via a health index that takes into account a number of critical, work-

6
Across the 10-year analysis period, there is very little changes in automatability of each occupation. In our skill
index, the within-occupation variations have a standard deviation of 0.
7
Race takes on a value 1 if Caucasians, 2 if African-Americans, and 3 if others. Marital statuses are categorized into
three: 1. If never married (reference), 2 if married or partnered, and 3 if separated, divorced, or widowed. Education
also has three groups: 1 if obtained GED or high school diploma (reference), 2 if obtained college degrees, and 3 if
obtained graduate degrees.
8
Net wealth includes the value of checking and savings accounts, bonds, stocks, deposits, mutual funds, primary
and secondary housing, and any other savings less debt. Income and wealth variables are entered in logs in the
regressions.
9
Individuals who are covered in the DB pension plans tend to face stronger early retirement incentives due to the
particular nature of the DB benefit accrual (Shoven and Slavov, 2014).
18

impeding medical conditions diagnosed by doctors: The index is normalized to have a mean of
zero and a standard deviation of one, with a higher value corresponding to better health.
10
If
individuals access health insurance through their employers, their retirement decisions are likely
influenced by whether or not—or for how long—they can keep the insurance in the post-
retirement years (Cutler and Zeckhauser, 2000; Nyce et al., 2013). Thus, we control for whether
individuals have insurance provided by employers during the active working-years (ESI) and for
whether they can keep it as retiree insurance benefits (RHB). Since retirement is likely a joint-
decision between married couples (Blau and Gilleskie, 2006), we use both respondents’ and their
spouses’ retirement expectations and construct indicators for whether the respondent expects to
retire earlier, at the same time or later than their spouse/partner.
By conditioning on all the aforementioned variables, we account for a variety of sources
that may induce individuals to take different retirement paths and estimate whether different
skills have an independent effect on retirement decisions, which could be attributed to
differential SBTC. Yet, a significant 𝛾 coefficient in equation (2) could simply result from
workers with different skills having systematically different preferences for leisure versus work.
The richness of the HRS data enables us to proxy for such unobserved preferences using
respondents’ work-related personality traits as elicited by the Leave Behind (LB) questionnaire.
11

We, therefore, amend our baseline specification, with five additional variables measuring the

10
The conditions include 1) high blood pressure or hypertension; 2) diabetes or high blood sugar; 3) cancer or a
malignant tumor of any kind except skin cancer; 4) chronic lung disease except asthma such as chronic bronchitis or
emphysema; 5) heart attack, coronary heart disease, angina, congestive heart failure, or other heart problems; 6)
stroke or transient ischemic attack (TIA); 7) emotional, nervous, or psychiatric problems; and 8) arthritis or
rheumatism.
11
Since 2004, the HRS interview includes the LB self-administered questionnaire, eliciting respondents’ subjective
well-being, life-style, and life events, and other psychosocial information. The LB is administrated to a rotating,
random 50% of the core HRS sample every other wave. Using the fact that the personality traits do not change
significantly for middle-aged and older individuals (Costa and McCrae, 1997; Costa et al., 2000), we calculate for
each HRS respondent the average personality trait scores over time and treat the score as constant in our analysis.
19

extent to which respondents rate themselves—on a 1-to-6 scale, with 6 indicating the highest
display of the trait—as organized, responsible, hardworking, intelligent, and thorough.
12

To gauge the effect of psychological motives to leave or remain in the labor force, we
estimate a third and richest specification. Alongside all the variables included in the second
specification, this specification features individuals’ varying extents of job satisfaction and age
discrimination at workplaces, both of which are shown to negatively influence individuals’
preferences for work (e.g., organizational commitment and work-withdrawal inclinations)
(Bibby, 2008; Ensher et al., 2001; Griffin et al., 2016; Macdonald and Levy, 2016). Job
satisfaction is measured using a 5-point scale index that represents the extent to which an
individual enjoys going to work. Perceived age discrimination at workplaces is measured using
two variables: whether age poses a hurdle for promotion opportunities and heightens retirement
pressures from employers.
Next, we move beyond planned retirement outcomes and focus an actual retirement
indicators: We construct a binary variable for retirement transitions that takes value 1 if an
individual works full-time in wave t and is fully retired in wave t+1 and 0 if an individual works
full-time in both waves t and t+1. We then estimate linear probability models and logit models,
and report average marginal effects with Delta Method standard errors clustered at the individual
level. All models are estimated for the entire sample as well as separately for men and women,
for whom retirement decisions may be determined by different factors.

12
Following the economic literature that has begun to use the five-factor model in psychology (Borghans et. al.,
2008; McCrae and John, 1992), we replaced these traits with the Big Five personality traits: openness to experience,
conscientiousness, extraversion, agreeableness, and neuroticism. The estimation results remain unaffected.
20

1.5 Results
1.5.1 Summary Statistics

Our sample consists of 11,880 person-year observations. The average number of
observations per individual across waves is three. Summary statistics of all the variables used in
the analysis are displayed in Table 1.3. In addition to the statistics for the entire sample, we
provide information on the male and female subsamples. To evaluate any systematic differences
between men and women included in the sample in terms of the covariates, we conduct t-tests.
The t-test results highlight that men and women vary in terms of their average skill levels—both
in terms of the overall score as well as the subcomponent (i.e., creativity, nonroutineness, and
social intelligence) scores—that dictate their level of automatability. Moreover, the subsamples
exhibit significant differences (i.e., at the 5% level) in terms of their average marital status,
educational attainment, individual earnings, access to employer-sponsored pension and health
insurance by type, current health, access to Social Security income, the extent of perceived age
discrimination at work, job satisfaction, and work-related personality traits. In the estimation
models, we control for such differences to examine whether there still exist differences in their
planned and actual retirement decisions induced by the automatability of men and women’s
skills.

1.5.2 Estimation Results
The regression results for the planned retirement outcomes are shown in Table 1.4. A
complete set of estimated coefficients is available in the appendix, Table 1.4A. Specification 1,
the baseline specification, includes demographics and determinants of retirement. Specification 2
includes all covariates from the Spec. 1 as well as the personality traits (i.e., proxying
21

unobserved preferences for work and leisure) as controls. Lastly, Specification 3 consists of the
richest set of controls as it includes all covariates from the Spec. 2 and additional controls for
work satisfaction and age discrimination. For columns (1)-(3), the outcome variable of interest is
the temporal distance to the planned timing of retirement. For columns (4)-(6) and columns (7)-
(9), the dependent variables are the self-reported probability (0-100) of working full-time after
age 62 and after age 65, respectively.
The estimated coefficient of the skill index in all said regressions show that higher skills
(i.e., lower automatability) of a worker is associated with longer (expected) working life. Using
the baseline specifications, or Spec. 1, a one standard deviation increase in the index is
associated with a postponement of retirement by 0.307 years, a 4.2% increase in the probability
of working full-time after age 62 from the average of 54.7%, and a 5.6% increase in the
probability of working full-time after age 65 from the average of 33.9%. The results are
statistically significant at 1% level. When we better account for the possible endogeneity induced
by individuals’ unobserved preferences for work and leisure in Spec. 2, we witness a slight
reduction of the estimated impact of the having higher automation-resistance (i.e., having a
higher skill index value) in columns (2), (5), and (8).
With the Spec. 3, we attempt to control for some of the psychological effect generated by
the SBTC. For all outcome variables, we observe that the estimated coefficients of the skill index
in Spec. 3 decreased from those using Spec. 2. It can be inferred here that the SBTC-induced
psychology indeed has a positive effect on the skilled as opposed to the unskilled—incentivizing
the skilled to lengthen their working lives. Among the covariates, age, gender, spouse retirement
expectation, household wealth, access to employer-sponsored health insurance, Social Security,
and private pension are significant predictors of retirement decisions across all specifications.
22

In Table 1.5, we present estimation results using the three planned retirement outcomes
separately for men and women. Spec. 3 from Table 1.4 is used for all. A full set of estimates on
covariates is available in the appendix, Table 1.5A. Regardless of gender and outcome variables,
it can be seen that the higher skills are once again associated with a postponement of the planned
timing of retirement. However, as shown in columns (1) and (2), one standard deviation increase
in the skill index is associated with a greater postponement of retirement by female workers than
male workers. Such a dynamic is reversed between men and women in columns (3)-(6): the
higher skill level seems to pose greater incentives to continue working full-time after ages 62 and
65 for men than women. It can be inferred from the results that while the female workers with
higher skills (i.e. higher automation-resistance) do plan on lengthening their working lives more
than their male counterparts, and they tend to plan on taking on part-time positions after ages 62
and 65.
In Table 1.6, we re-run the regressions by replacing the skill index with three sub-indices
that represent the degree of (1) creativity (2) non-routineness, and (3) social intelligence, all of
which contribute to immunizing the jobs from the pressures of automation. Detailed results are
shown in Table 1.6A. Throughout the models in Table 1.6, we use the distance to planned timing
of retirement as the outcome variable, and Spec. 3. The results in Table 1.6 clearly and
consistently illustrate that the more skilled and, hence, the more automation-resistant an
individual is—due to the degree of creativity, nonroutineness, and social skills he or she
exercises at work—the longer the individual plans to work compared to the unskilled
counterparts. Columns (1), (4), and (7) show that, for the total sample, a one standard-deviation
increase in the level of creativity, nonroutineness, and social intelligence is associated with an
increase of the distance to the planned retirement timing by 0.303 years, 0.281 years, and 0.289
23

years, respectively. The results are significant at 1% level. When the entire sample is under
inspection, it seems that the individuals’ level of creativity and social intelligence play stronger
roles than that of the non-routineness in shielding them and their jobs from the pressures of
automation. Between the two, creativity seems to be a slightly stronger influencer than social
intelligence. Both male and female subsamples emulate the results of the total sample.
Next, we re-run the regression models by replacing the planned retirement outcome with
a binary variable measuring individuals’ actual retirement decisions. Naturally, the sample for
the models using the actual retirement outcomes consists of individuals is different from the
sample used for estimations on planned retirement. In Table 1.7, the outcome variable takes on a
value 1 when an individual works in wave t and fully retires in t+1, and 0 when the individual
works in both waves t and t+1. For full results, see Table 1.7A. According to the linear
probability models (LPM) estimates shown in columns (1)-(3), the actual retirement is also
postponed with the rise in the skill index, with a one standard deviation increase in the skill index
leading to a reduction in actual retirement in the subsequent wave by 6.4% from the average of
14.0% (for men, 6.71% reduction from the average of 14.9%; for women, 5.19% from the
average of 13.5%). The marginal effects of the logit model estimates in columns (4)-(6) display
consistent results. With respect to the entire sample, it is shown that a one standard deviation
increase in the skill index is associated with 5.79% percentage point decrease in the chance that
the worker goes from working in wave t and fully retiring in t+1 from the average of 13.8%. In
other words, the higher skilled and more automation-resistant a worker is, the later she retires—
not merely by planning but by actually modifying her retirement behaviors in the subsequent
wave. This result is significant at 1% level. The results are consistently displayed both in terms
of direction and magnitude within the male- and female- subsamples: A one standard deviation
24

increase in the skill index is associated with a reduction in retiring in the subsequent wave by
6.21% from the mean of 14.5% for men and by 4.58% from the mean of 13.1% by women.
Unlike the results with the expected (i.e., planned) retirement outcomes, the results in Table 6 on
the actual retirement outcome suggest that male workers tend to lengthen their working lives
slightly more than their female counterparts do.

1.6 Discussion

As we suspected prior to launching the study, the process of automation not only
influences the labor market from the demand-side (i.e., initiating layoffs and structural
employment changes) but also has a direct impact on the labor supply behaviors in terms of their
work and retirement decisions. Overall, the panel regression results on both the expected and
actual timing of retirement consistently show that the skilled (i.e., automation-resistant) workers
retire later than their unskilled counterparts. However, in comparing male and female workers,
we observe a slight discrepancy between their retirement expectations and actual decisions. From
the results on planned retirement, it can be inferred that skilled female workers tend to plan on
lengthening their working lives more than skilled male workers—and, in doing so, take up part-
time positions after ages 62 and 65 as opposed to full-time jobs. However, from the results on
actual retirement, skilled male workers are shown to lengthen their working lives more than
skilled female counterparts. Moreover, from their actual retirement, it can be read that only the
skilled female workers are sufficiently disincentivized to transition from full-time to part-time
work due to the high automation-resistance of their skill-sets. While the dissonance between the
female workers’ retirement plans and actual behaviors is worth noting, we refrain from making
25

conclusive remarks on the implications of the observed discrepancy as we do not trace the same
individuals’ expected timing of retirement and their subsequent retirement behaviors over time.
From the disaggregation of the skill index into the three sub-indices, it can be inferred
that all three sub-indices influence the retirement decisions of the skilled workers in the same
direction: Workers who have high automation-resistance due to their exercising creativity,
nonroutineness, and social-intelligence in their jobs are more likely to postpone retirement both
in plans and in action than their counterparts with low resistance. Among the three, workers’
creativity and social-intelligence are shown to contribute more in raising automation-resistance
of individuals than does the ability to handle nonroutineness.
Linking the findings back to the discussion of the various effects associated with the
SBTC-induced psychology and wage growths in Section 1.5.2, the empirical results show that
the net effect of the SBTC encourages longer working lives and a postponement of retirement for
the skilled workers with high automation-resistance. From comparing results using different sets
of covariates, we observe a positive psychological effect on the parts of the skilled individuals:
In other words, the skilled individuals experience less discrimination and greater job-satisfaction
at work due to their high automation-resistance—which in turn results in their lengthening
working lives. However, even after controlling for the psychological- and wage- induced
retirement incentives in our models, we still witness the impact of SBTC remaining, continuing
to drive the skilled individuals to postpone retirement. The remaining effect could suggest that
our modeling of the psychological motives is insufficient. Future endeavors could focus on
devising a more comprehensive and nuanced index of workplace psychology that includes
factors other than job satisfaction and perceived discrimination to see if the remaining SBTC-
induced effect is reduced.
26

By revealing the opportunities and obstacles posed by the process of skill-biased
technology and the automation pressures it generates, the findings of this study add to the
discourse on how to effectively carry out the policy agenda to promote longer working lives. For
the policy-makers upholding the agenda to resolve the problem of population aging and a rapidly
shrinking tax base of the nation (Maestas and Zissimopoulos, 2010),
13
the fact that individuals
with higher automation-resistance lengthens their working lives compared to the low-resistance
counterparts can be a useful insight. A special attention ought to be paid to unskilled individuals
with low-automation resistance. To this particular group of individuals, promoting longer
working lives implies implementing more effective on-the-job training and/or retraining
programs that can help the workers enhance their adaptability to the skill-biased technological
growths. The focus of such training programs ought to be shifted to improving individuals’ soft
skills (e.g., creativity and social intelligence) that is shown in the literature to raise individuals’
automation-resistance as well as to raising individuals’ awareness of the value of their soft skills
in their race against the machines.
Lastly, it should be acknowledged that the notion of ‘retirement’ in this chapter is defined
by not mere exits from the current jobs—where there is a possibility for the individual to
transition to other jobs—but a transition into full retirement. Assuming that, for most individuals,
the full retirement implies a permanent end of their working lives, the findings of this study shed
lights on how the pressures of automation influence the duration of the working lives of the
aging labor force in the US. However, such an assumption may be too strict as a growing number
of retirees are choosing to return to the labor force in the recent decades (Berkovec and Stern,
1991; Cahill et al., 2010; Kanabar, 2012; Platts and Glaser, 2017). In order to better comment on

13
Past efforts include termination of the Social Security earnings test and a repeal of the mandatory retirement.
(Gustman et al., 2016).
27

how the pressures of automation and individuals’ level of automation-resistance influence the
duration of their working lives, explored are the association between automation and
unretirement practices in Chapter 2.

28

Table 1.1. Un-Automatable Skills
(1) Creativity and Expertise
Thinking Creatively
Developing Objectives and Strategies
(2) Unpredictability and Non-Routineness
Implementing Ideas, Programs, etc.
Analyzing Data or Information
Evaluating Information Against Standards
Interpreting the Meaning of Information for Others
Making Decisions and Solving Problems
Estimating Needed Characteristics
(3) Social Intelligence
Communicating with Supervisors, Peers, or Subordinates
Communicating with Persons Outside Organization
Establishing and Maintaining Interpersonal Relationships
Assisting and Caring for Others
Resolving Conflicts and Negotiations with Others
Provide Consultation and Advice to Others
Developing and Building Teams
Training and Teaching Others
Guiding, Directing, and Motivating Subordinates
Coaching and Developing Others
Judging the Qualities of Things, Services, or People
Source: Authors' inpsection of the O*NET data.
Table 1.2. Examples of Skilled and Unskilled Occupations
Automatable Occupations ("unskilled")
food preparation and serving related workers
mining machine operator
dental assistant
file clerks
clergy
payroll and timekeeping clerks
cashiers
Unautomatable Occupations ("skilled")
real estate managers
social scientists
medical and health services managers
counselors
secondary school teachers
engineering technicians
registered nurses
Source: Author's calculation using O*NET and HRS.
29

Table 1.3. Summary Statistics
Men Women t-test
Mean StDev Mean Mean p<0.05
time until the planned retirement 7.87 5.07 7.92 7.83
probability of working until age 62 33.86 33.94 36.47 32.25 +
probability of working until age 65 54.75 36.49 58.97 52.14 +
actual retirement (binary)* 0.13 0.34 0.14 0.13
skill index, standardized 0.39 1.24 0.50 0.32 +
sub-index 1: creativity, standardized 0.38 1.24 0.50 0.30 +
sub-index 2: nonroutineness, standardized 0.38 1.23 0.50 0.31 +
sub-index 3: social intelligence, standardized 0.40 1.24 0.49 0.33 +
age 56.72 3.31 57.02 56.53 +
male 0.38 0.49 1.00 0.00
race 1.33 0.62 1.33 1.33
marital status 2.14 0.46 2.06 2.19 +
educational attainment 1.43 0.69 1.45 1.42 +
household size 2.51 1.23 2.63 2.44 +
ln(income) 9.37 4.54 9.49 9.29
ln(HH income, less R's income) 7.77 6.20 7.78 7.76
ln(wealth) 11.38 3.41 11.35 11.40
spouse retirement expectations 2.70 0.53 2.74 2.67
pension type 2.15 1.13 2.11 2.18 +
health status 0.37 0.80 0.42 0.34 +
employer-sponsored health insurance 2.00 0.80 2.09 1.94 +
access to Social Security income 0.34 0.47 0.32 0.34 +
eligibility to early Social Security claims (age 62+) 0.07 0.26 0.08 0.07
wave 9.54 1.70 9.57 9.53
census division 5.04 2.40 5.07 5.02
age discrimination: promotion unlikely 2.07 0.73 2.11 2.04 +
age discrimination: retirement pressure 1.97 0.67 2.01 1.95 +
enjoyworking 3.10 0.65 3.07 3.12 +
personality: organized 3.13 0.78 3.04 3.19 +
personality: responsible 3.79 0.42 3.71 3.84 +
personality: hardworking 3.75 0.45 3.69 3.79 +
personality: intelligent 3.40 0.58 3.36 3.43
personality: thorough 3.21 0.69 3.10 3.28
N 11880 4539 7341
Source: Author's calculation using the HRS.
Total
Notes: The indicator variable for individuals aged 62 and older discerns eligibility for--but not necessarily receiving--the early Social Security
clams. For binary actual retirement (*), the count is 988 3 for total, 3766 for men, and 6117 for women. For multiple-outcome retirement (**),
count is 8034 for total, 3390 for men, and 4644 for women.
30

Table 1.4. Estimating the Impact of Automatability on Planned Retirement
(1) (2) (3) (4) (5) (6) (7) (8) (9)
Spec. 1 Spec. 2 Spec. 3 Spec. 1 Spec. 2 Spec. 3 Spec. 1 Spec. 2 Spec. 3
skill index 0.307*** 0.294*** 0.290*** 0.023*** 0.022*** 0.021*** 0.019*** 0.018*** 0.018***
(0.049) (0.049) (0.049) (0.360) (0.360) (0.358) (0.344) (0.343) (0.341)
baseline covariates Y Y Y Y Y Y Y Y Y
personality controls N Y Y N Y Y N Y Y
work satisfaction controls N N Y N N Y N N Y
age discrimination controls N N Y N N Y N N Y
dependent variable mean 7.866 7.866 7.866 0.547 0.547 0.547 0.339 0.339 0.339
observations 11880 11880 11880 11880 11880 11880 11880 11880 11880
R-squared 0.348 0.352 0.357 0.103 0.111 0.118 0.074 0.080 0.089
Source: Author's calculations using the HRS.
Y=time left till planned retirement
Y=expected probability of working at
age 62
Y=expected probability of working at
age 65
Note: Year- and state- fixed effects are included in the regression equations but not reported in the table. Standard errors are in parentheses. If included in the specification,
items are marked with a "Y". If not, "N" is listed. *p<0.1 **p<0.05 ***p<0.01
31

Table 1.5. Impact of Automatability on Planned Retirement, by Gender
(1) (2) (3) (4) (5) (6)
men women men women men women
skill index 0.229*** 0.315*** 0.023*** 0.019*** 0.016*** 0.017***
(0.076) (0.065) (0.533) (0.482) (0.516) (0.455)
baseline covariates Y Y Y Y Y Y
personality controls Y Y Y Y Y Y
work satisfaction controls Y Y Y Y Y Y
age discrimination controls Y Y Y Y Y Y
dependent variable mean 7.925 7.831 0.589 0.521 0.365 0.322
observations 4539 7341 4539 7341 4539 7341
R-squared 0.359 0.366 0.113 0.128 0.101 0.091
Y=time left till planned retirement
Y= expected probability of working
at age 62
Y=expected probability of working at
age 65
Note: Year- and state- fixed effects are included in the regression equations but not reported in the table. Standard errors are in parentheses. Controlled are baseline
covariaes, personality traits, and job satisfaction/ age discrimination. If included in the specification, items are marked with a "Y". If not, "N" is listed. *p<0.1
**p<0.05 ***p<0.01
Source: Author's calculations using the HRS.
32

Table 1.6. Impact of Automatability on Planned Retirement, by Index Components
(1) (2) (3) (4) (5) (6) (7) (8) (9)
total men women total men women total men women
subindex 0.303*** 0.246*** 0.327*** 0.281*** 0.216*** 0.312*** 0.289*** 0.230*** 0.311***
(0.050) (0.076) (0.067) (0.050) (0.075) (0.066) (0.049) (0.076) (0.064)
baseline covariates Y Y Y Y Y Y Y Y Y
personality controls Y Y Y Y Y Y Y Y Y
work satisfaction controls Y Y Y Y Y Y Y Y Y
age discrimination controls Y Y Y Y Y Y Y Y Y
dependent variable mean 7.866 7.925 7.831 7.866 7.925 7.831 7.866 7.925 7.831
observations 11880 4539 7341 11880 4539 7341 11880 4539 7341
R-squared 0.001 0.359 0.366 0.001 0.359 0.365 0.001 0.359 0.365
Source: Author's calculations using the HRS.
social intelligence
Note: Year- and state- fixed effects are included in the regression equations but not reported in the table. Standard errors are in parentheses. Controlled are baseline covariaes,
personality traits, and job satisfaction/ age discrimination. If included in the specification, items are marked with a "Y". If not, "N" is listed. *p<0.1 **p<0.05 ***p<0.01
creativity nonroutineness Y=time left until planned
retirement
33

Table 1.7. Automatability and Actual Retirement: A Binary-Outcome Regression
(1) (2) (3) (4) (5) (6)
all men women all men women
skill index -0.009*** -0.010*** -0.007** -0.008*** -0.009** -0.006**
(0.003) (0.004) (0.004) (0.002) (0.004) (0.003)
baseline covariates Y Y Y Y Y Y
personality controls Y Y Y Y Y Y
work satisfaction controls Y Y Y Y Y Y
age discrimination controls Y Y Y Y Y Y
dependent variable mean 0.140 0.149 0.135 0.138 0.145 0.131
observations 14577 6047 8530 14577 6047 8530
Source: Authors' calculations using the HRS.
Logit, marginal effects
Note: Year- and state- fixed effects are included in the regression equations but not reported in the table. Standard errors are in
parentheses. Controlled are baseline covariaes, personality traits, and job satisfaction/ age discrimination. The outcome
variable for all columns is a binary actual retirement that takes on a value 1 when working in time t and fully retiring in t+1, and 0
if working both in time t and t+1. If included in the specification, items are marked with a "Y". If not, "N" is listed. *p<0.1
**p<0.05 ***p<0.01
Linear Probability Models
34

1.7 Appendix

Table 1.4A. Estimating the Impact of Automatability on Planned Retirement (detailed)
(1) (2) (3) (4) (5) (6) (7) (8) (9)
Spec. 1 Spec. 2 Spec. 3 Spec. 1 Spec. 2 Spec. 3 Spec. 1 Spec. 2 Spec. 3
skill index 0.307*** 0.294*** 0.290*** 0.023*** 0.022*** 0.021*** 0.019*** 0.018*** 0.018***
(0.049) (0.049) (0.049) (0.360) (0.360) (0.358) (0.344) (0.343) (0.341)
age -1.379*** -1.366*** -1.333*** -22.639*** -22.554*** -22.156*** -12.009*** -11.877*** -11.418***
(0.338) (0.338) (0.337) (3.299) (3.294) (3.290) (3.104) (3.101) (3.089)
age-squared 0.005* 0.005* 0.005 0.217*** 0.216*** 0.212*** 0.117*** 0.116*** 0.111***
(0.003) (0.003) (0.003) (0.029) (0.029) (0.029) (0.028) (0.028) (0.028)
race (ref=1.Caucasian)
2. African-American -0.458*** -0.599*** -0.575*** -9.108*** -10.158*** -9.894*** -8.513*** -9.404*** -9.101***
(0.157) (0.160) (0.159) (1.237) (1.252) (1.241) (1.126) (1.141) (1.129)
3. Other -0.408* -0.367 -0.344 -6.367*** -5.950*** -5.671*** -5.745*** -5.424*** -5.081***
-0.224 (0.224) (0.224) (1.621) (1.614) (1.616) (1.480) (1.479) (1.475)
education (ref=1.highschool)
2. college degrees 0.271* 0.181 0.173 3.416*** 2.535** 2.450** 3.182*** 2.541** 2.437**
(0.155) (0.156) (0.155) (1.175) (1.179) (1.172) (1.098) (1.100) (1.092)
3. graduate degrees 0.179 0.032 -0.010 3.262** 1.920 1.471 3.922*** 2.927** 2.437*
-0.191 -0.192 (0.191) (1.478) (1.496) (1.486) (1.388) (1.403) (1.388)
male 0.633*** 0.616*** 0.631*** 6.498*** 6.695*** 6.862*** 3.953*** 4.010*** 4.198***
(0.123) (0.124) (0.124) (0.940) (0.953) (0.947) (0.880) (0.888) (0.883)
marital status (ref=1.never married)
2. married/partnered -1.412*** -1.428*** -1.408*** -3.590* -3.762* -3.615* -5.412** -5.554*** -5.388**
(0.303) (0.304) (0.303) (2.079) (2.083) (2.069) (2.116) (2.114) (2.102)
3. separated/divorced/widowed -0.794** -0.826*** -0.817*** 2.447 2.041 2.107 -0.473 -0.800 -0.707
(0.311) (0.312) (0.311) (2.118) (2.122) (2.107) (2.173) (2.174) (2.160)
spouse ret expectation (ref=1. retire at the same time)
2. spouse to retire earlier 1.980*** 1.978*** 1.959*** 6.434*** 6.447*** 6.250*** 8.194*** 8.180*** 7.969***
(0.173) (0.172) (0.171) (1.627) (1.625) (1.628) (1.500) (1.499) (1.497)
3. spouse to retire later 0.152 0.150 0.158 1.400 1.401 1.455 2.959** 2.934** 2.986**
(0.159) (0.159) (0.158) (1.578) (1.577) (1.580) (1.413) (1.412) (1.410)
household size 0.070** 0.069** 0.068** 0.595** 0.581* 0.563* 0.404 0.385 0.364
(0.031) (0.030) (0.031) (0.301) (0.299) (0.300) (0.305) (0.302) (0.301)
ln(income) -0.015* -0.015* -0.015* 0.209*** 0.202*** 0.212*** 0.096 0.089 0.099
(0.008) (0.008) (0.008) (0.067) (0.067) (0.067) (0.063) (0.063) (0.063)
ln(other HH income) -0.001 -0.001 -0.001 -0.048 -0.051 -0.059 -0.160*** -0.161*** -0.170***
(0.007) (0.007) (0.007) (0.063) (0.062) (0.062) (0.062) (0.062) (0.062)
ln(wealth) -0.048*** -0.048*** -0.049*** -0.352*** -0.374*** -0.377*** -0.440*** -0.454*** -0.458***
(0.013) (0.013) (0.013) (0.121) (0.122) (0.124) (0.110) (0.110) (0.110)
employer pension type (ref=4. access to none)
1. has DB -1.216*** -1.226*** -1.200*** 2.793** 2.582** 2.881*** -1.943* -2.109** -1.785*
(0.139) (0.139) (0.137) (1.119) (1.117) (1.113) (1.034) (1.033) (1.030)
2. has DC -0.771*** -0.784*** -0.761*** 6.548*** 6.321*** 6.555*** 1.992** 1.809* 2.050**
(0.135) (0.135) (0.133) (1.061) (1.058) (1.057) (0.996) (0.993) (0.990)
3. both DC and DB -1.132*** -1.150*** -1.118*** 3.457* 3.164 3.485* -0.179 -0.442 -0.094
(0.193) (0.193) (0.190) (1.946) (1.936) (1.931) (1.769) (1.758) (1.752)
health 0.234*** 0.251*** 0.231*** 1.258** 1.385*** 1.174** 0.576 0.723 0.502
(0.073) (0.074) (0.074) (0.528) (0.533) (0.530) (0.496) (0.499) (0.497)
health insurance (ref=1. no ESI, no RHI)
2. ESI, but no RHI -0.248*** -0.250*** -0.232** 6.066*** 6.032*** 6.280*** 2.332*** 2.297*** 2.557***
(0.096) (0.096) (0.095) (0.878) (0.876) (0.876) (0.845) (0.843) (0.839)
3. has both ESI and RHI -0.447*** -0.449*** -0.439*** 0.825 0.770 0.935 0.239 0.192 0.368
(0.100) (0.100) (0.100) (0.924) (0.922) (0.922) (0.901) (0.898) (0.896)
access to Social Security income -1.033*** -1.019*** -1.012*** -10.443*** -10.234*** -10.214*** -9.654*** -9.493*** -9.471***
(0.157) (0.157) (0.156) (1.274) (1.268) (1.262) (1.146) (1.144) (1.138)
eligibility to early Social Security claims (age 62+) 0.169 0.172 0.167 -4.483*** -4.440*** -4.541*** -0.519 -0.488 -0.594
(0.124) (0.124) (0.124) (1.116) (1.112) (1.116) (1.389) (1.388) (1.387)
enjoy going to work - - 0.265*** - - 2.751*** - - 2.832***
- - (0.051) - - (0.517) - - (0.476)
age discrimination: promotion - - -0.051 - - -0.650 - - -0.648
- - (0.052) - - (0.499) - - (0.465)
age discrimination: retirement pressure - - -0.117** - - -1.212** - - -1.510***
- - (0.050) - - (0.534) - - (0.486)
personality: organized - -0.350*** -0.359*** - -3.154*** -3.248*** - -2.969*** -3.068***
- (0.090) (0.089) - (0.670) (0.667) - (0.628) (0.624)
personality: responsible - -0.392** -0.405*** - -1.024 -1.167 - -1.518 -1.672
- (0.155) (0.156) - (1.194) (1.187) - (1.095) (1.090)
personality: hardworking - 0.346*** 0.336*** - 2.289** 2.198** - 2.558*** 2.481***
- (0.127) (0.126) - (1.073) (1.066) - (0.965) (0.960)
personality: intelligent - 0.533*** 0.506*** - 4.237*** 3.946*** - 3.457*** 3.139***
- (0.111) (0.111) - (0.862) (0.860) - (0.801) (0.799)
personality: thorough - 0.124 0.115 - 1.898** 1.793** - 1.452* 1.331*
- (0.109) (0.108) - (0.795) (0.792) - (0.745) (0.741)
constant 71.798*** 70.473*** 69.295*** 640.257*** 622.908*** 609.388*** 345.506*** 330.859*** 316.151***
(9.513) (9.498) (9.466) (92.181) (92.105) (92.003) (86.976) (86.864) (86.472)
dependent variable mean 7.866 7.866 7.866 0.547 0.547 0.547 0.339 0.339 0.339
observations 11880 11880 11880 11880 11880 11880 11880 11880 11880
R-squared 0.348 0.352 0.357 0.103 0.111 0.118 0.074 0.080 0.089
Note: Year- and state- fixed effects are included in the regression equations but not reported in the table. Standard errors are in parentheses. *p<0.1 **p<0.05 ***p<0.01
Source: Authors' calculations using the HRS.
Y=expected probability of working at age
65
Y=time left till planned retirement
Y=expected probability of working at age
62
35

Table 1.5A. Impact of Automatability on Planned Retirement, by Gender (detailed)
(1) (2) (3) (4) (5) (6)
men women men women men women
skill index 0.229*** 0.315*** 0.023*** 0.019*** 0.016*** 0.017***
(0.076) (0.065) (0.533) (0.482) (0.516) (0.455)
age -1.756*** -1.087*** -24.906*** -20.509*** -18.961*** -7.517**
(0.580) (0.416) (5.445) (4.187) (5.451) (3.820)
age-squared 0.009* 0.003 0.237*** 0.198*** 0.178*** 0.076**
(0.005) (0.004) (0.048) (0.037) (0.048) (0.034)
race (ref=1.Caucasian)
2. African-American -0.284 -0.770*** -6.686*** -12.046*** -6.353*** -10.973***
(0.282) (0.190) (2.159) (1.532) (1.975) (1.400)
3. Other -0.025 -0.634** -3.354 -7.906*** -3.242 -6.582***
(0.339) (0.294) (2.340) (2.239) (2.203) (2.023)
education (ref=1.highschool)
2. college degrees 0.132 0.184 3.199* 1.576 4.484** 0.704
(0.258) (0.192) (1.806) (1.536) (1.752) (1.389)
3. graduate degrees 0.312 -0.283 4.974** -1.260 7.760*** -1.576
(0.279) (0.258) (2.212) (1.985) (2.155) (1.811)
marital status (ref=1.never married)
2. married/partnered -0.693 -1.771*** -1.284 -4.899* -4.507 -6.144**
(0.455) (0.396) (3.313) (2.633) (3.392) (2.668)
3. separated/divorced/widowed -0.209 -1.151*** 2.763 1.077 -1.354 -1.262
(0.513) (0.392) (3.598) (2.596) (3.660) (2.676)
spouse ret expectation (ref=1. retire at the same tme)
2. spouse to retire earlier 1.933*** 2.009*** 6.599*** 6.242*** 7.098*** 8.881***
(0.244) (0.231) (2.244) (2.232) (2.355) (1.934)
3. spouse to retire later -0.123 0.383* 1.412 1.639 0.480 4.836***
(0.215) (0.221) (2.230) (2.157) (2.203) (1.829)
household size 0.083 0.048 0.541 0.542 0.650 0.138
(0.051) (0.038) (0.465) (0.388) (0.466) (0.394)
ln(income) -0.004 -0.023** 0.068 0.314*** -0.029 0.201**
(0.013) (0.010) (0.102) (0.088) (0.101) (0.080)
ln(other HH income) 0.002 -0.002 -0.037 -0.072 -0.150 -0.183**
(0.011) (0.009) (0.094) (0.083) (0.096) (0.081)
ln(wealth) -0.029 -0.060*** -0.024 -0.575*** -0.236 -0.578***
(0.022) (0.015) (0.175) (0.167) (0.155) (0.152)
employer pension type (ref=4. access to none)
1. has DB -1.527*** -1.012*** 0.807 4.018*** -4.615*** -0.100
(0.215) (0.175) (1.881) (1.387) (1.658) (1.314)
2. has DC -1.084*** -0.577*** 5.739*** 6.845*** 1.687 2.140*
(0.205) (0.172) (1.741) (1.329) (1.569) (1.268)
3. both DC and DB -1.425*** -0.976*** 1.045 4.562* -4.456 2.321
(0.316) (0.242) (2.900) (2.579) (2.775) (2.248)
health 0.290* 0.199*** 1.513* 0.970 0.468 0.539
(0.148) (0.077) (0.857) (0.670) (0.813) (0.628)
health insurance (ref=1. no ESI, no RHI)
2. ESI, but no RHI -0.371** -0.144 2.063 8.854*** -0.347 4.350***
(0.152) (0.122) (1.423) (1.116) (1.348) (1.078)
3. has both ESI and RHI -0.492*** -0.411*** -2.834* 3.391*** -1.837 1.800
(0.165) (0.125) (1.477) (1.188) (1.444) (1.156)
access to Social Security income -0.968*** -1.042*** -9.607*** -10.504*** -9.932*** -9.022***
(0.265) (0.192) (2.050) (1.599) (1.881) (1.429)
eligibility to early Social Security claims (age 62+) -0.022 0.250* -6.212*** -3.754** -1.924 -0.101
(0.222) (0.145) (1.682) (1.500) (2.153) (1.828)
enjoy going to work 0.234*** 0.271*** 2.244*** 2.955*** 2.260*** 3.055***
(0.081) (0.066) (0.824) (0.666) (0.771) (0.610)
age discrimination: promotion -0.133 -0.015 -2.061*** 0.110 -1.754** -0.060
(0.090) (0.065) (0.784) (0.648) (0.744) (0.597)
age discrimination: retirement pressure -0.216*** -0.058 -0.546 -1.657** -1.175 -1.691***
(0.083) (0.062) (0.814) (0.703) (0.761) (0.632)
personality: organized -0.568*** -0.232** -4.026*** -2.782*** -3.793*** -2.603***
(0.151) (0.109) (1.074) (0.844) (1.021) (0.785)
personality: responsible -0.137 -0.707*** 0.199 -2.151 -1.981 -1.066
(0.221) (0.224) (1.592) (1.794) (1.514) (1.583)
personality: hardworking 0.191 0.453*** 1.370 2.811* 2.256 2.689**
(0.194) (0.166) (1.569) (1.465) (1.439) (1.303)
personality: intelligent 0.806*** 0.318** 5.460*** 2.964*** 4.491*** 2.314**
(0.180) (0.140) (1.349) (1.120) (1.286) (1.026)
personality: thorough 0.102 0.147 1.181 2.203** 1.204 1.465
(0.178) (0.137) (1.267) (1.017) (1.235) (0.927)
constant 80.709*** 63.522*** 692.862*** 565.388*** 538.092*** 203.080*
(16.442) (11.664) (153.180) (116.757) (153.514) (106.747)
dependent variable mean 7.925 7.831 0.589 0.521 0.365 0.322
observations 4539 7341 4539 7341 4539 7341
R-squared 0.359 0.366 0.113 0.128 0.101 0.091
Y=time left till planned
retirement
Note: Year- and state- fixed effects are included in the regression equations but not reported in the table. Standard errors are in parentheses. Controlled are
baseline covariaes, personality traits, and job satisfaction/ age discrimination. *p<0.1 **p<0.05 ***p<0.01
Source: Authors' calculations using the HRS.
Y=expected probability of
working at age 65
Y= expected probability of
working at age 62
36

Table 1.6A. Impact of Automatability on Planned Retirement, by Index Components (detailed)
(1) (2) (3) (4) (5) (6) (7) (8) (9)
total men women total men women total men women
subindex 0.303*** 0.246*** 0.327*** 0.281*** 0.216*** 0.312*** 0.289*** 0.230*** 0.311***
(0.050) (0.076) (0.067) (0.050) (0.075) (0.066) (0.049) (0.076) (0.064)
age -1.332*** -1.757*** -1.087*** -1.333*** -1.757*** -1.088*** -1.333*** -1.756*** -1.088***
(0.337) (0.580) (0.416) (0.337) (0.580) (0.416) (0.337) (0.580) (0.416)
age-squared 0.005 0.009* 0.003 0.005 0.009* 0.003 0.005 0.009* 0.003
(0.003) (0.005) (0.004) (0.003) (0.005) (0.004) (0.003) (0.005) (0.004)
race (ref=1.Caucasian)
2. African-American -0.569*** -0.274 -0.767*** -0.576*** -0.288 -0.770*** -0.576*** -0.286 -0.771***
(0.159) (0.282) (0.190) (0.159) (0.282) (0.190) (0.159) (0.282) (0.190)
3. Other -0.338 -0.018 -0.629** -0.346 -0.025 -0.637** -0.345 -0.027 -0.634**
(0.223) (0.339) (0.294) (0.224) (0.339) (0.295) (0.224) (0.339) (0.294)
education (ref=1.highschool)
2. college degrees 0.161 0.118 0.176 0.170 0.134 0.179 0.176 0.134 0.188
(0.155) (0.258) (0.192) (0.155) (0.258) (0.192) (0.155) (0.258) (0.192)
3. graduate degrees -0.031 0.290 -0.303 -0.011 0.310 -0.283 -0.006 0.317 -0.281
(0.191) (0.278) (0.257) (0.191) (0.279) (0.258) (0.191) (0.279) (0.258)
male 0.619*** - - 0.628*** - - 0.635*** - -
(0.124) - - (0.124) - - (0.124) - -
marital status (ref=1.never married)
2. married/partnered -1.411*** -0.709 -1.765*** -1.411*** -0.696 -1.773*** -1.406*** -0.689 -1.771***
(0.303) (0.456) (0.396) (0.303) (0.455) (0.396) (0.303) (0.456) (0.396)
3. separated/divorced/widowed -0.822*** -0.224 -1.149*** -0.819*** -0.213 -1.152*** -0.815*** -0.204 -1.152***
(0.311) (0.513) (0.392) (0.311) (0.513) (0.392) (0.311) (0.513) (0.392)
spouse ret expectation (ref=1. retire at the same time)
2. spouse to retire earlier 1.958*** 1.931*** 2.009*** 1.959*** 1.934*** 2.009*** 1.959*** 1.933*** 2.009***
(0.171) (0.244) (0.231) (0.171) (0.244) (0.231) (0.171) (0.244) (0.231)
3. spouse to retire later 0.158 -0.123 0.383* 0.159 -0.122 0.384* 0.158 -0.123 0.383*
(0.158) (0.215) (0.221) (0.158) (0.215) (0.221) (0.158) (0.215) (0.221)
household size 0.068** 0.084* 0.048 0.068** 0.083 0.049 0.068** 0.083 0.048
(0.031) (0.051) (0.038) (0.031) (0.051) (0.038) (0.031) (0.051) (0.038)
ln(income) -0.015* -0.004 -0.023** -0.015* -0.004 -0.023** -0.015* -0.004 -0.023**
(0.008) (0.013) (0.010) (0.008) (0.013) (0.010) (0.008) (0.013) (0.010)
ln(other HH income) -0.001 0.002 -0.002 -0.001 0.002 -0.002 -0.001 0.002 -0.002
(0.007) (0.011) (0.009) (0.007) (0.011) (0.009) (0.007) (0.011) (0.009)
ln(wealth) -0.049*** -0.029 -0.060*** -0.049*** -0.029 -0.060*** -0.049*** -0.029 -0.060***
(0.013) (0.022) (0.015) (0.013) (0.022) (0.015) (0.013) (0.022) (0.015)
employer pension type (ref=4. access to none)
1. has DB -1.201*** -1.527*** -1.013*** -1.200*** -1.528*** -1.012*** -1.199*** -1.527*** -1.011***
(0.137) (0.215) (0.175) (0.137) (0.215) (0.175) (0.137) (0.215) (0.175)
2. has DC -0.762*** -1.085*** -0.578*** -0.761*** -1.084*** -0.578*** -0.760*** -1.084*** -0.576***
(0.133) (0.205) (0.172) (0.133) (0.205) (0.172) (0.133) (0.205) (0.172)
3. both DC and DB -1.118*** -1.424*** -0.977*** -1.118*** -1.427*** -0.977*** -1.117*** -1.425*** -0.975***
(0.190) (0.316) (0.242) (0.190) (0.316) (0.242) (0.190) (0.316) (0.242)
health 0.231*** 0.289* 0.201*** 0.231*** 0.289* 0.200*** 0.231*** 0.290* 0.199***
(0.074) (0.148) (0.077) (0.074) (0.148) (0.077) (0.074) (0.148) (0.077)
health insurance (ref=1. no ESI, no RHI)
2. ESI, but no RHI -0.232** -0.371** -0.144 -0.233** -0.372** -0.144 -0.232** -0.370** -0.143
(0.095) (0.152) (0.122) (0.095) (0.152) (0.122) (0.095) (0.152) (0.122)
3. has both ESI and RHI -0.438*** -0.491*** -0.409*** -0.441*** -0.494*** -0.412*** -0.439*** -0.491*** -0.411***
(0.100) (0.165) (0.125) (0.100) (0.165) (0.125) (0.100) (0.165) (0.125)
access to Social Security income -1.012*** -0.967*** -1.044*** -1.016*** -0.973*** -1.044*** -1.010*** -0.967*** -1.041***
(0.156) (0.265) (0.192) (0.156) (0.265) (0.192) (0.156) (0.265) (0.192)
eligibility to early Social Security claims (age 62+) 0.168 -0.021 0.251* 0.166 -0.023 0.250* 0.167 -0.021 0.251*
(0.124) (0.222) (0.145) (0.124) (0.222) (0.145) (0.124) (0.222) (0.145)
enjoy going to work 0.265*** 0.234*** 0.271*** 0.265*** 0.234*** 0.272*** 0.265*** 0.233*** 0.271***
(0.051) (0.081) (0.066) (0.051) (0.081) (0.066) (0.051) (0.081) (0.066)
age discrimination: promotion -0.051 -0.133 -0.014 -0.051 -0.133 -0.015 -0.051 -0.133 -0.014
(0.052) (0.090) (0.065) (0.052) (0.090) (0.065) (0.052) (0.090) (0.065)
age discrimination: retirement pressure -0.116** -0.215*** -0.057 -0.117** -0.217*** -0.057 -0.117** -0.216*** -0.058
(0.050) (0.083) (0.062) (0.050) (0.083) (0.062) (0.050) (0.083) (0.062)
personality: organized -0.359*** -0.568*** -0.231** -0.358*** -0.566*** -0.231** -0.360*** -0.569*** -0.232**
(0.089) (0.150) (0.109) (0.089) (0.151) (0.109) (0.089) (0.150) (0.109)
personality: responsible -0.406*** -0.137 -0.709*** -0.408*** -0.140 -0.711*** -0.404*** -0.136 -0.706***
(0.156) (0.221) (0.223) (0.156) (0.221) (0.224) (0.156) (0.221) (0.224)
personality: hardworking 0.336*** 0.191 0.450*** 0.338*** 0.192 0.457*** 0.336*** 0.190 0.452***
(0.126) (0.194) (0.166) (0.126) (0.194) (0.166) (0.126) (0.194) (0.166)
personality: intelligent 0.503*** 0.803*** 0.314** 0.508*** 0.811*** 0.317** 0.506*** 0.804*** 0.319**
(0.111) (0.180) (0.140) (0.111) (0.180) (0.140) (0.111) (0.180) (0.140)
personality: thorough 0.115 0.099 0.150 0.116 0.102 0.148 0.115 0.103 0.148
(0.108) (0.178) (0.137) (0.108) (0.178) (0.137) (0.108) (0.178) (0.137)
constant 69.319*** 80.741*** 63.538*** 69.315*** 80.712*** 63.543*** 69.286*** 80.705*** 63.513***
(9.467) (16.442) (11.666) (9.466) (16.443) (11.665) (9.465) (16.441) (11.663)
dependent variable mean 7.866 7.925 7.831 7.866 7.925 7.831 7.866 7.925 7.831
observations 11880 4539 7341 11880 4539 7341 11880 4539 7341
adjusted R-squared 0.001 0.359 0.366 0.001 0.359 0.365 0.001 0.359 0.365
creativity nonroutineness social intelligence Y=time left until planned retirement
Note: Year- and state- fixed effects are included in the regression equations but not reported in the table. Standard errors are in parentheses. Controlled are baseline covariaes, personality traits, and job
satisfaction/ age discrimination. *p<0.1 **p<0.05 ***p<0.01
Source: Authors' calculations using the HRS.
37

Table 1.7A. Automatability and Actual Retirement: A Binary-Outcome Regression (detailed)
(1) (2) (3) (4) (5) (6)
all men women all men women
skill index -0.009*** -0.010*** -0.007** -0.008*** -0.009** -0.006**
(0.003) (0.004) (0.004) (0.002) (0.004) (0.003)
age -0.210*** -0.274*** -0.174*** -0.098*** -0.169*** -0.054
(0.032) (0.055) (0.040) (0.032) (0.051) (0.042)
age-squared 0.002*** 0.003*** 0.002*** 0.001*** 0.002*** 0.001
(0.000) (0.000) (0.000) (0.000) (0.000) (0.000)
race (ref=1.Caucasian)
2. African-American 0.008 0.017 0.002 0.005 0.014 -0.001
(0.010) (0.017) (0.013) (0.008) (0.014) (0.010)
3. Other 0.013 0.031 -0.006 0.013 0.031* -0.008
(0.013) (0.020) (0.016) (0.012) (0.016) (0.016)
education (ref=1.highschool)
2. college degrees -0.022** -0.020 -0.020* -0.020*** -0.022* -0.016
(0.009) (0.013) (0.011) (0.008) (0.012) (0.010)
3. graduate degrees -0.031*** -0.037** -0.026* -0.034*** -0.045*** -0.030**
(0.011) (0.017) (0.014) (0.009) (0.016) (0.013)
male 0.013* - - - - -
(0.007) - - - - -
marital status (ref=1.never married)
2. married/partnered 0.018 0.003 0.025 0.013 0.002 0.018
(0.017) (0.026) (0.022) (0.015) (0.024) (0.020)
3. separated/divorced/widowed 0.006 0.025 0.000 0.002 0.018 -0.002
(0.017) (0.028) (0.021) (0.014) (0.025) (0.019)
spouse ret expectation (ref=1. retire at the same time)
2. spouse to retire earlier -0.027 -0.030 -0.026 -0.049** -0.054* -0.043*
(0.020) (0.033) (0.026) (0.022) (0.032) (0.023)
3. spouse to retire later -0.008 -0.015 -0.004 -0.029 -0.032 -0.024
(0.020) (0.032) (0.026) (0.021) (0.031) (0.023)
household size -0.006** -0.005 -0.006* -0.005* -0.003 -0.006*
(0.003) (0.004) (0.003) (0.003) (0.004) (0.003)
ln(income) 0.000 0.001 -0.000 -0.000 0.000 -0.000
(0.001) (0.001) (0.001) (0.001) (0.001) (0.001)
ln(other HH income) 0.000 -0.000 0.000 -0.000 -0.000 0.000
(0.001) (0.001) (0.001) (0.001) (0.001) (0.001)
ln(wealth) 0.000 -0.001 0.001 0.001 -0.000 0.001
(0.001) (0.002) (0.001) (0.001) (0.002) (0.001)
employer pension type (ref=4. access to none)
1. has DB 0.018* 0.028* 0.012 0.017** 0.022* 0.012
(0.010) (0.015) (0.012) (0.009) (0.013) (0.011)
2. has DC -0.014 -0.008 -0.017 -0.020*** -0.018 -0.023**
(0.009) (0.013) (0.012) (0.008) (0.012) (0.010)
3. both DC and DB -0.003 -0.037 0.023 0.000 -0.033 0.023
(0.020) (0.029) (0.028) (0.018) (0.030) (0.023)
health -0.024*** -0.030*** -0.020*** -0.017*** -0.021*** -0.015***
(0.005) (0.008) (0.006) (0.003) (0.006) (0.004)
health insurance (ref=1. no ESI, no RHI)
2. ESI, but no RHI -0.011 -0.015 -0.007 -0.010 -0.009 -0.011
(0.008) (0.013) (0.010) (0.007) (0.013) (0.010)
3. has both ESI and RHI 0.017** 0.012 0.022** 0.020*** 0.022* 0.018*
(0.009) (0.013) (0.011) (0.008) (0.012) (0.010)
access to Social Security income 0.121*** 0.130*** 0.114*** 0.106*** 0.109*** 0.099***
(0.010) (0.016) (0.013) (0.009) (0.013) (0.010)
eligibility to early Social Security claims (age 62+) -0.028 -0.043 -0.020 -0.031*** -0.047** -0.022
(0.017) (0.028) (0.023) (0.012) (0.021) (0.017)
enjoy going to work -0.022*** -0.023*** -0.022*** -0.022*** -0.025*** -0.020***
(0.005) (0.008) (0.006) (0.004) (0.007) (0.006)
age discrimination: promotion -0.001 0.004 -0.005 -0.003 0.001 -0.005
(0.005) (0.008) (0.006) (0.005) (0.007) (0.006)
age discrimination: retirement pressure 0.013** 0.018** 0.009 0.015*** 0.015* 0.015**
(0.005) (0.009) (0.007) (0.005) (0.008) (0.006)
personality: organized 0.010* 0.021** 0.002 0.009** 0.020*** 0.002
(0.005) (0.008) (0.007) (0.004) (0.007) (0.006)
personality: responsible 0.001 -0.005 0.008 0.001 -0.006 0.007
(0.010) (0.014) (0.016) (0.008) (0.012) (0.012)
personality: hardworking -0.055*** -0.039*** -0.070*** -0.040*** -0.031*** -0.048***
(0.009) (0.013) (0.014) (0.007) (0.009) (0.009)
personality: intelligent 0.008 -0.008 0.020** 0.007 -0.008 0.019**
(0.007) (0.010) (0.009) (0.006) (0.009) (0.008)
personality: thorough 0.004 0.002 0.004 0.003 0.004 0.003
(0.006) (0.010) (0.008) (0.006) (0.009) (0.007)
constant 5.855*** 7.654*** 4.846*** - - -
(0.897) (1.516) (1.119) - - -
dependent variable mean 0.140 0.149 0.135 0.138 0.145 0.131
observations 14577 6047 8530 14577 6047 8530
Logit, marginal effects
Note: Year- and state- fixed effects are included in the regression equations but not reported in the table. Standard errors are in parentheses. Controlled are baseline covariaes,
personality traits, and job satisfaction/ age discrimination. The outcome variable for all columns is a binary actual retirement that takes on a value 1 when working in time t and fully
retiring in t+1, and 0 if working both in time t and t+1. *p<0.1 **p<0.05 ***p<0.01
Source: Authors' calculations using the HRS.
Linear Probability Models
38

Chapter 2

Automation, Aging, and Unretirement Decisions

2.1 Introduction

Human workers and machines increasingly compete against each other for jobs, with
machines filling—automating—jobs that used to be carried out by human labor. In the literature,
two most prominent theories on the relationship between technological changes and human
workers define the nature of automation to be skill-biased and task-based, respectively.
The model of skill-biased technological change (SBTC) explains that the process of
automation replaces individuals based on their primary skills. If individuals possess skills at
which human workers have comparative advantage over the machines, then the technology will
complement the individuals. However, if individuals possess skills that can be more efficiently
carried out by the machines, then they will be substituted by the technology. In making the
argument, the model assumes that those who are complemented by the technology (i.e., high-
skilled) are those with excellent hard skills (e.g., skills obtained through years of education such
as math, reading and writing comprehension). According to the SBTC, then, it is that the high-
skilled—in terms of the level of hard skills—workers are preserved while the low-skilled are
displaced (Card and DiNardo, 2002; Goldin and Katz, 1998; MacCrory et al., 2014). Yet, the
empirical findings of the last few decades evidence that technology-induced automation
39

complements both the low- and high- skilled individuals while substituting the mid-skilled, a
phenomenon known as ‘job polarization’ (Acemoglu, 1999; Katz and Murphy, 1992).
Further generalizing the STBC model, the theory of task-based technological changes
(TBTC) suggests that automation replaces jobs based on their task contents, instead of displacing
workers. However, it is not the entire occupation but specific tasks within each occupation the
technology automates. Essentially, automation reshapes the nature of work (i.e., duties carried
out in each job) (Acemoglu and Autor, 2011; Acemoglu and Restrepo, 2018; Autor et al., 2003).
Whether or not individuals are threatened by the process of automation, then, depends not only
on their skill-sets but also on which occupational tasks they perform at work and on how much
their skill-sets match the occupational tasks.
In this study, we seek to understand the nature of automation through the lens of the
TBTC framework. Empirically, we create an automation resistance of individual skills (ARIS)
index that measures, as the name suggests, individuals’ abilities to resist automation based on
their primary skills. We, then, measure task-specific resistance to automation of occupations in
the US economy, generating an analogous index for jobs called the automation resistance of
occupational tasks (AROT) index. In creating the indices, we do not measure resistance to
automation solely by hard skills, as is done in the literature on the SBTC model. Rather, we
follow Acemoglu and Autor (2011) and others in estimating automation-resistance by the
presence of both hard- and soft- skills/tasks at which human workers have comparative
advantage over machines: creativity, social intelligence, abilities to analyze nonroutine
information, and manual dexterity. (Autor et al., 2003; Deming, 2017; Frey and Osborne, 2017;
McKinsey, 2017). Here, soft skills refer to (1) attributes (e.g., empathy) that characterize
40

individuals’ relationships with others, or (2) skills (e.g., creativity) that are difficult to measure
by traditional educational standards.
Using the two indices and thereby acknowledging both the skill-specific and task-specific
effects of technological changes, we explore the impacts of automation on unretirement (i.e., an
act of returning from full-retirement to the labor force) (Quinn 1999; Cahill et al., 2011, 2015).
For our empirical exploration, we draw data from the Health and Retirement Study (HRS), the
Dictionary of Occupational Titles (DOT), and the Occupational Information Network (O*NET)
from 1992 to 2014. We take advantage of the nature of the panel data that allows us to track the
transitions of individuals (1) in, out, and back into the labor force over time, and (2) across
tenured and temporary jobs.
First, we examine whether individuals’ unretirement decisions depend on the level of
ARIS. Controlling for a rich set of individual observables, we use logit models and find that a
one standard deviation increase in individuals’ ARIS leads to a 1.5% increase in unretirement
from the mean unretirement likelihood of 40%. Next, for unretirees, we estimate the relationship
between individuals’ ARIS and the AROT of the jobs they choose. We find that those with high
ARIS choose jobs that have higher task-specific automation resistance, or higher AROT. Finally,
examined are mental health and financial benefits of skill-to-task matching in unretirees’ job
choices. Our results show that a one standard deviation in this skill-to-task matching reduces
depressive symptoms using two distinct dependent variables: a CES-D score and doctor-
diagnosed depression.
14
Specifically, a one standard deviation in this skill-to-task matching

14
The Center for Epidemiological Studies-Depression (CES-D) score is a common metric for identifying whether
individuals experience depressive symptoms (Radloff, 1977). The measure is described in further detail in Section
2.3.3. Analysis Sample.
41

reduces depressive symptoms by 0.007 CES-D points and by 0.6-1.8% in the self-rated scale
across all cohorts.
The increase in the matching increases wages by 82.6-158.0% if and only if we take into
account cohort-specific valuations of automation-resistance. The mental health benefits of
individuals who achieve high skill-to-task matching persist even when I control for changes in
wages, suggesting that the mental health benefits are not resulting from the rising financial
benefits. In addressing the last two questions, we restrict our sample to only those who have
chosen unretirement, thereby exposing it to potential non-random sampling biases. In an effort to
remove the biases, we perform Heckman two-stage corrections to account for incidental
truncation (Vella, 1998; Semykina and Wooldridge, 2010).
In this study, we examine unretirement as a distinct phenomenon instead of treating it as
a part of the rest of the life cycle of work. As opposed to aging workers who are still in the labor
force and considering whether to switch to other jobs (e.g., part-time employment or self-
employment), unretirees make their decisions to come back to the labor force in the absence of
professional ties—thereby being subject to greater levels of uncertainties in securing jobs again.
Also, the fact that unretirees are likely to have experienced a sense of social isolation in
retirement, they may come back to the labor force for reasons (e.g., restore social ties or to enjoy
mental health benefits) different from current workers who are switching jobs (Adams et al.,
2004; Segel-Karpas et al., 2018).
This study serves as one of the very first attempts in the labor supply literature to discuss
potential determinants of unretirement. Such an effort is necessary in today’s society whereby
many nations are experiencing population aging and growing unretirement in the labor market.
In the United States, for instance, unretirement rates increase from 6 to 14 percent in the 1980s
42

(Berkovec and Stern, 1991; Hayward et al., 1994) to 15 to 26 percent in the 1990s (Cahill et al.,
2010; Maestas, 2010). Despite the growing number of unretirees in the nation’s labor force, little
is known about the rationale behind their decisions to return to work (Allen et al., 2014; Coile
and Levine, 2006; Maestas 2010). The job choices of unretirees deserve special attention because
the motivations for their decisions differ from those of younger workers who have not entered
initial retirement.
15

Another contribution this study makes lies in its attempt to explore as potential
determinants of unretirement the process of automation and the extent of automation-resistance
of individuals and jobs—factors that are understudied and yet highly relevant in the modern
society. In doing so, we not only conjoin the theoretical models of technological changes and of
labor supply, but we also acknowledge both the skill-specific and task-specific natures of the
process of automation via creating two distinct metrics (i.e., ARIS and AROT indices). Very few
existing studies delve into examining the linkage between technological changes and the labor
market. For instance, Deming (2017) studies the role of social skills—one of the main skills that
ensure individuals’ automation-resistance—on individuals’ job choices and wages. In this
chapter, we shed new light on the impact of technological changes by considering a fuller set of
skills at which human workers have comparative advantage over the machines: creativity,
abilities to analyze nonroutine information, and manual dexterity in addition to social
intelligence. Moreover, by studying the late-life job choices of retirees as they transition back to
the labor force, we deviate from the previous research that focuses on the occupational choices of

15
For instance, because many retirees are already claiming income from private pension plans or Social Security
benefits—neither of which is available for entry-level or mid-career workers—the financial aspects of retirees’
back-to-work decisions should be gauged not only by the extent of their household wealth but also by their pension
and Social Security wealth (Coile and Gruber, 2000; Feldstein, 2005; Shoven and Slavov, 2014).
43

job-market entrants or middle-aged individuals changing jobs mid-career (Cable and Judge,
1996; Deming, 2017; Eriksson and Kristensen, 2014).
While our work is written in the context of the US, its insights have implications for
many other nations faced with rapid population aging and increasing unretirement. The outcomes
of this study could assist policy-makers in many developed nations who seek to promote longer
working lives of the aging, particularly as the nations’ shrinking tax bases are increasingly
falling short of financially sustaining current levels of public pension and health benefits for
retirees. Finally, the findings of this chapter can yield insights for designing more effective
training programs for aging workers and retirees seeking employment that focus not only on
building the hard skills but also the soft skills that will help them survive in the workplaces that
are highly automated.

2.2 Conceptual Framework
2.2.1 Theories on the Nature of Technological Growths
The conceptual framework for this chapter is based on theories of the SBTC and the
TBTC literature. We start this section by discussing the intuitions behind the SBTC framework.
A more technical description of the SBTC framework is available in Appendix A.
The fundamental ideas behind the SBTC model are proposed by Welch (1973), Katz and
Murphy (1992), Card and Lemieux (2001), Card and DiNardo (2003), and Myck and Reed
(2006), among many others. In a simplified SBTC model, there are only two groups of workers
(i.e., skilled and unskilled) in an economy. The aggregate labor demand is generated by a
production function consisting of labor inputs by the skilled and the unskilled workers, with a
constant elasticity of substitution between the two. Technological parameters in the function are
44

set to vary across time. The relative demand for the skilled workers is determined by setting the
ratio of the marginal product of the skilled and the unskilled equal to the ratio of their wages.
Here, the model assumes that the changes in the relative wage must stem from either the (1)
changes in the relative supply of skilled workers or (2) changes in technology.
16
An increase in
the relative proportion of skilled workers would correspond to a decrease in their relative wage
holding constant the technology. If the relative supply of skilled-workers were to be held
constant, then the technological changes would generate a wage differential—with the skilled
workers’ relative wage being higher than that of the unskilled counterpart. Hence, the
technological changes are skill-biased.
17

Three shortcomings of the SBTC model merits attention. First, in the SBTC, the
technological growth is assumed to increase wages for all workers, albeit at a higher rate for the
high-skilled workers. This is widely at odds with the findings of the literature on the returns to
educational premium, which demonstrates that the real wages rose for the skilled workers while
falling steeply for the unskilled workers in the 1980s and in the 2000s (Katz and Murphy, 1992;
Acemoglu and Autor, 2011).
18
Second, as the SBTC model predicts a monotonic relationship
between technology-induced automation and skill levels (i.e., the more skilled an individual is,
the more complemented by technological changes the individual is), it fails to explain the job
polarization. The phenomenon of polarization started in the 1990s where the employment share
grew rapidly both at for the low-skilled and the high-skilled workers while falling for the mid-

16
Other features that influence relative wages such as premiums for efficiency or productivity have been ignored in
the model.
17
On the contrary, skill-neutral technological changes would generate a same level of wage increase across all skill-
levels, resulting in a proportional shift in the technological parameters for two skill groups and thus leaving the
relative productivity of the two unchanged.
18
We equate the level of skills with the years of educational attainment. The skilled refer to the highly-educated
individuals with post-college degrees while the unskilled refers to the less-educated, without four-year college
degrees.
45

skilled ones (Acemoglu, 1999).
19
In the 2000s, the polarization persisted, as the employment
share grew steeply only among the low-skilled individuals while stagnating for the high-skilled
ones—with the share of those mid-skilled falling (Acemoglu and Autor, 2011). Third, the SBTC
framework equates workers’ skills and occupational tasks. In doing so, the framework presumes
that workers achieve a perfect match (i.e., a one-to-one match) between their skills and the
occupational tasks they carry out at work, thereby ignoring imperfect skill-to-task matches that
exist in the economy (Acemoglu and Autor, 2011; Acemoglu and Zilibotti, 2001).
The limitations of the SBTC model have led to development of a more comprehensive
model of Task-Based Technological Changes (TBTC). It accounts for the fact that wages can
decrease for those whose skills are substituted by the technological changes. Also, the model
better reflects the ‘job polarization’, or the non-monotonic relationship between technology-
induced automation and skill levels. Lastly, the TBTC framework brings into the picture the
notion of occupational tasks. The theoretical model has been developed predominantly by
scholars including but not limited to Autor et al (2003), Acemoglu and Zilibotti (2001), Autor
and Dorn (2010), and Acemoglu and Autor (2011). The remainder of this section summarizes the
model of TBTC, following the notations of Acemoglu and Autor (2011). A more technical
explanation of the TBTC framework is presented in Appendix B.
Indeed, the model differentiates workers’ skills (i.e., stock of capabilities) and tasks (i.e.,
units of activity performed at each job that produces output). An occupation consists of a
continuum of tasks—all of which are needed to produce a final good off the job.
20
Allowing no
trade in tasks and assuming a closed economy, the production of a unique final good is achieved

19
Ibid.
20
The TBTC model accounts for the fact that recent surge in offshoring and outsourcing of labor influences the
labor market. In the end, framework sets it so that the occupational tasks can be outsourced (i.e., produced outside
the US), automated, or upheld by workers (Autor et al. 2003).
46

via combining a range of occupational tasks among workers of varying skill levels. To better
encapsulate the job polarization empirically observed in the past decades, the model divides
individuals with three skill groups: low-skilled, mid-skilled, and high-skilled. The supply of each
skill level is inelastic.
The TBTC model accounts for the fact that there can be imperfect matching between
workers’ skills and tasks they carry out at their jobs (i.e., imperfect skill-to-task matching). Thus,
the framework acknowledges the possibility of a mid-skilled and a high-skilled individuals to
hold the same job—even though the two workers’ extent of skill-to-task matching differs.
Moreover, workers in the three skill groups could be allotted the same tasks. For the remainder
of this section, we consider occupational tasks, rather than jobs, as our unit of analysis.
Regardless of the possibility of holding the same job or carrying out the same tasks,
workers have different comparative advantages over different tasks depending on the specific
skills they possess. The varying comparative advantages of workers often require them to
collaborate to produce the final good. The production function of each task is written as,

(5) 𝑌 = 𝐴 𝐿 𝛼 𝐿 (𝑖 )𝑙 (𝑖 ) + 𝐴 𝑀 𝛼 𝑀 (𝑖 )𝑚 (𝑖 ) + 𝐴 𝐻 𝛼 𝐻 (𝑖 )ℎ(𝑖 ) + 𝐴 𝐾 𝛼 𝐾 (𝑖 )𝑘 (𝑖 ) .

𝐴 𝐻 is the technology favoring the high-skilled workers, 𝐴 𝑀 favors the mid-skilled, and so on.
The 𝛼 𝐿 (𝑖 ), 𝛼 𝑀 (𝑖 ), and 𝛼 𝐻 (𝑖 ) are the task-specific productivity parameters of each skill group in
carrying out different tasks 𝑖 .
21
Lastly, 𝑙 (𝑖 ), 𝑚 (𝑖 ), and 𝑞 (𝑖 ) are the number of low-skilled, mid-
skilled, and high-skilled workers carrying out task 𝑖 . Due to the different comparative advantages

21
The comparative advantages have the following structure: (1) As for 𝛼 𝐿 (𝑖 )/𝛼 𝑀 (𝑖 ), and 𝛼 𝑀 (𝑖 )/𝛼 𝐻 (𝑖 ). Higher
indices suggest more complicated tasks whereby the skill group whose productivity schedules are on the numerator
perform better than those in the denominator.
47

of the three skill groups, some tasks (𝑇 𝐿 ) will be best performed by the low-skilled; 𝑇 𝑀 , by the
mid-skilled; and 𝑇 𝐻 , by the high-skilled.
22
Naturally, tasks will be divided among workers of
three distinct skill levels. The extent of substitutability of skills across the occupational tasks
depend on (1) the quantity of labor supplied for each skill level and (2) the nature of
technological changes. The allocation of workers to occupational tasks—and hence the extent of
skill-to-task matching—is decided jointly by technological growths as well as hiring decisions of
the demand side.
23,24
The biased nature of technological growths dictates that workers with
varying skill levels have different levels of productivity in carrying out certain tasks.
Workers use skills to accomplish occupational tasks and to receive salaries in return. Any
changes in workers’ earnings as well as the differences in earnings across the skill groups are
determined by the quality of the skill-to-task matches and identification of threshold tasks (𝐼 𝐿 and
𝐼 𝐻 ).
25
Intuitively, the threshold tasks 𝐼 𝐿 are a set of tasks that can either be carried out by low-
skilled or mid-skilled workers; 𝐼 𝐻 can be understood in a similar manner. In that wages are
simply the value of the marginal product of different skill levels, the wage of a low-skilled
worker (𝑤 𝐿 ) is equivalent to the product of price index of tasks performed by low-skilled
workers and technology favoring the low-skilled. The wages of the mid- and the high-skilled are
defined in the same way.

22
Employers hiring workers from the same skill group (e.g., the mid-skilled group) must pay them the same wage in
an equilibrium (i.e., the value of the marginal product of individuals in the same skill group will be the same in all
tasks they carry out).
23
An equilibrium is reached when the demands-side maximize profits and the labor markets clear.
24
The model of TBTC does not include labor supply behaviors as potential determinants of allocations of skills to
tasks. In other words, the TBTC model lets the optimal allocation of skill-to-task to be done by the employers alone.
25
There exist a range of tasks. The low-skilled workers perform a portion of them, the mid-skilled perform another
portion, and the high-skilled perform the rest. Here, given the existence of substitutability of skills across the
occupational tasks, the threshold tasks refer to the ones that are shared by the workers across skill groups.
48

The equilibrium in the TBTC equalizes the cost of carrying out the threshold tasks 𝐼 𝐿 by
low-skilled and mid-skilled individuals as well as the cost of carrying out 𝐼 𝐻 by mid-skilled and
high-skilled individuals. Such conditions known as the ‘no arbitrage’ conditions. From the two
‘no arbitrage’ conditions, Acemoglu and Autor (2011) demonstrate that there exists a unique
intersection between the two, signifying the optimal allocation point of skills to occupational
tasks. Based on these equilibrium conditions of the TBTC model, we can anticipate how the
relative wages of workers in the three skill groups change in response to the particular nature of
the technological changes unraveling in the society.
High-Skill Biased Technological Changes: Technological changes, if favoring the high-
skilled (an increase in 𝐴 𝐻 ), increases the productivity of the high-skilled workers—thereby
expanding their set of tasks at the expense of those of the mid- and the low-skilled (i.e., lowering
the threshold tasks 𝐼 𝐿 and 𝐼 𝐻 ). Following the re-allocation of tasks, a rise in 𝐴 𝐻 raises the wage of
the high-skilled relative to that of both the mid-skilled and the low-skilled. In other words, both
𝑤 𝐻 𝑤 𝑀 and
𝑤 𝐻 𝑤 𝐿 increases. The direct effect of an increase in 𝐴 𝐻 is a shrinkage of tasks performed by
the mid-skilled. While the indirect effect of the rise in 𝐴 𝐻 is a reduction of tasks performed by
the low-skilled, the indirect effect never dominates the direct effect because the mid-skilled
workers become cheaper—incentivizing employers to expand tasks performed by the mid-
skilled. In the end, the wage of the mid-skilled decrease relative to that of the low-skilled, or
𝑤 𝑀 𝑤 𝐿
decreases.
Mid-Skill Biased Technological Changes: If technological changes are middle-skill
biased (an increase in 𝐴 𝑀 ), they increase the productivity of the mid-skilled. The technology,
then, reduces 𝐼 𝐿 and increases 𝐼 𝐻 , expanding the set of tasks performed by the mid-skilled by
shrinking those of the low-skilled and the high-skilled. Due to the task re-allocation, the increase
49

in 𝐴 𝑀 raises the wage of the mid-skilled relative to that of both the high- and the low-skilled—
decreasing
𝑤 𝐻 𝑤 𝑀 and increasing
𝑤 𝑀 𝑤 𝐿 . Unlike the case of high-skilled biased technological changes,
the rise in 𝐴 𝑀 has a direct effect on both the high-skilled and the low-skilled—exerting
downward pressures on their wages. Thus, the impact of the rise of 𝐴 𝑀 on the relative wages of
the high-skilled to the low-skilled (
𝑤 𝐻 𝑤 𝐿 ) are ambiguous.
Low-Skill Biased Technological Changes: If technological changes favor the low-
skilled (an increase in 𝐴 𝐿 ), they raise both 𝐼 𝐿 and 𝐼 𝐻 , expanding the set of tasks performed by the
low-skilled while contracting those of the mid- and the high-skilled individuals. The re-
allocation of tasks due to the increase in 𝐴 𝐻 raises the wage of the low-skilled relative to that of
both the high- and mid-skilled. Both
𝑤 𝑀 𝑤 𝐿 and
𝑤 𝐻 𝑤 𝐿 decreases. The direct effect of an increase in 𝐴 𝐿
is a shrinkage of tasks performed by the mid-skilled. While the indirect effect here is a
contraction of the set of tasks performed by the high-skilled, the indirect effect never
overshadows the direct effect. The mid-skilled workers become significantly cheaper. In the end,
the wage of the mid-skilled decrease relative to that of the high-skilled—resulting in an increase
in
𝑤 𝐻 𝑤 𝑀 .
In this chapter, based on the empirical evidence of the past few decades, technological
changes are set to favor both the low-skilled and the high-skilled—while automating the mid-
skilled. When both the 𝐴 𝐿 and 𝐴 𝐻 rise, their net effect on the skill allocation (i.e., changes in the
threshold tasks 𝐼 𝐿 and 𝐼 𝐻 ) is ambiguous. While the rise of 𝐴 𝐿 and 𝐴 𝐻 consistently increase
𝑤 𝐻 𝑤 𝑀
and decrease
𝑤 𝑀 𝑤 𝐿 , their effects on
𝑤 𝐻 𝑤 𝐿 clash. The net change in
𝑤 𝐻 𝑤 𝐿 is therefore unclear.
Compared to the relative wages, unmentioned above are the effect of technological
changes on the actual wage levels of each skill group. Unlike the SBTC model that assume that
50

the wages of all skill-groups rise, the model of TBTC sufficiently shows that the wage level can
be reduced for certain group of workers whose comparative advantages disappear and their set of
tasks contract. For instance, the high-skill biased technological changes can significantly reduce
the wage of the mid-skilled (Acemoglu and Autor, 2011). Nevertheless, changes in the wage
levels of low-, mid-, and high-skilled workers due to the three types of technological changes are
much harder to predict than the relative wages. In the estimation section of the chapter, we use
data at hand to evaluate the net effect of the technological changes—favoring both high- and
low-skill—on individuals’ wage levels. Of note, in studying the linkage between technology and
wages, we deem that the wage level to be a good proxy for how much employers value the future
employees’ skills and resistance to automation.
26

In showing that the skill-to-task allocations are determined by the labor demands, the
TBTC model is silent on the role of the labor supply. We believe that the labor supply could
contribute to the re-allocation of skills to tasks in the face of the technological changes—
generating behavioral changes (e.g., workers of a particular skill group flocking to certain level
of occupational tasks) even in the absence of re-allocation incentives provided by the employers.
In Section 2.4-2.5 of the chapter, we empirically test the extent of the skill-to-task allocations
driven by the supply side while controlling for the financial allocation incentives (e.g., wages)
offered by the demand side. In doing so, we posit that the technological changes generate
different psychological returns to workers with varying skill levels, thereby generating a job-
satisfaction inequality among the high-, mid-, and low-skilled workers similar to the technology-

26
One may question that the real wage alone does not represent the total compensation of the labor, contending that
the wage inequality observed between the skilled and the unskilled workers could be reduced or eliminated if other
fringe benefits (e.g., healthcare, sick time, bonuses) are considered. However, Pierce (2001) who assesses the
evolution of wages as well as other fringe benefits of the skilled and the unskilled find that the trends using the total
compensation package does not produce a significant deviation from the trends observed using just the real wages
(Katz and Murphy, 1992; Acemoglu and Autor, 2011).
51

induced wage inequalities. Accounting for the fact that the psychological returns could follow
the changes in wage levels induced by the technological changes, we examine whether there are
psychological returns while controlling for the financial rewards.
27

2.2.2 Linking Technological Growths and Unretirement

In a standard life cycle model of labor supply, once the technology-induced process of
automation determines the relative changes in the wage levels of the skill groups, a forward-
looking individual seeks to choose the optimal work and retirement paths that can maximize her
lifetime utility determined by consumption and leisure, or U(C,L), subject to budget
constraints.
28
When the costs of consumption and leisure as well as the purchasing power change
with alterations in the individual’s wage trajectory and assets, the individual re-evaluates how
much time to exert to work and to leisure (Blundell and MaCurdy, 1999).
Given the way in which the life cycle model is set up, unretirement can enter the model in
two ways. First, unretirement may be unforeseen prior to an individual’s initial retirement. In this
case, unretirement enters the life cycle model through uncertainty, where an individual receives
new information after the initial retirement that alters her budget constraint (e.g., unexpected
health problems, or changes in her finances) and re-optimizes her labor supply paths. Second,
unretirement could instead be planned ahead of an individual’ initial retirement can be viewed as
a part of her original work- and retirement- paths chosen to maximize her lifetime utility—based
on the information she has on the determinants of her intertemporal asset evolution constraint for

27
As for the psychological returns that succeeds monetary returns, for instance, it could be that a skilled worker’s
level of psychological health or job satisfaction heightens as she observes her wage growing at a higher rate than
that of the unskilled counterpart.
28
The full-time devoted to leisure is equivalent to retirement.
52

the remainder of her life (Blundell and MaCurdy, 1999; Myck and Reeds, 1999).
29
Examples of
the determinants are Social Security and private pension claiming rules, health insurance
provision terms of their current employers, etc. (Coile and Gruber, 2007; French and Jones,
2011; Garber and Skinner, 2008; Gruber and Wise, 1999; Lusardi and Mitchell, 2007).
In this chapter, in addition to the widely-known determinants listed above, we explore the
potential role of the task-based technological changes—proxied by individuals’ varying levels to
resist automation—as determinants of the wage- trajectories and subsequent unretirement
likelihoods of aging workers. Changes in individuals’ wage levels induced by the TBTC will
generate income- and substitution- effects. Let us briefly hypothesize about the possible
unretirement likelihoods of a high-skilled retiree (i.e., displaying a higher resistance to
automation than her counterparts), faced with a high-skill biased technological change (rise in
𝐴 𝐻 ):
Hypothesis 1 (dominant income effect): In experiencing higher wage growths
compared to those of the mid- and low-skilled workers, high-skilled individuals increase
their demand of leisure (i.e. time spent in retirement). If this income effect predominates
over the substitution effect, and if other things are set to be equal, the skilled individuals
will be less inclined to unretire—thereby spending more time in retirement than their
counterparts.

Hypothesis 2 (dominant substitution effect): In experiencing higher wage growth, a
high-skilled individual finds that her leisure has become more expensive (i.e. a higher

29
Indeed, previous research provides empirical evidence that unretirement is often predicted prior to the initial
retirement (Maestas, 2010).
53

opportunity cost of not working) and tends to want less of it. If this substitution effect
dominates the income effect, then the high-skilled individual will be more likely to
unretire after their initial-retirement.

The unretirement inclinations portrayed by the hypotheses 1 and 2 bear point to opposing
directions. Depending on which of the two effects prevails over the other, unretirement decisions
of the high-skilled worker will be decided.
Ascertaining the predictability of unretirement prior to initial retirement is beyond the
scope of this chapter. In this work, we simply acknowledge the roles played by the information
sets obtained prior to and after the initial-retirement by including both in our estimation models
as determinants of unretirement.
Next, when retirees decide to unretire, what types of jobs would they choose? We explore
the question using the utility maximization framework. Assume that a worker i in sorting into a
job j seeks to maximize the following utility function 𝑈 𝑖𝑗
= 𝑈 (𝑊 𝑖𝑗
, 𝑅𝐴
𝑖𝑗
). Here, 𝑅𝐴 is the extent
of resistance to automation embedded in the tasks performed at the job j and W is the wage.
30
We
assume that there is no saving, implying that wage is used entirely on consumption. Utility from
the 𝑅𝐴 —distinct from that of the 𝑊 —comes from psychological rewards to having a job whose
task contents have a low resistance to automation such as an increase in job satisfaction, a
growing sense of self-worth, and reduced stress and anxiety levels that exist in the absence of
any changes in the wage level.

30
Here, we are trying to make an argument that W and RA—the elements of our utility function--are independent.
Here, RA is the extent of automation-resistance of task contents of jobs, not of workers’ skills. Workers’ having
skills that have high resistance to automation (i.e., high ARIS) will increase productivity, which will influence W.
But an occupation’s consisting of tasks with high resistance (high RAT) in itself does not touch upon a job holder’s
productivity. The productivity has to do with her skills.
54

While the utility rises with W for all workers, we posit that the utility associated with the
𝑅𝐴 rises for some while not for others. This is because the psychological effect of performing
occupational tasks with high resistance to automation is positive only when workers’ main skill-
sets render them capable of performing such tasks. For instance, one of the social skills that
enhance the extent of resistance to automation is social intelligence. For an individual who has
high social skills, it may increase her job satisfaction if she were to hold a job whose main
duties—dictated by the task contents—require dealing with people. For someone with low social
skills, holding such a job could generate the opposite effect.
31

For simplicity’s sake, assume that 𝑅𝐴 only takes values 0 and 1: The economy consists of
jobs whose task contents consist entirely of tasks that have high resistance to automation (i.e.,
𝑅𝐴
𝑖𝑗
=1) and jobs devoid of such tasks (i.e., 𝑅𝐴
𝑖𝑗
=0). Applying Rosen’s (1986) model (Currie
and Madrian, 1999; Gruber, 2000), we set ∆𝑊 𝑖𝑗
as the compensating differentials. In other
words, ∆𝑊 𝑖𝑗
is the tradeoff between wage and requiring high-resistance tasks in a job that is
necessary for a worker to be indifferent between choosing jobs with 𝑅𝐴
𝑖𝑗
=1 and 𝑅𝐴
𝑖𝑗
=0.
32
The
tradeoff can be both positive and negative for a worker depending on whether her utility
increases or decreases, respectively, with 𝑅𝐴
𝑖𝑗
. For an individual who possesses more of the
high-resistance skills, obtaining a job with 𝑅𝐴 =1 has a positive payout both in terms of W and
𝑅𝐴 . For this individual to be indifferent between the job with 𝑅𝐴 =1 and another job with 𝑅𝐴 =0,
the employer offering the job with 𝑅𝐴 =0 will have to set the wage level sufficiently high so as to

31
A stream of literature states that the basis of psychological tolls (e.g., intensified stress) at workplaces lies in the
discrepancy between workers’ capabilities and tasks demanded (Caplan and Harrison, 1993; Sharit and Czaja, 2007)
32
The ∆W ij adjusts for not only the compensating differentials but the degree to which the society rewards
resistance to automation—which seems to have grown over time (Katz and Murphy, 1992; Acemoglu and Autor,
2011). Then, the sorting for an individual with certain preferences and skill-sets in cohort A will be different from a
comparable person in the next cohort B. This will be accounted empirically in this chapter by taking cohort fixed-
effects.
55

reflect the utility-loss in 𝑅𝐴 the individual bears in choosing the job with 𝑅𝐴 =0 over the job with
𝑅𝐴 =1. Here, the 𝑅𝐴 is assumed to be experience-rated at the worker level, indicating that each
worker faces a distinct rate of compensation based on the extent of high-resistance tasks they
perform at work.
33
The wage rate of this individual is equal to her marginal productivity prior to
considering the compensation differentials.
Given the utility formulation, an individual i will desire a job with 𝑅𝐴
𝑖𝑗
=1 if there exists
a ∆W
ij
such that 𝑈 (𝑊 𝑖𝑗
− ∆𝑊 𝑖 𝑗 , 𝑅𝐴
𝑖𝑗
= 1) − 𝑈 (𝑊 𝑖𝑗
, 𝑅𝐴
𝑖𝑗
= 0) = 𝑉 𝑖𝑗
≥ 0. Here, 𝑉 𝑖𝑗
represents
individuals’ heterogeneous valuation of jobs with high-resistance tasks. Intuitively, when
workers’ skills and occupational tasks they carry out match closely in terms of the contents as
well as the extent of automation-resistance, the workers are likely to assign a higher value of 𝑉 𝑖𝑗

than their counterparts. If we define 𝐶 𝑖 as the cost the firms incur in offering jobs with 𝑅𝐴
𝑖𝑗
=1,
the firms will set it so that ∆𝑊 𝑖𝑗
= 𝐶 𝑖 assuming that they live in a perfectly competitive
economy. The fact that jobs vary from one another in terms of the level of 𝑅𝐴
𝑖𝑗
incentivizes
workers to optimize their labor supply paths via sorting themselves into firms that match their
preferences over 𝑊 and 𝑅𝐴 given their primary skill-sets, automation-resistance (determined by
the skill sets), and other work and retirement related considerations.
Then, the supply of individuals that sort into jobs with 𝑅𝐴
𝑖𝑗
=1 and jobs with 𝑅𝐴
𝑖𝑗
=0 can
be written in terms of the probability density function of 𝑉 𝑖𝑗
across all workers in the economy,
g(𝑉 𝑖𝑗
) and its cumulative distribution, G(𝑉 𝑖𝑗
). If we normalize the total labor force to 1, the
supply of 𝑅𝐴
𝑖𝑗
=1 workers is

33
We borrow the notion of ‘experience-rating’ from the context of employer-sponsored health insurance and wage
tradeoffs. The experience-rating is a way for private insurance providers to charge the enrollees based on the
enrollees’ actual—both projected and the past—cost experience (Gruber, 2000).
56

(6) ∫ g(𝑉 𝑖𝑗
)
∆𝑊 𝑖𝑗
0
𝑑𝑉 = 𝐺 (∆𝑊 𝑖𝑗
),

and the supply of their counterparts is 1−G(∆𝑊 𝑖𝑗
).
Based on this model, the relative supply of workers’ sorting into 𝑅𝐴
𝑖𝑗
=1 jobs depends on
the distribution of ∆𝑊 𝑖𝑗
employers are willing to inject to modify the wage rates of the jobs.
Also, the model is set so that individuals’ job sorting depends on their heterogeneous internal
valuation of automation-resistance. Those with higher valuation will sort into jobs whose duties
consist mainly of high-resistance tasks, and vice versa. The unretirement occupational sorting is
thus endogenously determined by their preferences—which can also dictate the types of primary
skill-sets the individuals obtain. In our estimation models, we examine the extent of skill-to-task
matching driven by individuals’ varying levels of automation-resistance while accounting for the
unobserved individual-specific preferences for work, retirement, and skill-choices.

2.3 Study Data
2.3.1 Measuring the Extent of Automation-Resistance
The occupation-specific information (e.g., task-contents) is drawn from the Dictionary of
Occupational Titles (DOT) and Occupational Information Network (O*NET) for years between
1992 and 2014.
34
Using the information from the DOT (4
th
ed.) for years preceding 1998 and

34
Both the revised 4
th
edition of the DOT and O*NET are created and maintained by the U.S. Department of Labor.
The DOT is created by the U.S. Department of Labor Employment and Training Administration in 1991. See
https://www.oalj.dol.gov/LIBDOT.HTM. This is the only version of the DOT drafted in the 1990s; its predecessor—
the 3
rd
edition—was published in 1965. Starting in 1998, the information available in the DOT were transferred to
the O*NET. The occupation-specific information in the O*NET are identified by the experts endorsed by the US
Department of Labor, and is updated annually. Some deviations in the definition of the non-automatability are
unavoidable due to the changes in the occupation evaluation system from the DOT to the O*NET. I do the best to
keep the selection criteria of the skill and task items that comprise the non-automatability indices to be closely
comparable across the DOT and O*NET and across waves.
57

O*NET—the successor of DOT—for years between 1998 and 2014, we calculate the ARIS of
workers based on their primary skill-sets.
35

Based on the empirical evidence on the technology-induced job polarization described in
Section 2.2.1., we set the process of automation to be biased towards both the low- and the high-
skilled—substituting only those who are mid-skilled (Acemoglu, 1999). What kind of skills are
possessed by the mid-skilled workers and not by the high- or the low-skilled individuals? We
find an answer to the question from the routinization hypothesis (Acemoglu and Autor, 2011).
The hypothesis states that the technology holds comparative advantage at handling highly
routinized work (i.e., codified and programmed into a set of rules) such as clerical duties,
repetitive data input, and production work than human workers. Such routinized work is mostly
carried out by mid-skilled workers in, for instance, office administrative jobs—and not by the
low- and the mid-skilled (Acemoglu and Autor, 2011).
Based on these guiding rules, we select from the task pool used by the DOT and the
O*NET both the abstract and the manual skills that are nonroutine to create the ARIS index of
individuals. The abstract nonroutine skills are related to problem-solving and exercise of
intuition, persuasion, creativity, and social intelligence. Such skills are known to be possessed
mostly by high-skilled individuals in professional, managerial, STEM, or creative occupations.
The manual nonroutine skills are those involving situational adaptability, visual and language
recognition, manual dexterity (i.e., accuracy of physical work), and in-person interactions. Such
skills are possessed by those in blue-collar jobs including in-person health assistant, food-

35
Ideally, it would be the best to obtain skill-specific information for each respondent in the HRS. Unfortunately,
variables on the skills of individuals are not available through the HRS. Given the limitation of data, we resort to
inferring what individuals’ primary skills are based on their career jobs (i.e., jobs at which individuals worked for
the longest time), using the job-specific information available in the O*NET. In making use of the O*NET data on
the representative skill-sets possessed by an average job holder of each occupation, we equate the HRS respondents’
primary skill-sets with those of the average job holder of their career jobs.
58

servers, janitors, etc.—often concentrated in service industry. In the end, given the criteria for
calculating the ARIS index, a higher ARIS score is not necessarily equivalent to having more
years of education, having a white-collar as opposed to blue-collar job, or being equipped with
more hard skills.
36

For each occupation, the DOT and O*NET rate the extent to which the skill is used in an
scale of 0-5, with those with 5s being the most frequently exercised.
37
For example, an average
human resources manager possesses skills in reading comprehension (4.25/5), monitoring
(4.38/5), and social perceptiveness (4.12/5), while knowing very little about technology design
(1.12/5), programming (1.62/5), or equipment troubleshooting (0/5). From the total pool of skills
available in the DOT and the O*NET, we select the skills that help individuals increase their
automation-resistance—skills related to handling nonroutine analytical- and interpersonal- duties
for high-skilled individuals as well as nonroutine adaptability-related and interpersonal duties for
manual for the low-skilled individuals.
38
Then, using the skill ratings, we sum the ratings of the
selected skills to calculate the ARIS of the HRS respondents based on the primary skills (i.e.,
skills used in their career jobs). From the sum, subtracted are the averaged ratings of skills that
lower the automation-resistance: abilities to handle routinized work and skills displaying a high
offshorability.
39
TABLE 2.1 contains a complete list of skills used to calculate the ARIS index.
In the end, the ARIS index is calculated as follows,

36
Once again, hard skills are those garnered through years of education such as math, reading and writing
comprehension.
37
There are two types of ratings that DOT and O*NET use: an ‘importance’ rating and a ‘level’ rating. The former
rating measures how valuable each skill is for an average job-holder. The latter measures how much of the skill is
needed and used by the average job-holder. While Acemoglu and Autor (2011) use the importance rating, we use the
levels.
38
While the pool of the DOT’s task elements differs from that of the O*NET, the DOT criteria are selected so that
they closely resemble that of the O*NET.
39
A high offshorability indicates that the job is likely to be removed from the U.S. economy, to be performed abroad
at a location where the costs of production is lower.
59

(7) 𝐴𝑅𝐼𝑆 𝑖 =
1
𝑇 (∑ 𝑋 𝑖𝑘
𝑁 𝑅 𝐴𝑁𝐴𝐿 𝑇 𝑘 + ∑ 𝑋 𝑖𝑘
𝑁 𝑅 𝑆𝑂𝐶𝐼𝐴𝐿 𝑇 𝑘 + ∑ 𝑋 𝑖𝑘
𝑁 𝑅 𝐷𝐸𝑋 𝑇 𝑘 +
∑ 𝑋 𝑖𝑘
𝑁𝑅 _𝐼𝑁𝑇𝐸𝑅𝑃𝐸𝑅𝑆
− ∑ 𝑋 𝑖𝑘
𝑅 _𝐻𝐼𝐺𝐻 −
𝑇 𝑘 ∑ 𝑋 𝑖𝑘
𝑅 _𝐿𝑂𝑊 − ∑ 𝑋 𝑖𝑘
𝑂𝐹𝐹𝑆𝐻𝑂𝑅 ),
𝑇 𝑘
𝑇 𝑘 𝑇 𝑘

where the subscript i indicates each individual in the labor force, and 𝑋 𝑖 𝑁𝑅 _𝐴𝑁𝐴𝐿 refers to
nonroutine analytical skills for the high-skilled individuals; 𝑋 𝑖 𝑁𝑅 _𝑆𝑂𝐶𝐼𝐴𝐿 is a vector of nonroutine
social intelligence possessed by the high-skilled; 𝑋 𝑖 𝑁𝑅 _𝐷𝐸𝑋 is a vector of nonroutine manual
dexterity owned by the low-skilled individuals; 𝑋 𝑖 𝑁𝑅 _𝐼𝑁𝑇𝐸𝑅𝑃𝐸𝑅𝑆 consists of nonroutine
interpersonal skills of the low-skilled workers; 𝑋 𝑖 𝑅 _𝐻𝐼𝐺𝐻 and 𝑋 𝑖 𝑅 _𝐿𝑂𝑊 are skills to handle
routinized work, owned by both the high- and low-skilled individuals; and 𝑋 𝑖 𝑂𝐹𝐹𝑆𝐻𝑂𝑅 are
offshorable skills. For easier interpretations, we standardize the index by subtracting its mean
and dividing by its standard deviation. The final index is a continuous index whereby all
workers’ levels of automation-resistance are rated in a spectrum, with a higher index value
indicates a greater resistance.
In that the skills honed through completing the career job tasks become stocks for the
individuals, the ARIS index lacks within-individual variations across time. Recent literature
measuring the skill levels of workers maintain a similar strategy in preserving the lack of
variations over time in the stock of skills (Deming, 2017; Frey and Osborne, 2017).
Next, we create the AROT index that ranks the automation resistance of jobs based on
their tasks: The higher the AROT index of an occupation is, the more resistant the job is to the
pressures of automation. The AROT index is calculated in a similar manner with the ARIS index
for individuals. From the DOT and the O*NET, from the total pool of occupational tasks, we
choose the tasks that help increase the automation-resistance of the jobs. The selection criteria is
60

the same as what is listed in Table 2.1. In the end, the AROT index for each occupation is
estimated by summing the usage ratings of nonroutine analytical tasks, nonroutine social
intelligence in tasks, nonroutine manual dexterity of tasks, and nonroutine interpersonal tasks—
and subtracting from it the extent of routinization and offshorability of the tasks as explained in
Eq. (7) above.
Unlike the ARIS index that does not vary across our analytic period, the AROT index is
set to change over time, thereby reflecting any changes in the automation-resistance of jobs
individuals currently hold in our sample.

2.3.2 Exploring the Automation-Resistance Indices

The ARIS and AROT indices devised in this chapter find their credibility in the fact that
their selection criteria (e.g., non-routine analytical skills, non-routine interpersonal skills, non-
routine physical adaptability, etc.) sufficiently reflects the technology-induced job polarization
observed in the last decades (Autor and Acemoglu, 2011). This section is devoted to explore
insights contained in the two indices regarding the evolution of automation-resistance of
skills/tasks.
Using the AROT index, we show in Figure 2.1 the evolution of the automation-
resistance by occupational groups as observed among the individuals in the HRS data. The figure
on the top shows occupational groups whose AROT scores are on the rise. It should be noted that
both the high-skilled (e.g., management jobs) and low-skilled occupations (e.g.,
construction/extraction jobs) are experiencing the increase in AROT over time. While on the
rise, the AROT of life-, physical-, and social-scientists fluctuates rather heavily.
61

The figure on the bottom shows occupations whose AROT scores decreased most
significantly in the past three decades. While it is difficult to ascertain how many jobs in each of
the three occupational groups are taken strictly by high-skilled or mid-skilled workers. However,
we should note that the occupations most vulnerable to the threats of automation are not manual
jobs mostly carried out by low-skilled individuals.
Next, we show the evolution of ARIS index of workers by gender in Figure 2.2. In our
data, we have three cohorts. The oldest cohort (Cohort 1) consists of individuals born before
1930. The middle cohort (Cohort 2) is made up of individuals born between 1931 and 1941.
Lastly, the youngest (Cohort 3) include individuals born between 1942 and 1959. While the
reason behind the irregularity is unknown, we can infer from the variations in the evolution of
ARIS across cohorts that the degree to which the society rewards the automation-resistance
changes over time—as suggested by Katz and Murphy (1992) and Acemoglu and Autor (2011).
Lastly, as can be seen in Figure 2.3, we observe slight differences in the way the ARIS
index scores change for individuals grouped based on their educational attainment. Those earned
up to high-school diplomas or GED certificates are equivalent to the ‘low-skilled’ individuals
discussed in the SBTC and TBTC frameworks in Section 2.2.1. Individuals who have bachelor’s
degrees from colleges can be viewed as the ‘mid-skilled’. Those who has J.D., M.D., or Ph.D.
are the ‘high-skilled’ workers. ARIS has risen the most for the low-skilled, while remaining
relatively stagnant for the mid-skilled. ARIS has fluctuated the most—risen up to early 2000s
and is now falling—for the high-skilled workers.

62

2.3.3 Analysis Sample

Individual-specific information (e.g., skills, demographics, or income) is drawn from the
Health and Retirement Study (HRS), a biennial survey representative of aging American
population.
40
The HRS contains information on retirement plans, retirement expectations,
employment history, health, health insurance, financial and housing wealth, income, Social
Security and pension, family structure, and basic demographics of age-eligible respondents and
their spouses/partners. Our analytic period of 1992-2014 allows our results to reflect the impact
of at least two decades of technological change. Restricted HRS with geographic and
occupational identifiers has been linked to the public HRS. The restricted HRS occupation
information allows me to identify the respondents’ occupations at the disaggregated, 4-digit
Census Occupational Classification (COC) System, thereby obtaining a comprehensive coverage
of most occupations available in the US economy.
41
The restricted geographic identifiers reveal
the states of residence of the HRS respondents, which are used to create state fixed-effects in the
estimation models.
In this study, sample data is restricted to include individuals between age 50 and 80 who
are fully retired. Individuals’ transitions from work to retirement—and back to work—are
identified via their official labor force statuses in each wave. Retirees are those who are fully-
retired and out of the labor force.
42
Following the definition used in the literature, unretirees are

40
Currently, a total of fourteen HRS waves are available from 1992 to 2016. The HRS includes the following
cohorts: HRS cohort born between 1931 and 1941, AHEAD cohort born before 1924, Children of Depression
(CODA) cohort born 1924 to 1930, War Baby (WB) cohort born 1942 to1947, Early Baby Boomer (EBB) cohort
born 1948 and 1953, and lastly, Mid Baby Boomer born between 1954 and 1959.
41
Once the COC codes are converted into 2010 Standard Occupation Codes (SOC), the occupation-specific task
contents information from the DOT and the O*NET are linked.
42
The HRS does not distinguish between individuals who leave the labor force voluntarily and those who are forced
to retire (e.g., due to layoffs). It could be that those who are forced to retire are more prone to unretiring than those
whose initial retirement decisions were their own. While such differences are not clearly delineated given the
63

identified in this chapter as individuals who return from full-retirement to partial-retirement,
part-time employment, and full-time employment as well as those returning from partial
retirement to full-time employment (Maestas, 2010). Here, the retirement-to-work transitions are
captured by the wave-to-wave changes in the labor statuses.
43
For estimation models, a dynamic
unretirement indicator is devised whereby it takes on a value 0 if respondent is fully-retired in
wave t-1 and t, and 1 if the respondent is fully-retired in wave t-1 and working in wave t.
44

From the rich set of variables provided by the HRS, we select the potential drivers of
unretirement decisions to be used as covariates. Among the demographic and socioeconomic
statuses, used are respondents’ age and its quadratic polynomial, gender, race, cohort,
educational attainment, marital status, number of dependents in the household, individual annual
earnings, annual household earnings less the respondent’s, and the household total net
wealth.
45,46,47
Once we convert the earnings variables into real dollars indexed to 2012 dollars
using the Consumer Price Index, we take a natural log of the variables. The wealth variable is
converted to real dollars as well, but we refrain from taking a natural log of the variable to avoid
dropping individuals with negative wealth (i.e., having greater debts than assets). Instead, the

limitation of the data at hand, we try to control for some of the consequences of such differences by controlling for
the individuals’ heterogeneous preferences for work and leisure in some of the models.
43
Without the preceding wave available, we impute the unretirees in the 1992, the first wave: We refer to questions
in the first wave that inquire about the year in which the respondent retired from the last job. If the year answered is
before between 1990-1991, and if the same person records that she/he is currently working full-time, then we count
such an individual as an unretiree.
44
Note that the HRS is conducted biennially. Thus, there is a two-year gap between wave t-1 and wave t.
45
Race categories are Caucasians, African-Americans, or others. Educational attainment is grouped into three: group
one if obtained GED or high-school diploma, group two if obtained college degrees, and group three if obtained
graduate degrees. Marital statuses are categorized into three: 1. If never married (reference), 2 if married or
partnered, and 3 if separated, divorced, or widowed. As for the cohorts, we group individuals into three: the first
cohort is the oldest, consisting of those born before 1930; the second cohort includes individuals born between 1931
and 1941; lastly, the third cohort—the youngest—consists of those born between 1942 and 1959.
46
As for the individual earnings, all retirees’ earnings are set to be zero—until they unretire and resume working.
The household income includes total net family income from all sources including employment, state benefits,
pensions and other assets. Respondents’ own earnings are excluded.
47
The net wealth includes information on the checking and savings accounts, bonds, stocks, deposits, mutual funds,
primary and secondary housing, and any other savings less debts.
64

wealth variable is Winsorized at the 0.1 percent level to avoid outliers—following the approach
of Shoven and Slavov (2014).
48
Also devised are binary indicator variables that identify whether
individuals have access to employer-sponsored private pension income, and whether they are
eligible for monthly Social Security benefit claims.
Current health is measured by two variables: First, we estimate an index of health
conditions that takes into account a number of critical, work-impeding medical conditions based
on doctors’ diagnoses. The index is normalized to have a mean of zero and a standard deviation
of one, with a higher value indicating worse health.
49
Additionally, presence of impairments in
daily living is traced as we estimate the average of the self-rated levels of difficulty associated
with six activities of daily-living (ADLs) that lasted at least six months.
50
As the employer-
sponsored health insurance coverage is one of the primary reasons for individuals to stay
working if they will be devoid of coverage in retirement (Blau and Gilleskie, 2006; French and
Jones, 2011), we identify whether or not individuals have access to Medicare and Medicaid, the
public insurance alternatives that are likely to reduce the individuals’ unretirement incentives.
As our sample consists of older individuals, their stock of hard- and soft- skills can
undergo attrition due to the rapid cognitive declines in the post-retirement years. Such declines
lead to a deteriorating of information processing capabilities at work, inadequate functioning in
daily life such as planning sequences of activities, an inability to make sound decisions, and

48
For retirees who remain retired instead of returning to work, their earnings will be set to zero. Net wealth includes
the value of checking and savings accounts, bonds, stocks, deposits, mutual funds, primary and secondary housing,
and any other savings less debt.
49
Eight conditions are considered: 1) high blood pressure or hypertension; 2) diabetes or high blood sugar; 3) cancer
or a malignant tumor of any kind except skin cancer; 4) chronic lung disease except asthma such as chronic
bronchitis or emphysema; 5) heart attack, coronary heart disease, angina, congestive heart failure, or other heart
problems; 6) stroke or transient ischemic attack (TIA); 7) emotional, nervous, or psychiatric problems; and 8)
arthritis or rheumatism.
50
The six items that jointly represent the level of difficulty associated with the ADLs are walking, dressing, bathing,
eating, getting in and out of bed, and using toilet. Each item has a 0-2 scale, with the value 2 representing most
difficulty.
65

inefficient use of economic and social resources—all of which can influence an individual’s
unretirement decisions (Park, 1999; Fillenbaum et al., 1988; Salthouse, 1999). Hence, we create
an indicator for the individuals’ current state of cognitive abilities using the information
available in the HRS. In each wave, we collect scores for the following five elements of
cognition known to decline with age: memory, mental status, vocabulary, dementia, and
numeracy (Lindenberger, Mayr and Kliegl, 1993; Park, 1995; Salthouse, 1985).
51
Once we
standardize each of the scores to have a mean of zero and a standard deviation of one, we take
the average of the five to generate an overarching index of cognitive abilities. Depending on the
rate of declines, the level of cognitive abilities is allowed to change (i.e., bearing within-
individual variations) over the course of the analytic period.
In examining the degree of psychological returns to unretirement and to skill-to-task
matching in unretirement job sorting, we make use of (1) a depression index known as the Center
for Epidemiologic Studies Depression Scale (CES-D; Radloff, 1977) and (2) doctor-diagnoses.
The CES-D score consists of a sum of the following eight items that require respondents’ self-
assessment: How much of the past week they felt (1) depressed, (2) like everything was an effort,
(3) sleep was restless, (4) happy, (5) lonely, (6) sad, (7) that they could not get going, and (8) that
they enjoyed life. Each of the items is in a 0-1 scale, and the overall CES-D score is, naturally, in
a scale of 0-8.
52
As the components of the CES-D are based on self-reports, we create an

51
Memory is measured by the number of words immediately recalled, number of words recalled after a delay, and a
serial-seven test whereby individuals are asked to keep subtracting seven and provide accurate answers; mental
status is ascertained by backward counting, date naming, object naming, and president naming; vocabulary is
measured by the number of words correctly defined; dementia is identified by individuals’ health status, by whether
or not they have ever been diagnosed with dementia; lastly, numeracy is ascertained by the accuracy of individuals’
calculation of lottery splits and interest on savings.
52
The CES-D index is normalized to have a mean of 0 and standard deviation of 1 for easier interpretations.
66

alternative depression measure that is based on doctors’ diagnoses of depression. This particular
variable takes on a binary form, with a value of 1 if diagnosed with depression and 0 if not.
Lastly, information on HRS respondents’ work-related personality traits are elicited from
the HRS Leave Behind (LB) questionnaire.
53
Using the fact that the personality traits do not
change significantly for middle-aged and older individuals (Costa and McCrae, 1997; Costa et
al., 2000), we calculate for each HRS respondent the average personality trait scores over time
and treat the score as constant in our analysis. Among the personality traits, we make use of the
following five that are work-related: organized, responsible, hardworking, intelligent, and
thorough. Each personality trait is measured on a 1-to-4 scale—with 4 indicating the highest
display of the trait. The LB is administrated to a rotating, random 50% of the core HRS sample
every other wave.

2.3.4 Summary Statistics

Our sample consists of 115,432 person-year observations. Once again, the sample
consists solely of older individuals who have already entered full, initial retirement from the
labor force. The average number of observations per individual across waves is six. Displayed in
Table 2.2 are the summary statistics of all variables used in the analyses. Assuming the sample is
devoid of egregious outliers, the sample data consists of approximately a half having access to
the employer-sponsored pension income and eligible for Social Security monthly benefits—and
another half without such inflows of money. Once again, assuming there are no outliers skewing
the sample average, slightly more individuals do not have Medicare coverage—due to their not

53
Since 2004, the HRS interviewers started conducting the LB self-administered questionnaires by leaving the LB
surveys with respondents upon the completion of an in-person Core Interview. Questions included in the LB surveys
are related to respondents’ subjective well-being, life-style, and life events, and other psychosocial information.
67

yet reaching age 65—than those who do. Such a story is possible if we acknowledge that many
choose to enter initial retirement before reaching the Medicare- and Social Security eligibility
ages of 65 and 62, respectively.
As for individual earnings, it is unsurprising to see a negative mean for the log earnings
given that all individuals in the sample are retirees who start out the analytic period by being in
the initial retirement and by earning zero earnings on their own. Based on the average value for
the challenges to the ADLs indicator—a higher value indicates more hardships experienced in
day-to-day living—individuals in the sample are fairly healthy and self-sufficient given their age
range. The fact that our sample is devoid of anyone who has stopped working due to disabilities
may be contributing to the low average of the indicator for the challenges to ADLs.
The average of 0.46 for the binary unretirement indicator suggests that a fair number of
individuals who have undergone initial retirement return to the labor force. As the weighted
average falls short of providing a clear picture on whether we have a sufficient pool of unretirees
in the sample, we show in Table 2.3 the number of unretirees present in each wave. It can be
seen that the phenomenon of unretirement grows steadily over the analytic period except in the
last two waves. Despite the two economic downturns between years 1992-1994 and 2006-2008,
no unusual spikes in the trends of unretirement are observed during the analytic period. We
further account for potential year-specific peculiarities (e.g., the recessions) that may encourage
or discourage unretirement by including year fixed-effects in all our estimation models. Details
on the modeling strategies are to follow in Section 2.4.
Going back to Table 2.2, cohort-specific weighted averages reveal that there is a
considerable difference among the three cohorts. The average unretirement likelihood seems to
grow from the oldest cohort (Cohort 1) to the youngest (Cohort 3), which is intuitive given that
68

the practice of retirees’ returning to the labor force is undergoing a continuous growth over the
years. The fact that the average ARIS- and AROT- index scores switch number sign—dropping
suddenly to a negative value—only for the middle cohort deserves attention. While the reason
behind the irregularity is unknown, we can infer that the degree to which the society rewards the
resistance to automation varies over time—as suggested by Katz and Murphy (1992) and
Acemoglu and Autor (2011). Given the variations in the unretirement likelihood, the average
level of resistance to automation, and other covariates, unretirement decisions and job choices of
an individual with certain preferences and skill-sets in cohort 1 are likely to be different from a
comparable person in cohorts 2 or 3. Such differences will be accounted for in the empirical
estimations by our including cohort fixed-effects.

2.4 Empirical Approach
2.4.1 Automation-Resistance and Unretirement Decisions
First, restricting our sample to include only those retired from the labor force, we
examine whether individuals’ unretirement decisions depend on the level of ARIS. Our linear
probability model is written as follows,

(8) 𝑈𝑛𝑟𝑒𝑡𝑖𝑟𝑒𝑑 𝑖𝑡𝑠 = 𝑎 + 𝛾 𝐴𝑅𝐼𝑆 𝑖 + 𝛽 𝑋 𝑖𝑡𝑠
+ 𝛿 𝑡 + 𝜑 𝑠 + 𝑢 𝑖𝑡𝑠 ,

where the subscripts i , t, and s, indicate individual, time and state, respectively. The dynamic
outcome (i.e., 𝑈𝑛𝑟𝑒𝑡𝑖𝑟𝑒𝑑 𝑖𝑡
) takes on a value 𝑗 = 0, 1 where the value 1 is equivalent to when an
individual i returns from full-retirement in time t-1 to partial-retirement, part-time employment,
and full-time employment in t as well as those transitioning from partial retirement in t-1 to full-
69

time employment in t.
54
The main explanatory variable of interest is the 𝐴𝑅𝐼𝑆 𝑖 index whereby a
higher index value represents a higher resistance to automation. In that the retirees—the sole
constituents of our sample data—no longer hold career jobs, the 𝐴𝑅𝐼𝑆 𝑖 index constructed from
their career-jobs’ skill-specific information does not undergo changes during the analytic period.
In short, the 𝐴𝑅𝐼𝑆 𝑖 index lacks within-individual variations across time.
In the baseline specification, 𝑋 𝑖𝑡
is a vector of baseline covariates consisting of
demographics, current health measures, and institutional drivers of unretirement: gender, race,
age and its quadratic polynomial, marital status, educational attainment, number of chronic
conditions diagnosed, difficulties associated with the ADLs , and whether respondents currently
have access to (1) Social Security benefits, (2) private pension income, and (3) Medicare.
55,56
For
the health conditions indicator, we sum the number of critical, work-impeding medical
conditions diagnosed by doctors and normalize it to have a mean of zero and a standard deviation
of one—with a higher value corresponding to worse health.
57
The number of limitations with the
ADLs has also been normalized—with a higher value indicating a greater difficulty.
58
The

54
Our sample is restricted to include only the retirees. However, as retirement is what we typically consider as the
final phase of labor supply life cycle, we do not expect any incidental truncation of the sample as we consider only
the aging individuals who have chosen to retire as our analysis pool.
55
Gender takes a value 1 if male, 0 if female. Race takes on a value 1 if Caucasians, 2 if African-Americans, and 3
if others. Marital statuses are categorized into three: 1. If never married (reference), 2 if married or partnered, and 3
if separated, divorced, or widowed. Education also has three groups: 1 if obtained GED or high school diploma
(reference), 2 if obtained college degrees, and 3 if obtained graduate degrees.
56
Whether or not retirees have access to Social Security benefits, private pension income, and Medicare influence
their unretirement decisions. The first two relate to how much income the retirees have—and a shortage may drive
them back to work. The last component is about having access to health insurance to brace for unforeseen future
medical expenses (Cutler and Zeckhauser, 2000). Unless individuals are covered by Medicare, their lack of
insurance may incentivize the retirees to seek for jobs that include employer-sponsored insurance.
57
The conditions include 1) high blood pressure or hypertension; 2) diabetes or high blood sugar; 3) cancer or a
malignant tumor of any kind except skin cancer; 4) chronic lung disease except asthma such as chronic bronchitis or
emphysema; 5) heart attack, coronary heart disease, angina, congestive heart failure, or other heart problems; 6)
stroke or transient ischemic attack (TIA); 7) emotional, nervous, or psychiatric problems; and 8) arthritis or
rheumatism.
58
The criteria included in the ADLs are difficulties associated with (1) walking across room, (2) dressing, (3)
bathing or showering, (4) eating, (5) getting in/out of bed, and (6) using the toilet—without help from others.
70

socioeconomic variables are purposefully excluded in the baseline model so that any impact of
the automation-resistance on unretirement inclinations captured by the model encapsulates both
the financial- and non-financial drives back to work. Included are year (𝛿 𝑡 ) and state (𝜑 𝑡 ) fixed-
effects to account for year-specific and region-specific labor market and economic peculiarities.
Standard errors are clustered at the individual level.
Several robustness checks are conducted. Because we are unable to include in the
baseline specification individual fixed-effects due to the lack of within-individual variations in
the ARIS index, we account for some of the unobserved preferences dictating the extent of
resistance to automation as well as unretirement decisions by including as controls five work-
related personality variables.
59
The personality controls measure the extent to which respondents
rate themselves—on a 1-to-6 scale, with 6 indicating the highest display of the trait—as
organized, responsible, hardworking, intelligent, and thorough. Next, we assign a new
specification by including the following socioeconomic controls so that the remaining effect
represents a non-financial impact of the extent of resistance to automation on unretirement
decisions: individual earnings, total household income less the respondent’s earnings, and
household wealth.
We also try regressing unretirement on 𝐴𝑅𝐼𝑆 𝑖 , baseline covariates, and the personality
controls while replacing the age variable with cohort fixed-effects (ν
c
). Our including the cohort
fixed-effects controls for inter-generational differences in estimating the impact of the 𝐴𝑅𝐼𝑆 𝑖 .
60

59
The richness of the HRS data enables us to proxy for such tastes using respondents’ work-related personality traits
as elicited by the Leave Behind (LB) questionnaire. Since 2004, the HRS interview includes the LB self-
administered questionnaire, eliciting respondents’ subjective well-being, life-style, and life events, and other
psychosocial information. The LB is administrated to a rotating, random 50% of the core HRS sample every other
wave. Using the fact that the personality traits do not change significantly for middle-aged and older individuals
(Costa and McCrae, 1997; Costa et al., 2000), we calculate for each HRS respondent the average personality trait
scores over time and treat the score as constant.
60
The extent to which the society rewards automation-resistance has grown over time. Thus, the unretirement
sorting tendencies of an individual in one cohort is likely to be different from a comparable person in the subsequent
71

Next, we estimate the magnitude of the cross-cohort differences by interacting the 𝐴𝑅𝐼𝑆 𝑖 index
with cohort dummies (i.e., 𝐴𝑅𝐼𝑆 𝑖 × 1(𝐶𝑂𝐻𝑂𝑅 𝑇 𝑖 ))—thereby testing how much the importance
of the resistance to automation has grown over time. Lastly, we try interacting the 𝐴𝑅𝐼𝑆 𝑖 index
with a time-variant indicator for the individuals’ current level of cognitive abilities—which are
known to decline with age. Included in the last specification are individual fixed-effects (𝜏 𝑖 ) to
account for individuals’ unobserved preferences for work, leisure, and automation that can
potentially have an effect both on their extent of resistance to automation and their unretirement
decisions.
Next, in lieu of the linear probability models, we re-run all specifications using the
following logit model,

(9) 𝑃𝑟 𝑖𝑡𝑘 = 𝑃𝑟 (𝑈𝑛𝑟𝑒𝑡𝑖𝑟𝑒𝑑 𝑖𝑡
= 𝑘 |𝑋 𝑖𝑡
, 𝐴𝑅𝐼𝑆 𝑖 ) =
exp (𝑉 𝑖𝑡𝑘 )
∑ exp (𝑉 𝑖𝑡𝑗 𝑗 )
; k ∈ 𝐽

where 𝑉 𝑖𝑡𝑠𝑗 = 𝛼 𝑗 𝑋 𝑖𝑡𝑗 + 𝛽 𝑗 𝐴𝑅𝐼𝑆 𝑖 + 𝛿 𝑡 + 𝜑 𝑠 . The outcome variable indicating unretirement
likelihood, the baseline vector of 𝑋 𝑖𝑡
, and the 𝐴𝑅𝐼𝑆 𝑖 index are specified in the same as those in
Eq. (8). All covariates specifications remain the same as those of the linear probability model.
We report average marginal effects with Delta Method standard errors clustered at the individual
level.

cohort. To empirically account for such inter-generational differences, we include the cohort fixed-effects. For the
same reason, cohort fixed-effects are included in a robustness check in the subsequent estimations in Sections 2.4.2.
and 2.4.3.
72

2.4.2 Unretirement Job Choices

Next, for unretirees, we explore the extent to which their 𝐴𝑅𝐼𝑆 𝑖 dictates the types of jobs
they choose. The primary nature of the jobs examined is the task-specific automation resistance
(i.e., AROT) of the jobs. The degree to which unretirees put their primary skills to use in their
new occupations is worth examining in that the skill-to-task matching—and thereby matching
between the 𝐴𝑅𝐼𝑆 𝑖 and 𝐴𝑅𝑂𝑇 𝑖𝑡
—is often imperfect, despite individuals’ efforts to secure jobs
that best rewards their skills and their automation-resistance.
Given that we can study the job choices of individuals if and only if they unretire, our
pool of subjects may now suffer from non-random sampling biases. Thus, we perform the
Heckman two-stage correction to account for potential incidental truncation (Briggs, 2004;
Vella, 1998; Semykina and Wooldridge, 2010). First, we write the outcome equation whereby
the unretirees’ job choices are dictated by 𝐴𝑅𝐼𝑆 𝑖 ,

(10) 𝐴 𝑅𝑂𝑇 𝑖𝑡𝑠 = 𝛼 + 𝛾 𝑋 𝑖𝑡
+ 𝛽 𝐴𝑅𝐼𝑆 𝑖 + 𝛿 𝑡 + 𝜑 𝑠 + 𝜖 𝑖𝑡𝑠 ,

where 𝐴𝑅𝑂𝑇 𝑖𝑡𝑠 is the extent of resistance to automation embedded in the newly-obtained
occupation of an unretiree i at wave t. 𝑋 𝑖𝑡
is a vector of covariates that dictate unretirees’ job
choices, which is a subset of the covariates used in the selection equation shown below
(Eq.(11)): gender, race, age and its quadratic polynomial, marital status, highest educational
attainment, number of chronic conditions diagnosed, and difficulties associated with the ADLs.
61

61
The 𝑋 𝑖𝑡
is the subset of that of the selection equation to abide by the exclusion restriction of the Heckman-type
correction method. The estimator requires the availability of at least one regressor in the selection equation that is
conditionally independent from the dependent variable in the outcome equation (Greene, 1993; Lalonde, 1986;
Lewbel, 2005).
73

Also included are the year (𝛿 𝑡 ) and state (𝜑 𝑠 ) fixed-effects. The formulation of the covariates is
consistent with that of the Eq.(8) or Eq.(9) presented in Section 2.4.1.
The outcome equation shown in Eq.(10) is conditional on that individuals have chosen to
unretire, which is in turn determined by a selection equation. Assumed is that errors in the
selection function are normally distributed, making the Heckman correction estimator parametric
in nature. We specify a selection function for unretirement decisions, which is then used to
estimate the parameters using maximum likelihood (ML):

(11) P(𝑈𝑛𝑟𝑒𝑡𝑖𝑟𝑒 𝑑 𝑖𝑡𝑠 = 1|𝑍 𝑖𝑡𝑠 , 𝐴𝑅𝐼𝑆 𝑖 )=P(-𝜇 𝑖𝑡𝑠 < 𝑉 𝑖𝑡𝑠 |𝑍 𝑖𝑡𝑠 , 𝐴𝑅𝐼𝑆 𝑖 ) = 𝛷 (𝑉 𝑖𝑡𝑠 ),

where 𝑉 𝑖𝑡𝑠 = 𝛼 𝑍 𝑖𝑡
+ 𝛽 𝐴𝑅𝐼𝑆 𝑖 + 𝛿 𝑡 + 𝜑 𝑠 , and the dynamic unretirement outcome 𝑈𝑛𝑟𝑒𝑡𝑖𝑟𝑒 𝑑 𝑖𝑡𝑠
as well as the 𝐴𝑅𝐼𝑆 𝑖 index are specified in the same manner as those of Eq.(8) in Section 2.4.1.
The 𝜇 𝑖𝑡𝑠 is the latent covariate, and Φ is standard normal cumulative distribution function.
62
The
vector of baseline covariates, 𝑍 𝑖𝑡𝑠 , is also the same as that of the estimation Eq.(8) in Section
2.4.1.: demographics, current health measures, and institutional drivers of unretirement: gender,
race, age and its quadratic polynomial, marital status, educational attainment, number of chronic
conditions diagnosed, difficulties associated with the ADLs, and whether respondents currently
have access to Social Security benefits, private pension income, and Medicare. Also included are
the year (𝛿 𝑡 ) and state (𝜑 𝑠 ) fixed-effects. The formulation of the covariates is consistent with
that of the linear probability- and logit- models presented in Section 2.4.1.

62
The Heckman model requires three statistical assumptions: (1) (𝜖 𝑖𝑡
, 𝜇 𝑖𝑡
) are iid in I with a standard normal
distribution, (2) {𝑋 𝑖𝑡
: 𝑖 = 1, … , 𝑁 } is independent of {𝜖 𝑖𝑡
: 𝑖 = 1, … , 𝑁 }, and (3) {𝑍 𝑖𝑡
: 𝑖 = 1, … , 𝑁 } is independent of
{𝜇 𝑖𝑡
: 𝑖 = 1, … , 𝑁 }.
74

Assuming normal distributions of the outcome equation (Eq. (10)) and the selection
function (Eq. (11)), we use the estimated parameters from the ML to compute the inverse Mills
ratio when 𝑈𝑛𝑟𝑒𝑡𝑖𝑟𝑒 𝑑 𝑖𝑡𝑠 = 1 and when 𝑈𝑛𝑟𝑒𝑡𝑖𝑟𝑒 𝑑 𝑖𝑡𝑠 = 0 based on the symmetry of the normal
distribution:

(12) λ
its
=
𝜙 (𝑡 )
1−Φ(𝑡 )
=
𝜙 (𝑡 )
Φ(𝑡 )
,

where t is the point at which the distribution has been truncated, 𝜙 is the standard normal density
function, and Φ is standard normal cumulative distribution function. Assuming the error term in
the outcome equation and the latent variable in the selection function—𝜖 𝑖𝑡𝑠 and 𝜇 𝑖𝑡𝑠 ,
respectively—have a correlation ρ, it is that

(13) 𝐸 (𝜖 𝑖𝑡𝑠 |𝑋 𝑖𝑡𝑠 , 𝐴𝑅𝐼𝑆 𝑖 ) = 𝜌 𝜆 𝑖𝑡𝑠 (𝐴𝑅𝐼𝑆 𝑖 , 𝑉 𝑖𝑡𝑠 )

where the individual-specific Mill’s Ratio is equal to

(14) λ
its
(𝐴𝑅𝐼𝑆 𝑖 , 𝑉 𝑖𝑡𝑠 ) = 𝐴𝑅𝐼𝑆 𝑖 (
𝜙 (𝑉 𝑖𝑡𝑠 )
1−Φ(𝑉 𝑖𝑡𝑠 )
) + (1 − 𝐴𝑅𝐼𝑆 𝑖 )(
−𝜙 (𝑉 𝑖𝑡𝑠 )
Φ(𝑉 𝑖𝑡𝑠 )
).

Once the Eq. (14) is estimated, we include λ
its
(𝐴𝑅𝐼𝑆 𝑖 , 𝑉 𝑖𝑡𝑠 ̂
) in the outcome equation,

(15) 𝐴𝑅𝑂𝑇 𝑖𝑡𝑠 = 𝛼 + 𝛾 𝑋 𝑖𝑡𝑠 + 𝛽 𝐴𝑅𝐼𝑆 𝑖 + 𝛿 𝑡 + 𝜑 𝑠 + 𝜂 λ
its
(𝐴𝑅𝐼𝑆 𝑖 , 𝑉 𝑖𝑡𝑠 ̂
) + 𝜖 𝑖𝑡𝑠 ∗
,

where 𝜖 𝑖𝑡𝑠 ∗
= 𝜎 𝜖 𝑖𝑡𝑠 − 𝜂 λ
its
(𝐴𝑅𝐼𝑆 𝑖 , 𝑉 𝑖𝑡𝑠 ̂
), 𝜎 is estimated as a function of residuals from the
regression equation, and 𝜂 is estimated, 𝜂 ̂ = 𝜎 ̂ 𝜌 ̂ where 𝜌 is the correlation coefficient between
75

the error term of the outcome equation and the latent variable in the selection function. Finally,
using GLS for efficient standard errors and consistently-estimated coefficients, the coefficients
as well as the selection bias parameter (i.e., 𝜂 ̂ ) of the outcome equation are estimated.
As for covariates in the models, the baseline specification as well as additional
specifications used for robustness checks used for Eq.(8) in Section 2.4.1. have been applied
once again to the selection equations. Subsets of the specifications have been used for the
corresponding outcome equations.
63
Next, we explore whether unretirees with higher 𝐴𝑅𝐼𝑆 𝑖 sort
into occupations that are white-collar jobs in lieu of blue-collar, in the same industry as the
career jobs or in a different industry, or in STEM jobs as opposed to non-STEM. Setting each of
the said traits as a binary outcome variable in the outcome equation Eq.(10), we re-run the
Heckman correction estimations with the same selection equation as Eq.(11). As for covariates,
the baseline specification and personality controls are used. For all regressions on unretirement
job sorting, standard errors are clustered at the individual level.
64

2.4.3 Returns to Skill-to-Task Matching

In the final part of the estimation, we study the monetary and mental health benefits of
skill-to-task matching. Here, we are interested in examining the degree to which unretirees’
automation-resistance (dictated by skills) matches that of their jobs (determined by the
occupational task contents)—or simply, their ARIS-AROT matching. Since such aspects can be

63
The list of covariates for the outcome equation is a subset of those used for its selection equation. The subsets are
almost identical to their matching selection equations’ covariates except the access to private pension benefits,
Social Security benefits, and Medicare. Nevertheless, unlike the case in Eq.(6) in Section 2.4.1., individual fixed
effects are not included in the last specification for Eq.(10).
64
The level of the robustness in standard errors is somewhat reduced as the two-stage correction is conducted based
on a bivariate normality. The two-stage correction is conducted based on a bivariate normality. In lay terms, the two-
stage method used in this study assumes the errors in the outcome equation (i.e., second stage) and the selection
equation (i.e., first stage) are simultaneously normal, unlike a univariate normality that relaxes such an assumption.
76

discussed only for those who choose to unretire, two-stage Heckman type correction methods
similar to the ones described in in Section 2.4.2. are used. We run the following outcome
equation in the second stage of the correction:

(16) ln(wage )
𝑖𝑡𝑠 = 𝛼 + 𝛽 𝐴𝑅𝐼𝑆 𝑖 + ς𝐴𝑅𝑂𝑇 𝑖𝑡
+ 𝜌 (𝐴𝑅𝐼𝑆 𝑖 × 𝐴𝑅𝑂𝑇 𝑖𝑡
) + 𝛾 𝑋 𝑖𝑗𝑡 +
𝛿 𝑡 + 𝜑 𝑠 + 𝜖 𝑖𝑡𝑠 .

The specification of variables and parameters are consistent with that of Eq. (10) in
Section 2.4.2. The selection function is the same as the Eq. (11) in Section 2.4.2. except all is
conditional not solely on 𝐴𝑅𝐼𝑆 𝑖 but also on 𝐴𝑅𝑂𝑇 𝑖𝑡
and (𝐴𝑅𝐼𝑆 𝑖 × 𝐴𝑅𝑂𝑇 𝑖𝑡
). A main reason that
we include the interaction term is that comparing wage levels across different individuals who
choosing jobs with varying 𝐴𝑅𝑂𝑇 𝑖𝑡
is not an optimal way of measuring the quality of job-
matching. Since all individuals are inclined to sort into jobs that best reward their skills, the
difference across different individuals’ wages does not always signify the differences in their
quality of skill-to-job matching. Hence, we attempt to resolve this issue by conducting a type of a
‘within-individual’ examination whereby we compare the wage levels of the same individual as
she transitions from a worse- to a better- matching job, using the interaction term symbolizing
the ARIS-AROT matching. Specifically, in Eq.(16), our including the interaction term lets us
measure whether the monetary returns to 𝐴𝑅𝐼𝑆 𝑖 increases when the same unretiree switches from
a worse matching job to a better matching job.
As for the covariates, the same baseline covariates used for Eq.(8) in Section 2.4.1. are
employed—with the exception of private pension income, Social Security benefits, and Medicare
coverage. Then, in another specification, we try adding to the baseline covariates the work-
related personality controls and socioeconomic controls. Next, we include in a third specification
77

the baseline covariates, personality controls, cohort fixed-effects, and an interaction between the
𝐴𝑅𝐼𝑆 𝑖 index and the cohorts fixed-effects.
For the mental health returns, the outcome variables used are (1) the CESD index that
measures self-reported levels of depression and (2) a binary indicator for the doctor-diagnosed
depression (i.e., 1 yes, 0 no). For covariates, included first are the baseline covariates, work-
related personality controls, and socioeconomic controls. Next, we try including the baseline
covariates, personality controls, and cohort fixed-effects. Lastly, we add to the previous
specification an interaction term between the 𝐴𝑅𝐼𝑆 𝑖 index and the cohorts fixed-effects.

2.5 Results
2.5.1 Who Returns to Work?
Table 2.4 presents the association between retirees’ 𝐴𝑅𝐼𝑆 𝑖 and their decisions to return to
the labor force. The table only includes results from linear probability models. We only show the
results for regressors directly related to individuals’ automation resistance. Detailed results are
shown in Table 2.4A in the Appendix C. Controlling for a rich set of baseline covariates, we
show that a one standard deviation increase in the ARIS index leads to a 0.7 percentage point
increase, or a 1.9% increase from the mean of 35.9% unretirement, holding all else constant
(column 1). When the work-related personality traits are controlled to account for some of the
endogeneity stemming from unobserved preferences influencing both individuals’ skill choices
as well as unretirement decisions (column 2), the impact of a one standard deviation increase in
𝐴𝑅𝐼𝑆 𝑖 increases the unretirement probability by 0.7 percentage point, again a 1.9% increase.
Despite the significant role played by the personality variables as shown in Table 2.4A, the
impact of the 𝐴𝑅𝐼 𝑆 𝑖 on unretirement likelihood remains unaltered with the inclusion of the
78

personality traits. Assuming that our work-related personality controls do a fair job capturing the
previously-unobserved preferences for work and leisure, the estimated coefficient shown in
column 2 implies that the association between individuals’ 𝐴𝑅𝐼𝑆 𝑖 and unretirement decisions is
robust to heterogeneous preferences. As for the estimated coefficients of the personality controls,
on one hand, it looks as if the more hardworking and intelligent an individual is, the more she is
likely to come back to work from retirement. On the other hand, a retiree’s extent of being
thorough, organized, and responsible do not increase her unretirement chances. Such patterns
portrayed by the personality traits remain consistent through all model specifications in columns
2-5.
When we control for individual- earnings, other household- earnings, and household
wealth in addition to the personality controls and the baseline covariates (column 3), a one
standard-deviation increase in 𝐴𝑅𝐼𝑆 𝑖 increases the unretirement probability by 0.6 percentage
point, or 1.7%. increase in unretirement from the mean of 35.9%. This indicates that the income
and wealth have an overall negative effect on unretirement likelihood. By taking a closer look at
the estimated coefficients of the socioeconomic controls in Table 2.4A, we notice that the
estimated coefficients of the other household earnings and household wealth have negative
associations with the unretirement likelihood. In that a sense of financial security induced by
sufficient household income and wealth would give a retiree less reasons to unretire, the negative
association is reasonable.
In columns 4, our controlling for inter-generational differences in the unretirement trends
via including the cohort fixed-effects wipes away the statistical significance from the 𝐴𝑅𝐼𝑆 𝑖
index. The opposing directionality and the significance of the two cohort fixed-effects included
in columns 4 and 5 indicate that the unretirement decisions depend on cohort-specificity (e.g.,
79

the culture or the perception of unretirement during the time of each generation). In column 5,
when we include in our model the interaction term between the 𝐴𝑅𝐼 𝑆 𝑖 and the cohort dummies,
we see that the degree to which retirees’ 𝐴𝑅𝐼𝑆 𝑖 influences their unretirement decisions grows
from a preceding cohort to a following cohort—confirming that the society’s valuation of the
resistance to automation has grown over time. Nevertheless, such inter-generational differences
are not shown to be statistically significant.
Lastly, in the last specification shown in column 6, we add to the baseline covariates (1)
an interaction term estimating the joint effect between the 𝐴𝑅𝐼𝑆 𝑖 index and a time-variant
indicator for the individuals’ current level of cognitive abilities, and (2) individual fixed-effects
to better account for individuals’ unobserved preferences for work, leisure, and automation.
65

The estimated coefficient of the interaction term (𝐴𝑅𝐼𝑆 𝑖 ) x (𝑐𝑜𝑔𝑛𝑖𝑡𝑖𝑣𝑒 𝑎𝑏𝑖𝑙𝑖𝑡𝑖𝑒𝑠 𝑖𝑡
) suggests that
a one standard-deviation rise in the cognitive abilities increases the effect of 𝐴𝑅𝐼𝑆 𝑖 on
unretirement likelihood by 0.6 percentage point, or 1.7% increase. The estimation is significant
at 10-percent level.
As all results presented in Table 2.4 are linear probability models, we re-run all
regressions using logit models. Both the main and the marginal effects of the logit models are
shown in Table 2.5.
66
Detailed results are shown in Table 2.5A in Appendix C. The panel logit
estimations in Table 2.5 and the linear probability models in Table 2.4 show consistent results
with respect to the impact of the 𝐴𝑅𝐼𝑆 𝑖 (i.e., consistent statistical significance and the
directionality) on unretirement likelihood regardless of covariate specifications. The marginal
effect of the logit estimations show that one standard deviation increase in 𝐴𝑅𝐼𝑆 𝑖 is associated

65
In that the estimation using the specification shown in column 6 if a fixed-effect model, coefficients of the all
time-invariant covariates are no longer estimated.
66
Given the limitation of the logit model, the specification shown in column 6 no longer includes individual fixed-
effects. All time-invariant covariates are preserved in the model.
80

with a 0.6 percentage point increase in unretirement likelihood (1.5% increase from the mean of
40% unretirement) with baseline covariates, a 0.6 percentage point increase (1.5% increase) with
baseline covariates and personality controls, and a 0.5 percentage point increase (1.3% increase)
with baseline-, personality-, and socioeconomic controls. Also notable is the result in column 6
where a one standard deviation increase in 𝐴𝑅𝐼𝑆 𝑖 leads to a 1.4 percentage point increase (3.5%
increase) in unretirement likelihood as individuals’ cognitive abilities improve.

2.5.2 What Types of Jobs Do the Unretirees Choose?

Next, using the two-stage Heckman correction method, we evaluate whether individuals’
𝐴𝑅𝐼𝑆 𝑖 dictate their job choices as they unretire. Summary results are shown in Table 2.6.
Detailed results for all covariates are listed in the Appendix C, in Table 2.6A. All models shown
in Table 2.6 examine one particular nature of the newly-obtained jobs: the automation resistance
of occupational tasks (𝐴𝑅𝑂𝑇 𝑖𝑡
).
Overall, regardless of covariates specifications, we find that unretirees with high
𝐴𝑅𝐼𝑆 𝑖 choose jobs that have higher task-specific automation resistance. Specifically, using the
baseline covariates in column 1, we observe that a one standard-deviation increase in an
individual’s 𝐴𝑅𝐼𝑆 𝑖 is associated with getting a job with a higher 𝐴𝑅𝑂𝑇 𝑖𝑡
by 0.772 standard-
deviations. While the estimated coefficient of 𝐴𝑅𝐼𝑆 𝑖 lowers slightly as we include work-related
personality controls (column 2), the directionality and the statistical significance of the estimated
coefficient stays consistent.
According to the TBTC framework discussed in Section 2.2.1., the extent of skill-to-task
allocations in the economy are set by the labor demand side as they react to the task-based nature
and the pace of the technological changes. As the framework is silent on how much of the
81

matching is driven by the voluntary choice of the labor supply (i.e. unretirees), we test the
existence and the magnitude of the supply-driven matching by controlling for the financial
incentives injected by the demand side. In column 3 of Table 2.6, as we control for wages—
thereby holding constant the financial incentives offered by the employers to reward and attract
certain workers based on their levels of 𝐴𝑅𝐼𝑆 𝑖 —the sorting tendencies of an individual with
higher 𝐴𝑅𝐼𝑆 𝑖 to flock into jobs with higher 𝐴𝑅𝑂𝑇 𝑖𝑡
are reduced, and yet are still positive and
statistically significant.
67
Specifically, holding constant the effect of the monetary rewards
injected by the labor demand side constant, we still observe that unretirees with one higher
standard-deviation of 𝐴𝑅𝐼𝑆 𝑖 flock to jobs with greater 𝐴𝑅𝑂𝑇 𝑖𝑡
by 0.755 standard-deviation. This
suggests a possibility that the remaining sorting tendencies are driven by non-financial
incentives. Moreover, the result indicates that the wages have a positive sorting effect, and the
coefficients associated with the socioeconomic controls as shown in Table 2.6A confirms the
conjecture: An increase in the log earnings—indicating a higher valuation of resistance to
automation by the labor demand side—by one percentage point induces sorting into an
occupation with greater resistance to automation by 0.011 standard deviation.
Based on the results shown in columns 5, it can be seen that the high-𝐴𝑅𝐼𝑆 𝑖 -to-high-
𝐴𝑅𝑂𝑇 𝑖𝑡
sorting tendencies decrease as we move from the youngest cohort (born before 1930) to
the 2
nd
and the 3
rd
cohorts (born in 1931-1941 and 1942-1959, respectively). Such a result draws
a contrast from that of Table 2.4 whereby the extent to which the 𝐴𝑅𝐼𝑆 𝑖 encourages unretirement
is shown to grow across cohorts. Digesting the results of the cohort-specific unretirement
decisions and unretirement job sorting together, we think it could be that (1) a higher 𝐴𝑅𝐼𝑆 𝑖

67
We assume that wage represents the sole financial rewards embedded in a job by the demand side, thereby
ignoring any fringe benefits.
82

increases unretirement likelihood, and (2) the society’s valuation of 𝐴𝑅𝐼𝑆 𝑖 has grown over time,
but (3) the younger generations of unretirees are less likely to find jobs whose task-specific
automation resistance best match their 𝐴𝑅𝐼𝑆 𝑖 due to rapid job-destruction, automation, etc. In
other words, the reduction in the sorting tendencies of unretirees with high 𝐴𝑅𝐼𝑆 𝑖 to jobs with
high 𝐴𝑅𝑂 𝑇 𝑖𝑡
across cohorts could be involuntary.
Lastly, as we include an interaction term between 𝐴𝑅𝐼𝑆 𝑖 and the current level of
cognitive abilities (column 6), we observe that a one standard-deviation increase in 𝐴𝑅𝐼𝑆 𝑖 and
the cognitive abilities jointly increase the 𝐴𝑅𝑂𝑇 𝑖𝑡
of the newfound job by 0.080 standard
deviations. The synergy effect is statistically significant, implying that the aging individuals’
cognitive abilities help accentuate their 𝐴𝑅𝐼𝑆 𝑖 -induced unretirement job preferences.
Next, we examine additional features of unretirees job choices. Results are shown in
Table 2.7, with detailed results in Table 2.7A in Appendix C. There are three criteria we use:
whether or not unretirees (1) obtain blue- versus white- collar occupations, (2) return to the same
industry where they held career jobs, and (3) find jobs in the field of science, technology,
engineering, and mathematics (STEM) that are rapidly growing. We re-run separate heckman
two-stage correction estimations using the three criteria as outcome variables. As for covariates,
we are consistent in including in the models shown in columns 1-3 the baseline covariates and
the work-related personality controls to account for the unobserved preferences.
It can be seen that, on average, a one standard-deviation increase in an individual’s 𝐴𝑅𝐼𝑆 𝑖
is associated with a rise in the likelihood of her choosing a white-collar job by 4.2 percentage
point (or 61.8% increase from the mean of 6.8%), a reduction in the likelihood of her choosing a
job in the same industry as her career job by 1.1 percentage point (2.52% decrease from the
mean of 43.6%), and a rise in her choosing a STEM job as opposed to non-STEM occupations
83

by 2.9 percentage point (74.4% increase from the mean of 39.0%). From the results, it can be
inferred that an unretiree’s high automation-resistance enables her to seek employment in white-
collar and STEM-related professions—both of which are financially more rewarded by the
demand side. Moreover, the high 𝐴𝑅𝐼𝑆 𝑖 may be opening doors to employment that bear little
relevance to unretirees’ career jobs, outside their main industry and specialization.

2.5.3 Financial and Mental Health Returns to Skill-to-Task Matching

The estimated financial and mental health returns to skill-to-task matching (i.e., 𝐴𝑅𝐼𝑆 𝑖 -to-
𝐴𝑅𝑂𝑇 𝑖𝑡
matching) are shown in Table 2.8. For detailed results including the estimated
coefficients for all covariates in the models, refer to Table 2.8A in Appendix C. From our
models on the financial returns, we run three estimations shown in columns 1-3. As we include
baseline covariates (column 1) and additional personality controls (column 2), we see that a
higher 𝐴𝑅𝐼𝑆 𝑖 -to-𝐴𝑅𝑂𝑇 𝑖𝑡
matching results in a decrease in earnings by 65.1% and 61.8%,
respectively. In plain words, when an unretiree with a given 𝐴 𝑅𝐼𝑆 𝑖 switches from a worse- to a
better- matching job in terms of the automation-resistance, her securing a better matching job
results in a decrease in earnings.
We run an additional model exploring the monetary return to the skill-to-task matching
while taking into account the differences across cohorts (i.e., the evolution of how the society
and members of each cohort values the resistance to automation). As shown in column 3, we find
that the 𝐴𝑅𝐼𝑆 𝑖 -to-𝐴𝑅𝑂𝑇 𝑖𝑡
matching is rewarded more and more across time from a preceding
cohort to next: Compared to the earliest cohort, a member of the middle cohort is likely to
experience a growth in earnings by 82.6% as she moves from a worse to a better matching job; a
member of the youngest cohort sees a rise in earnings by 158.0%. Based on the drastic
84

differences in the financial returns to the 𝐴𝑅𝐼𝑆 𝑖 -to-𝐴𝑅𝑂𝑇 𝑖𝑡
matching across columns 1-3, we
deduce that the cohort-specificity (i.e., cohort fixed-effects and cohort specific skill-to-task
interactions) plays a crucial role in estimating the monetary returns to the 𝐴𝑅𝐼𝑆 𝑖 -to-
𝐴𝑅𝑂𝑇 𝑖𝑡
matching. The extent to which the society rewards the 𝐴𝑅𝐼𝑆 𝑖 -to-𝐴𝑅𝑂𝑇 𝑖𝑡
matching has
been growing significantly over time.
Next, we move to examine the mental health returns to 𝐴𝑅𝐼𝑆 𝑖 -to-𝐴𝑅𝑂𝑇 𝑖𝑡
matching, using
two indicators of workers’ mental states: (1) C-ESD scores and (2) a binary indicator for
diagnoses of depression. As for the C-ESD score, it can be considered a comprehensive self-
evaluated measure of depressive symptoms. A higher score indicates having worse depressive
symptoms. The doctor-diagnosis of depression implies a more objective measure based on
experts’ medical opinion. Both sets of results are discussed simultaneously. We run the two-
stage estimations by including in the models the baseline covariates, personality controls, and
socioeconomic controls (columns 4 and 7), and observe that a better 𝐴𝑅𝐼𝑆 𝑖 -to-𝐴𝑅𝑂𝑇 𝑖𝑡
matching
results in a reduction in the depressive symptoms by 0.005 CES-D points, and a reduction in the
diagnosis of depression by 0.1 percentage points (0.6% decrease from the mean of 17.1%).
Between the two, only the coefficient of the CES-D is significant. However, as we take into
consideration the cohort-specificities in columns 5-6 and 8-9, the statistical significance of the
results associated with the CES-D scores and the diagnosed depression grows. In the end, we can
infer that an unretiree’s achieving a better skill-to-task matching in terms of automation
resistance results in positive mental health benefits, as shown by the reduced depressive
symptoms.

85

2.6 Discussion

Despite the growing number of unretirees in the labor market, little is known about the
phenomenon or the motives behind individuals’ unretirement decisions (Cahill et al., 2010;
Maestas, 2010). In this chapter, we have explored the effect of automation on retirees’
unretirement decisions by creating metrics for individuals’ varying extents of automation-
resistance. Then, for unretirees, we have examined how much individuals’ skill-specific
automation resistance dictates their job choices—especially the task-specific automation
resistance of the jobs themselves. In doing so, we were interested in finding out whether an
unretiree with a high 𝐴𝑅𝐼𝑆 flocks to a job with high 𝐴𝑅𝑂𝑇 , thereby resulting in a good skill-to-
task match. Lastly, we have explored potential financial and mental health returns to better skill-
to-task matching in terms of automation resistance.
Drawing data from the 1992 to 2014 Health and Retirement Study and the Occupational
Information Network and employing logit models, we find that retirees with higher 𝐴𝑅𝐼𝑆 𝑖 are
more likely to return to the labor force than those with low automation-resistance. Our two-stage
Heckman correction estimations show that, among retirees who return to work, those with higher
𝐴𝑅𝐼𝑆 𝑖 are more likely to choose jobs whose tasks are more resistant to automation (i.e., jobs with
higher 𝐴𝑅𝑂𝑇 𝑖𝑡
). The finding persists even after our controlling for the financial incentives
provided by employers who value employees’ 𝐴𝑅𝐼𝑆 𝑖 . This implies that the 𝐴𝑅𝐼𝑆 𝑖 -induced
unretirement job choices are driven at least partially by the voluntary will of the unretirees (i.e.,
the supply side) themselves rather than determined entirely by the demand side. Lastly, in
studying the financial and mental health returns to the skill-to-task matching in terms of
automation-resistance (i.e., the 𝐴𝑅𝐼𝑆 𝑖 -to-𝐴𝑅𝑂𝑇 𝑖𝑡
matching), we find positive financial returns to
matching if and only if we control for within-cohort variations in the valuation of automation-
86

resistance. In contrast, we find positive psychological returns to matching regardless of specific
cohort characteristics. For all analyses, the main findings undergo multiple sensitivity checks to
show that they are robust to different model specifications.
One of the main implications of the findings in this chapter is that financial need is not
the sole reason for unretirement. Retirees’ skill-specific automation resistance is shown to
enhance their unretirement likelihood even in the absence of financial reasons. One non-financial
return to unretirement could be an improvement in mental health. Via both the financial and non-
financial channels, retirees’ strong resistance to automation renders the notion of ‘returning to
work’ more attractive.
Another key implication of our findings is the value of soft skills in today’s society. Most
of us spend early stages of our lives honing hard skills through years of education and job
training. Likewise, our human capital is typically evaluated based on the extent of our hard skills.
Yet, the literature on task-based technology have begun to shed light on the fact that human
workers’ resistance to automation is determined not by hard skills but soft skills such as social
intelligence and creativity. This chapter makes a small contribution to this literature by providing
empirical evidence that workers’ soft skills have far-reaching impacts on the labor supply and
demand behaviors in the face of rapid technological changes. Our findings suggest that unretirees
make job choices based on the extent of match between the automation-resistance of their skills
and the task contents of the new job, and it is the soft skills/tasks that determine the level of
automation-resistance of the workers themselves and the occupations. These findings challenge
the traditional belief that individuals’ job choices are determined almost entirely by their hard
skills in math, engineering, science, reading, and writing. In highlighting the significant role of
the soft skills in strengthening the automation-resistance of individuals, we suggest that U.S.
87

education be revamped to help provide means and opportunities to hone individuals’ soft skills in
addition to the hard skills. In addition to considering how to develop these skills in secondary
and post-secondary education, scholars and policy-makers should discuss how to design training
programs for aging workers so as to improve their soft skills.
Finally, our findings underscore the intermediary role of cognitive abilities for aging
individuals in their making unretirement decisions. As shown in Section 2.4., aging individuals’
current level of cognitive abilities generates a synergy effect with their soft skills—which
determine the level of resistance to automation—in gauging their unretirement- likelihood and
job choices. Considering that individuals’ cognitive abilities rapidly decline with age, especially
among the retirees 65 or older, our findings suggest that unretirement incentives provided by
retirees’ automation-resistance may be offset by weakening cognitive abilities and deteriorating
decision-making abilities.
Among myriad aspects of individuals’ working lives, we have chosen to focus on
unretirement not only because of our incomplete understanding of the phenomenon itself in the
literature but also because of its policy implications. For instance, unretirement may be a boon to
policymakers faced with a shrinking tax base unable to provide pensions and health insurance to
a growing older population (Gruber and Wise, 1999; Munnell and Rutledge, 2013). Such
pressures are leading policymakers to promote longer working lives and thereby increase the
self-sustainability of workers (Maestas and Zissimopoulos, 2010).
68
Hence, our attempt in this
chapter to understand the motives behind unretirement and how the phenomenon unravels with
respect to the rising pressures of automation will help policymakers promote unretirement as a
new means for aging individuals to lengthen their working lives.

68
Past efforts on encouraging workers to extend their working lives include terminating the Social Security earnings
test and repealing mandatory retirement (Gustman et al., 2018).
88

Another benefit of voluntary unretirement (e.g., unretirement that is not tied to financial
needs) seems to lie in its ability to improve aging individuals’ mental health. Given the rapidly
increasing longevity in later years (Anderson and Hussey, 2000), retirees who stop working in
their 50s and 60s may find the inactivity and loneliness associated with the post-retirement years
to be exasperating. Gerontology research finds that late-life loneliness is directly related to
mental health problems, including worsening depressive symptoms and reduced self-esteem
(Blazer, 2002; Holmen, Ericsson, and Winblad, 1999; Reitzes et al., 1996). Faced with such
risks, unretirement serves a function other than improving financial security: working in old age
provides social stimulation and a sense of self-worth that are helpful in fighting depression
(Adams, Sanders, and Auth, 2004; van Solinge, 2007).
Given the limitation of currently-available data, we have explored the linkage between
work and automation with a proxy measurement: the extent of skill-specific and task-specific
automation resistance. We suggest that future research re-evaluate this topic by using data that
allows analysis of labor supply and demand behaviors in reaction to the variations across
industries, locations, or occupations in the levels of exposure to automation directly measured by
usage of robots or artificial intelligence.

89

Table 2.1. Elements of Automation-Resistance
Raises Automation-Resistance
Non-routine analytical adaptability (high-skilled)
Analyzing data/information
Thinking creatively
Interpreting information for others
Non-routine social intelligence (high-skilled)
Establishing and maintaining personal relationships
Guiding, directing and motivating subordinates
Coaching/developing others
Non-routine physical adaptability/dexterity (low-skilled)
Operating vehicles, mechanized devices, or equipment
Spend time using hands to handle, control or feel objects, tools or controls
Manual dexterity
Spatial orientation
Non-routine interpersonal adaptability (low-skilled)
Social Perceptiveness
Lowers Automation-Resistance
Routinization (high-skilled)
Importance of repeating the same tasks
Importance of being exact or accurate
Structured v. Unstructured work (reverse)
Routinization (low-skilled)
Pace determined by speed of equipment
Controlling machines and processes
Spend time making repetitive motions
Offshorability (both high- and low- skilled)
Face-to-Face Discussions
Assisting and Caring for Others
Performing for or Working Directly with the Public
Inspecting Equipment, Structures, or Material x 0.5
Handling and Moving Objects x 0.5
Repairing and Maintaining Mechanical Equipment x 0.5
Repairing and Maintaining Electronic Equipment x 0.5
Source: O*NET
Notes: For weights, O*NET's 'levels' are used. As for the offshorability measure, I have
taken the average of onsite- and face-to-face work without overlapping items, and multiplied
equipment repair categories by 0.5. All criteria closely emulate that of the index proposed by
Acemoglu and Autor (2011).
90

Figure 2.1. Evolution of the AROT Index Scores by Occupational Groups

Source: Author’s calculations based on the DOT, O*NET, and the HRS.

91

Figure 2.2. Evolution of the ARIS Index Score by Gender

Source: Author’s calculations based on the DOT, O*NET, and the HRS.

Figure 2.3. Evolution of the ARIS Index Score by Education

Source: Author’s calculations based on the DOT, O*NET, and the HRS.

92

Table 2.2. Summary Statistics
cohort 1 cohort 2 cohort 3 ANOVA t-test
mean sd mean mean mean cohort gender
unretired (1 yes 0 no) 0.360 0.498 0.132 0.192 0.660 *** ***
ARIS index, for workers 0.008 1.478 0.022 -0.031 0.146 *** ***
AROT index, for jobs -0.021 0.762 0.056 -0.082 0.103 ***
age 64.856 7.848 75.023 67.785 58.71 *** ***
cohort 2.229 0.666 1 2 3 ***
male 0.412 0.500 0.352 0.425 0.528 ***
race 1.252 0.515 1.126 1.145 1.245 *** ***
educ, highest degree obtained 1.304 0.673 1.259 1.282 1.443 *** ***
marital status 2.218 0.512 2.366 2.236 2.164 *** ***
household size 2.275 1.168 1.866 2.122 2.425 *** ***
access to employer-sponsored pension 0.436 0.500 0.044 0.468 0.567 ** ***
access to Social Security benefits 0.617 0.500 0.997 0.815 0.278 *** ***
ln(earnings), indexed to 2012 dollars -2.291 9.610 -6.721 -4.668 2.630 ** ***
ln(other HH income), indexed to 2012 dollars 9.364 4.220 9.867 9.876 8.725 *** ***
wealth (in 00,000), indexed to 2012 dollars 1.742 1.555 1.554 1.918 1.895 ** ***
access to Medicare 0.541 0.496 0.984 0.708 0.203 *** ***
health conditions, standardized 0.693 1.019 0.722 0.792 0.489 *** ***
challenges to ADLs, std 0.073 0.941 0.036 0.065 0.044 *** ***
state 24.200 15.294 24.530 25.086 24.605 * *
wave 7.791 2.829 5.400 7.299 9.144 *** **
work-related personality: organized 3.105 0.787 3.099 3.141 3.093 ** ***
work-related personality: responsible 3.717 0.473 3.650 3.725 3.749 *** ***
work-related personality: hardworking 3.460 0.691 3.256 3.407 3.598 *** ***
work-related personality: intelligent 3.299 0.627 3.235 3.264 3.369 *** ***
work-related personality: thorough 3.121 0.712 3.092 3.140 3.166 *** ***
post-retirement cognitive abilities (total), std 0.030 0.540 0.346 0.118 -0.082 *** ***
CES-D score 1.414 0.937 1.352 1.375 1.494 *** ***
diagnosed depression (1 yes 0 no) 0.171 0.376 0.108 0.159 0.186 *** ***
N 115432 115432 14310 60390 40732
total
Notes: ARIS refers to the automation-resistance of individual skills index, for each worker. AROT refers to the automation-resistance of
occupational tasks, for each job. The oldest cohort (Cohort 1) consists of individuals born before 1930. The middle cohort (Cohort 2) is made up
of individuals born between 1931 and 1941. Lastly, the youngest (Cohort 3) include individuals born between 1942 and 1959. *p<0.05
Source: Author's calculations from the HRS.
93

Table 2.3. Unretirement Across the Analytic Period
Year Stay Retired Unretired
1992 834 561
1994 3284 1142
1996 3763 1438
1998 5522 3168
2000 6113 3100
2002 6671 3024
2004 7348 4686
2006 7889 4277
2008 7893 4054
2010 8308 6149
2012 8396 5333
2014 7936 4543
Source: Author's calculations using the HRS.
Table 2.4. Unretirement Decisions by ARIS, Linear Probability Model
(1) (2) (3) (4) (5) (6)
outcome = Unretired (1/0)
ARIS index i 0.007** 0.007*** 0.006*** -0.002 -0.018 -
(0.003) (0.003) (0.002) (0.003) (0.016) -
cohort (ref= cohort 1, born before 1930)
1(cohort 2, born in 1931-1941) - - - -0.038*** -0.040*** -
- - - (0.006) (0.006) -
1(cohort 3, born in 1942-1959) - - - 0.249*** 0.247*** -
- - - (0.008) (0.008) -
(ARIS index i ) x 1(cohort 2 i ) - - - - 0.011 -
- - - - (0.016) -
(ARIS index i ) x 1(cohort 3 i ) - - - - 0.024 -
- - - - (0.016) -
(cognitive abilities it ) - - - - - 0.013***
- - - - - (0.003)
(ARIS index i ) x (cognitive abilities it ) - - - - - 0.006*
- - - - - (0.003)
baseline covariates Y Y Y Y Y Y
work-related personality controls N Y Y Y Y N
income and wealth controls N N Y N N N
cohort fixed-effects N N N Y Y N
post-retirement cognitive abilities N N N N N Y
individual fixed-effects N N N N N Y
dependent variable mean 0.359 0.359 0.359 0.359 0.359 0.359
observations 115432 115432 115432 115432 115432 115432
Note: ARIS refers to the automation-resistance of individual skills index, for each worker. The oldest cohort (Cohort 1) consists of
individuals born before 1930. The middle cohort (Cohort 2) is made up of individuals born between 1931 and 1941. Lastly, the youngest
(Cohort 3) include individuals born between 1942 and 1959. Year- and state- fixed effects are included in the regression equations but not
reported in the table. Standard errors are in parentheses. If included in the specification, items are marked with a "Y". If not, items are
marked with a "N". For column 4, the reference category is the earliest cohort--those born before 1930. For column 5, the reference
category is (ARIS index)x(cohort 1). *p<0.1 **p<0.05 ***p<0.01
Source: Author's calculations using the HRS.
94

Table 2.5. Unretirement Decisions by ARIS, Logit Model
outcome = Unretired (1/0) logit margins logit margins logit margins logit margins logit margins logit margins
ARIS index i 0.067* 0.006** 0.070** 0.006** 0.063** 0.005** -0.015 -0.001 -0.296 -0.003 - -
(0.034) (0.002) (0.034) (0.003) (0.028) (0.002) (0.033) (0.003) (0.186) (0.003) - -
cohort (ref= cohort 1, born before 1930)
1(cohort 2, born in 1931-1941) - - - - - - -0.205** -0.017** -0.225***-0.019*** - -
- - - - - - (0.084) (0.007) (0.085) (0.007) - -
1(cohort 3, born in 1942-1959) - - - - - - 3.043***0.305*** 3.023*** 0.303*** - -
- - - - - - (0.110) (0.010) (0.110) (0.010) - -
(ARIS index i ) x 1(cohort 2 i ) - - - - - - - - 0.232 -0.005 - -
- - - - - - - - (0.192) (0.004) - -
(ARIS index i ) x 1(cohort 3 i ) - - - - - - - - 0.356* 0.006 - -
- - - - - - - - (0.191) (0.004) - -
(cognitive abilities it ) - - - - - - - - - - 0.351***0.029***
- - - - - - - - - - (0.035) (0.003)
(ARIS index i ) x (cognitive abilities it ) - - - - - - - - - - 0.111***0.014***
- - - - - - - - - - (0.036) (0.004)
baseline covariates Y Y Y Y Y Y Y Y Y Y Y Y
work-related personality controls N N Y Y Y Y Y Y Y Y N N
income and wealth controls N N N N Y Y N N N N N N
cohort fixed-effects N N N N N N Y Y Y Y N N
post-retirement cognitive abilities N N N N N N N N N N Y Y
individual fixed-effects N N N N N N N N N N Y Y
dependent variable mean 0.400 0.400 0.400 0.400 0.400 0.400
observations 115432 115432 115432 115432 115432 115432
(4) (5) (6)
Note: ARIS refers to the automation-resistance of individual skills index, for each worker. The oldest cohort (Cohort 1) consists of individuals born before 1930. The middle cohort
(Cohort 2) is made up of individuals born between 1931 and 1941. Lastly, the youngest (Cohort 3) include individuals born between 1942 and 1959. Year- and state- fixed effects
are included in the regression equations but not reported in the table. Standard errors are in parentheses. If included in the specification, items are marked with a "Y". If not, items
are marked with a "N". For column 4, the reference category is the earliest cohort--those born before 1930. For column 5, the reference category is the interaction term (ARIS
index) x (cohort 1). *p<0.1 **p<0.05 ***p<0.01
Source: Author's calculations using the HRS.
(1) (2) (3)
95

Table 2.6. Unretirement Job Choices by ARIS: Task-Specific Automation Resistance of Unretirees' Jobs
(1) (2) (3) (4) (5) (6)
outcome = AROT index it
ARIS index i 0.772*** 0.763*** 0.755*** 0.761*** 1.118*** 0.759***
(0.020) (0.020) (0.020) (0.020) (0.113) (0.020)
cohort (ref= cohort 1, born before 1930)
1(cohort 2, born in 1931-1941) - - - -0.109 -0.101 -
- - - (0.078) (0.076) -
1(cohort 3, born in 1942-1959) - - - -0.021 -0.010 -
- - - (0.095) (0.094) -
(ARIS index i ) x 1(cohort 2 i ) - - - - -0.413*** -
- - - - (0.118) -
(ARIS index i ) x 1(cohort 3 i ) - - - - -0.345*** -
- - - - (0.114) -
(cognitive abilities it ) - - - - - 0.063**
- - - - - (0.027)
(ARIS index i ) x (cognitive abilities it ) - - - - - 0.080***
- - - - - (0.026)
baseline covariates? Y Y Y Y Y Y
work-related personality controls? N Y Y Y Y N
income and wealth controls? N N Y N N N
cohort fixed-effects? N N N Y Y N
post-retirement cognitive abilities? N N N N N Y
individual fixed-effects? N N N N N N
dependent variable mean 0.008 0.008 0.008 0.008 0.008 0.008
observations 115432 115432 115432 115432 115432 115432
Note: ARIS refers to the automation-resistance of individual skills index, for each worker. AROT refers to the automation-resistance of
occupational tasks, for each job. The oldest cohort (Cohort 1) consists of individuals born before 1930. The middle cohort (Cohort 2) is
made up of individuals born between 1931 and 1941. Lastly, the youngest (Cohort 3) include individuals born between 1942 and 1959.
Only the second stage (outcome equation estimation) is shown. Year- and state- fixed effects are included in the regression equations but
not reported in the table. Standard errors are in parentheses. If included in the specification, items are marked with a "Y". If not, items are
marked with a "N". For column 4, the reference category is the earliest cohort--those born before 1930. For column 5, the reference
category is (ARIS index)x(cohort 1). *p<0.1 **p<0.05 ***p<0.01
Source: Author's calculations using the HRS.
96

Table 2.7. Unretirement Job Choices by ARIS: White-Collar Status, Industry, and STEM jobs
(1) (2) (3)
white-collar same industry STEM
ARIS index i 0.042*** -0.011* 0.029***
(0.005) (0.006) (0.003)
baseline covariates? Y Y Y
work-related personality controls? Y Y Y
income and wealth controls? N N N
individual fixed-effects? N N N
dependent variable mean 0.068 0.436 0.039
observations 115432 115432 115432
Note: ARIS refers to the automation-resistance of individual skills index, for each worker. Outcomes of interest for columns 1, 2, and
3 are whether or not the job is white-collar, in the same industry as the worker's career job, and in STEM, respectively. Only the
second stage results are shown. Year- and state- fixed effects are included in the regression equations but not reported in the table.
Standard errors are in parentheses. If included in the specification, items are marked with a "Y". If not, items are marked with a
"N". *p<0.1 **p<0.05 ***p<0.01
Source: Author's calculations using the HRS.
Table 2.8. Monetary and Mental Health Returns to Skill-to-Task Matching
(1) (2) (3) (4) (5) (6) (7) (8) (9)
Y
ARIS index i , for individuals 0.359*** 0.349*** -0.003 0.003 0.002 -0.018 -0.002* -0.001 -0.002
(0.040) (0.039) (0.254) (0.008) (0.008) (0.050) (0.001) (0.001) (0.004)
AROT index it , for occupations -0.041 -0.059 1.400** -0.003 -0.005** 0.005 0.002 0.003 -0.006
(0.089) (0.088) (0.583) (0.002) (0.003 (0.010 (0.004) (0.004) (0.022)
(ARIS i ) x (AROT it ) -0.651*** -0.618*** -1.872*** -0.005* -0.007*** -0.018* -0.001 -0.003*** -0.008*
(0.039) (0.038) (0.200) (0.003) (0.003) (0.010) (0.001) (0.001) (0.005)
cohort (ref= cohort 1, born before 1930)
1(cohort 2, born in 1931-1941) - - -2.919*** - 0.062** 0.059** - 0.058*** 0.055***
- - (0.283) - (0.027) (0.028) - (0.013) (0.013)
1(cohort 3, born in 1942-1959) - - -11.003*** - 0.153*** 0.148*** - 0.084*** 0.081***
- - (0.294) - (0.031) (0.032) - (0.014) (0.014)
(AROT index it ) x 1(cohort 2 i ) - - 0.618** - - -0.008 - - 0.003
- - (0.263) - - (0.010) - - (0.004)
(AROT index it ) x 1(cohort 3 i ) - - 0.239 - - -0.013 - - -0.001
- - (0.256) - - (0.010) - - (0.004)
(ARIS index i ) x 1(cohort 2 i ) - - -1.402** - - 0.028 - - -0.001
- - (0.600) - - (0.052) - - (0.023)
(ARIS index i ) x 1(cohort 3 i ) - - -1.539*** - - 0.018 - - 0.015
- - (0.592) - - (0.051) - - (0.022)
(AROT index it ) x (ARIS index i) x 1(cohort 2 i ) - - 0.826*** - - 0.010 - - 0.004
- - (0.212) - - (0.010) - - (0.005)
(AROT index it ) x (ARIS index i) x 1(cohort 3 i ) - - 1.580*** - - 0.013 - - 0.005
- - (0.203) - - (0.010) - - (0.005)
baseline covariates? Y Y Y Y Y Y Y Y Y
work-related personality controls? N Y Y Y Y Y Y Y Y
income and wealth controls? N N N Y N N Y N N
cohort fixed-effects? N N Y N Y Y N Y Y
individual fixed-effects? N N N N N N N N N
dependent variable mean -2.291 -2.291 -2.291 -0.113 -0.113 -0.113 0.171 0.171 0.171
observations 115432 115432 115432 115432 115432 115432 115432 115432 115432
ln(income), RE CES-D, RE diag. depression, RE
Note: ARIS refers to the automation-resistance of individual skills index, for each worker. AROT refers to the automation-resistance of occupational tasks, for each job.
The oldest cohort (Cohort 1) consists of individuals born before 1930. The middle cohort (Cohort 2) is made up of individuals born between 1931 and 1941. Lastly, the
youngest (Cohort 3) include individuals born between 1942 and 1959. Only the second stage (outcome equation estimation) is shown. Year- and state- fixed effects are
included in the regression equations but not reported in the table. Standard errors are in parentheses. If included in the specification, items are marked with a "Y". If not,
items are marked with a "N". For columns 5 and 8, the reference category is the earliest cohort--those born before 1930. For columns 3, 6, and 9, the reference category is
(ARIS index)x(AROT index)x(cohort 1). *p<0.1 **p<0.05 ***p<0.01
Source: Author's calculations using the HRS.
97

2.7 Appendix A: Theory of Skill-Biased Technological Changes (SBTC)

In understanding the nature of the automation, researchers often rely on the SBTC model
known as the theory of skill-biased technological change (SBTC). The model is based on a
simplified world whereby workers are grouped into two—skilled and unskilled—based on their
levels of educational attainment. The SBTC model states that the relationship between
technology and human beings is such that technology favors only the skilled, serving as a
complement for the skilled while substituting the unskilled. In the SBTC model, following the
notations of Card and DiNardo (2002) and Myck and Reed (2006), the aggregate labor demand is
generated by a constant elasticity of substitution production function,
Y = 𝐴 [𝛼 (𝑔 𝑆 𝑁 𝑆 )
(𝜎 −1)
𝜎 ⁄
+ (1 − 𝑎 )(𝑔 𝑈 𝑁 𝑈 )
(𝜎 −1)
𝜎 ⁄
]
𝜎 (𝜎 −1)
⁄
.
Again, workers fall into either the skilled or the unskilled groups. Y is the value of output, NS
and NU are the labor inputs (i.e., hours worked) of skilled and unskilled workers, respectively,
and 𝜎 is the elasticity of substitution between NS and NU such that 𝜎 ≥ 0. Lastly, 𝐴 , 𝑎 , 𝑔 𝑆 and
𝑔 𝑈 are time-variant technological parameters. A key feature of this model is the skill-biased
nature of technology that generates a wage differential between the skilled and the unskilled.
69

Going back to the production function in Eq. (1), the relative demand for skilled workers
is determined by making the ratio of the marginal product of the skilled and the unskilled equal

69
A simple technological growth that is not skill-biased would hold the same effect across all skill-levels, resulting
in a shift in the parameter α or a proportional shift in 𝑔 𝐻 and 𝑔 𝐿 and thus leaving the relative productivity of the
two skill groups unchanged. On the contrary, the SBTC differs from the notion of non-relative technological growth
in that the SBTC involves either an increase in α or an increase in 𝑔 𝑆 with respect to 𝑔 𝑈 .
98

to the ratio of their wages. The evolution of relative wages between the two groups can be
written by taking logarithms and first-differencing over time,

∆ log (
𝑤 𝑆 𝑤 𝑈 ) = ∆ log [
𝛼 (1−𝑎 )
] + [
𝜎 −1
𝜎 ∆ log (
𝑔 𝑆 𝑔 𝑈 )] − [(
1
𝜎 ) ∆ log (
𝑁 𝑆 𝑁 𝑈 )].

The above-equation shows that changes in the relative wage must stem from either the changes
in the relative supply of skilled workers or changes in technology.
70
As an individual decides on
how long to work and how much time to spend in leisure (e.g., retirement) so as to maximize her
utility, and in that the wage trajectory influences the value of work and leisure through income
and substitution effects, the SBTC-induced wage gap holds an impact on the labor supply
behaviors of the skilled and the unskilled (Blundell and MaCurdy, 1999; Myck and Reed, 2006).
Three shortcomings of the SBTC model merits attention. First, the technological growth
in the model is assumed to be factor-augmenting, thereby increasing wages for all workers—
albeit at a much higher rate for the skilled workers. This is widely at odds with the empirical data
on the returns to educational premium shown in the literature (Katz and Murphy, 1992;
Acemoglu and Autor, 2011). In plotting the evolution of real log earnings by the skill levels
between 1964 to 2009, Acemoglu and Autor (2011) notice that the real wages rose for the highly
educated (i.e. ‘skilled) workers while falling steeply for the less educated workers in the 1980s.
Such a wage inequality across the skill groups undergoes a continuous widening until 2009,
which marks the end of their analytic period. The fact that the major source of the inequality can

70
Regardless of the estimated value of σ, an increase in the relative proportion of the skilled workers would
corresponds to a decrease in their relative wage holding constant the technology. If the relative supply of skilled-
workers were to be held constant, then the technological growths will generate a wage differential—with the skilled
workers’ relative wage rising at a higher rate than that of the unskilled. Other features that influence relative wages
such as premiums for efficiency or productivity have been ignored in the model.
99

be attributed more to the rapidly falling wage of the unskilled and less to the increasing wage of
the skilled undermines the factor-augmenting technology portrayed by the SBTC model.
71

Second, the SBTC model falls short of explaining the job polarization, an empirically
observed phenomenon starting in the 1990s. While the SBTC model predicts a monotonic
relationship between technology-induced automation and skill levels (i.e., the more skilled an
individual is, the more complemented by the process of automation she is), scholars assessing the
relationship between skill levels and change in employment as a share of total US employment
for each decade observe a pattern of job polarization in the 1990s: The employment share grows
rapidly both at the highest and lowest percentiles (i.e., the low-skilled and the high-skilled
workers) while falling at the intermediate percentiles (i.e., the mid-skilled) (Acemoglu, 1999).
72

Acemoglu and Autor (2011) note that, in 2000s, the employment share grows steeply only at the
lowest percentiles while stagnating at the highest percentiles—with the share of those in the
middle still falling. Linking the trend of polarization to automation, it could be said that
technology is substituting the mid-skilled workers while complementing the low-skilled and the
high-skilled ones. Such is a phenomenon cannot be explained by the SBTC model.
Another shortcoming of the SBTC framework lies in its having no mention of
occupational tasks. In doing so, the model implicitly assumes that there is a one-to-one matching
between a worker’s skills and tasks she carries out in her job. Nevertheless, an occupation
consists of multiple tasks—some of which fall within the job holder’s pool of skills.
Furthermore, it is jobs and not workers that technology automates: The one-to-one matching
between workers’ skills and occupational tasks carried out by the workers is not always

71
Here, the highly-educated refer to those with post-college degrees while the less-educated refer to those without
four-year college degrees.
72
Here, we still equate the level of skills with the years of educational attainment.
100

guaranteed, and the assignment of skills to tasks is determined by the nature of the technological
changes, labor demands (i.e., task demands to be more accurate), and labor supplies (Acemoglu
and Autor, 2011; Acemoglu and Zilibotti, 2001).

2.8 Appendix B: Theory of Task-Based Technological Changes (TBTC)

The newer model, known as the model of task-based technological change (TBTC),
clearly distinguishes between skills (i.e., stock of capabilities) and tasks (i.e., the unit of activity
performed at each job that produces output). The TBTC model assumes that an occupation
consists of a continuum of tasks—all of which are needed to produce a final good off the job.
73

Workers use skills to accomplish occupational tasks and to receive salaries in return. Workers
holding the same occupation can have different comparative advantages over different tasks,
which often require them to collaborate to produce the final good. The remainder of this section
summarizes the model of TBTC, following the notations of Acemoglu and Autor (2011).
Allowing no trade in tasks and assuming a closed economy, the production function of a
unique final good—generated by combining a range of occupational tasks falling within [0,1]—
is written as

𝑌 = 𝑒𝑥𝑝 [∫ 𝑙𝑛 𝑦 (𝑖 )𝑑𝑖 1
0
] ,

where 𝑦 (𝑖 ) is the production level of task 𝑖 . To better encapsulate the job polarization empirically
observed in the past decades, the model of TBTC is written with respect to individuals with three

73
The TBTC model accounts for the fact that recent surge in offshoring and outsourcing of labor influences the
labor market. In the end, framework sets it so that the occupational tasks can be outsourced (i.e., produced outside
the US), automated, or upheld by workers (Autor et al. 2003).
101

skill levels: low-skilled, mid-skilled, and high-skilled. The supply of each skill level (i.e., 𝐿 , 𝑀
and 𝐻 ) is inelastic. As workers possess different comparative advantages across tasks based on
their skill levels, the production function of each task is written as,

𝑌 = 𝐴 𝐿 𝛼 𝐿 (𝑖 )𝑙 (𝑖 ) + 𝐴 𝑀 𝛼 𝑀 (𝑖 )𝑚 (𝑖 ) + 𝐴 𝐻 𝛼 𝐻 (𝑖 )ℎ(𝑖 ) + 𝐴 𝐾 𝛼 𝐾 (𝑖 )𝑘 (𝑖 ) ,

where A is the factor-augmenting technology (i.e., 𝐴 𝐻 is the technology favoring the high-skilled
workers), and 𝛼 𝐿 (𝑖 ), 𝛼 𝑀 (𝑖 ), and 𝛼 𝐻 (𝑖 ) are the task-specific productivity parameters of each skill
group in carrying out different tasks 𝑖 .
74
Lastly, 𝑙 (𝑖 ), 𝑚 (𝑖 ), and 𝑞 (𝑖 ) are the number of low-
skilled, mid-skilled, and high-skilled workers carrying out task 𝑖 . Among the tasks, some tasks
(𝑇 𝐿 ) will be best performed by the low-skilled. The same logic applies to 𝑇 𝑀 for the mid-skilled
and 𝑇 𝐻 for the high-skilled.
75
Naturally, tasks will be divided among workers of three distinct
skill levels. The extent of substitutability of skills across the occupational tasks depend on (1) the
quantity of labor supplied for each skill level and (2) the nature of technological changes.
Within the TBTC framework, the demand side (i.e. employers) choose how to optimally
allocate employees’ skills to occupational tasks.
76
An equilibrium is reached when the demands-
side maximize profits and the labor markets clear. The TBTC framework allows for imperfect
skill-to-task matching, thereby allowing the same tasks could be allotted to workers in different
skill groups. The extent of matching is decided jointly by technological growths as well as hiring

74
The comparative advantages have the following structure: (1) As for 𝛼 𝐿 (𝑖 )/𝛼 𝑀 (𝑖 ), and 𝛼 𝑀 (𝑖 )/𝛼 𝐻 (𝑖 ). Higher
indices suggest more complicated tasks whereby the skill group whose productivity schedules are on the numerator
perform better than those in the denominator.
75
Employers hiring workers from the same skill group (e.g., the mid-skilled group) must pay them the same wage in
an equilibrium (i.e., the value of the marginal product of individuals in the same skill group will be the same in all
tasks they carry out).
76
The model of TBTC does not include labor supply behaviors as potential determinants of allocations of skills to
tasks. In other words, the TBTC model lets the optimal allocation of skill-to-task to be done by the employers alone.
102

decisions of the demand side. Technological growths can change the nature of jobs to be carried
out by human workers by automating certain tasks within the jobs’ task contents. Additionally,
the biased nature of technology dictates that workers with varying skill levels have different level
of productivity in carrying out certain tasks.
It is the quality of the matches and the extent of delineations among the threshold tasks
(𝐼 𝐿 and 𝐼 𝐻 ) that determine the differences in earnings across the skill groups and influence the
subsequent increase or reduction in earnings.

Intuitively, the threshold tasks 𝐼 𝐿 are a set of tasks
that can either be carried out by low-skilled or mid-skilled workers; 𝐼 𝐻 can be understood in a
similar manner. In that wages are simply the value of the marginal product of different skill
levels, the wage of a low-skilled worker (𝑤 𝐿 ) is equivalent to 𝑃 𝐿 𝐴 𝐿 , the product of price index of
tasks performed by low-skilled workers and technology favoring the low-skilled. The relative
wage between the mid- and high- skilled workers is

w
H
𝑤 𝑀 =
𝑃 𝐻 𝐴 𝐻 𝑃 𝑀 𝐴 𝑀 = (
1−𝐼 𝐻 𝐼 𝐻 −𝐼 𝐿 ) (
𝐻 𝑀 )
−1
.

Here, 𝐻 is defined by the factor market clearing assumption: ∫ ℎ(𝑖 )𝑑𝑖 ≤ 𝐻 1
0
. 𝑀 and L are
defined analogously. The last term expresses the relative wages in terms of the relative supplies
of workers in each skill-level to the entire range of tasks falling within [0,1] and threshold tasks.
In a similar manner, the relative wages of a mid- and a low-skilled worker is,

w
M
𝑤 𝐿 =
𝑃 𝑀 𝐴 𝑀 𝑃 𝐿 𝐴 𝐿 = (
1−𝐼 𝐻 𝐼 𝐻 −𝐼 𝐿 ) (
𝐻 𝑀 )
−1
.

103

If we were to assume that there exist only two skill levels—high-skilled and low-skilled—as well
as a perfect matching between skills and tasks (i.e., low-skilled workers carrying out lower-level
tasks, which do not overlap with the tasks carried out by high-skilled individuals), then the
TBTC model reverts back to the SBTC model.
77

The equilibrium in the TBTC equalizes the cost of carrying out the threshold tasks 𝐼 𝐿 by
low-skilled and mid-skilled individuals, and that the cost of handling tasks 𝐼 𝐻 by mid-skilled and
high-skilled are the same. Such conditions known as the ‘no arbitrage’ conditions. The no-
arbitrage condition between the low-skilled and the mid-skilled is written as

𝐴 𝐿 𝛼 𝐿 (𝐼 𝐿 )𝐿 𝐼 𝐿 =
𝐴 𝑀 𝛼 𝑀 (𝐼 𝐿 )𝑀 𝐼 𝐻 −𝐼 𝐿 .

Likewise, the condition between the mid-skilled and the high-skilled is

𝐴 𝑚 𝛼 𝑀 (𝐼 𝐻 )𝑀 𝐼 𝐻 − 𝐼 𝐿 =
𝐴 𝐻 𝛼 𝐻 (𝐼 𝐻 )𝐻 1−𝐼 𝐻 .

By taking the logs of the two equations above, we can derive the following two equations
whereby both curves are upward sloping in the (𝐼 𝐻 , 𝐼 𝐿 ) space:

𝑙𝑛 𝐴 𝑀 − 𝑙𝑛 𝐴 𝐻 + (𝑙𝑛 𝛼 𝑀 (𝐼 ) − 𝑙𝑛 𝛼 𝐻 (𝐼 ))(𝐼 𝐻 ) + 𝑙𝑛𝑀 − 𝑙𝑛𝐻 − 𝑙𝑛 (𝐼 𝐻 − 𝐼 𝐿 ) + 𝑙𝑛 (1 − 𝐼 𝐻 ) = 0

77
But the reverted SBTC model now has a Cobb-Douglas production function whereby the elasticity of substitution
between the high-skilled and the low-skilled is equal to one. By changing the functional form of the comparative
advantage of the high- and the low-skilled workers, we can obtain any constant returns to scale production function
(i.e., elasticity of substitution >=1)—which is exactly in line with the set-up of the SBTC model (Acemoglu and
Autor, 2011; Acemoglu and Zilibotti, 2001).
104

and

𝑙𝑛 𝐴 𝐿 − 𝑙𝑛 𝐴 𝑀 + (𝑙𝑛 𝛼 𝐿 (𝐼 ) − 𝑙𝑛 𝛼 𝑀 (𝐼 ))(𝐼 𝐿 ) + 𝑙𝑛𝐿 − 𝑙𝑛𝑀 − 𝑙𝑛 (𝐼 𝐻 − 𝐼 𝐿 ) + 𝑙𝑛 (𝐼 𝐿 ) = 0

Between the two equations above, the first equation is steeper than the second, indicating an
existence of a unique intersection between the two curves that signifies the optimal allocation of
skills to occupational tasks. From these two equations, we can predict how the relative wages of
the low-, mid-, and high-skilled workers could change in response to the particular natures of the
technological changes.

105

2.9 Appendix C: Detailed Tables

Table 2.4A. Unretirement Decisions by Resistance to Automation Based on Skills, Linear Probability Models
(1) (2) (3) (4) (5) (6)
outcome = Unretired (1/0) FE
ARIS index i 0.007** 0.007*** 0.006*** -0.002 -0.018 -
(0.003) (0.003) (0.002) (0.003) (0.016) -
cohort (ref= cohort 1, born before 1930)
1(cohort 2, born in 1931-1941) - - - -0.038*** -0.040*** -
- - - (0.006) (0.006) -
1(cohort 3, born in 1942-1959) - - - 0.249*** 0.247*** -
- - - (0.008) (0.008) -
(ARIS index i ) x 1(cohort 2 i ) - - - - 0.011 -
- - - - (0.016) -
(ARIS index i ) x 1(cohort 3 i ) - - - - 0.024 -
- - - - (0.016) -
(cognitive abilities it ) - - - - - 0.013***
- - - - - (0.003)
(ARIS index i ) x (cognitive abilities it ) - - - - - 0.006*
- - - - - (0.003)
age -0.073*** -0.072*** -0.056*** - - -0.035***
(0.003) (0.003) (0.003) - - (0.004)
age-squared 0.000*** 0.000*** 0.000*** - - 0.000***
(0.000) (0.000) (0.000) - - (0.000)
male 0.067*** 0.068*** 0.047*** 0.045*** 0.045*** -
(0.004) (0.004) (0.003) (0.004) (0.004) -
race (ref=1.Caucasian)
2. African-American -0.009 -0.014** -0.025*** -0.016*** -0.016*** -
(0.006) (0.006) (0.005) (0.006) (0.006) -
3. Other 0.016* 0.013 0.012* 0.010 0.010 -
(0.009) (0.009) (0.007) (0.009) (0.009) -
marital status (ref=1.single)
2. married/partnered 0.014 0.006 0.028*** 0.010 0.010 -0.028
(0.010) (0.010) (0.008) (0.009) (0.009) (0.019)
3. separated/divorced/widowed 0.043*** 0.037*** 0.029*** 0.034*** 0.034*** -0.001
(0.010) (0.010) (0.008) (0.010) (0.010) (0.019)
education (1. highschool diploma, GED)
2. college degree 0.041*** 0.036*** 0.024*** 0.025*** 0.025*** -
(0.006) (0.006) (0.005) (0.006) (0.006) -
3. graduate degree 0.028*** 0.020** 0.013** 0.010 0.011 -
(0.008) (0.008) (0.006) (0.008) (0.008) -
access to Medicare -0.150*** -0.149*** -0.091*** -0.162*** -0.162*** -0.108***
(0.005) (0.005) (0.004) (0.004) (0.004) (0.005)
health conditions -0.046*** -0.043*** -0.032*** -0.044*** -0.044*** -0.024***
(0.002) (0.002) (0.001) (0.002) (0.002) (0.003)
challenges to ADLs, std -0.020*** -0.019*** -0.015*** -0.018*** -0.018*** -0.008***
(0.001) (0.001) (0.001) (0.001) (0.001) (0.001)
access to Social Security benefits -0.077*** -0.078*** -0.066*** -0.138*** -0.138*** -0.080***
(0.005) (0.005) (0.004) (0.004) (0.004) (0.005)
access to employer-pension 0.124*** 0.116*** 0.017*** 0.136*** 0.136*** 0.038***
(0.005) (0.004) (0.003) (0.004) (0.004) (0.012)
personality: organized - -0.021*** -0.014*** -0.022*** -0.022*** -
- (0.003) (0.002) (0.003) (0.003) -
personality: responsible - -0.009* -0.009*** -0.009* -0.008* -
- (0.005) (0.004) (0.005) (0.005) -
personality: hardworking - 0.078*** 0.050*** 0.082*** 0.082*** -
- (0.003) (0.002) (0.003) (0.003) -
personality: intelligent - 0.012*** 0.010*** 0.011*** 0.010*** -
- (0.004) (0.003) (0.003) (0.003) -
personality: thorough - -0.004 -0.001 -0.004 -0.004 -
- (0.003) (0.003) (0.003) (0.003) -
ln(income) - - 0.021*** - - -
- - (0.000) - - -
ln(other HH income) - - -0.004*** - - -
- - (0.000) - - -
totalHH wealth (Winsorized at 0.1) - - -0.011*** - - -
- - (0.001) - - -
constant 3.196*** 2.957*** 2.409*** 0.215*** 0.215*** 1.625***
(0.099) (0.101) (0.085) (0.021) (0.021) (0.198)
observations 115432 115432 115432 115432 115432 115432
dependent variable mean 0.359 0.359 0.359 0.359 0.359 0.359
Note: ARIS refers to the automation-resistance of individual skills index, for each worker. The oldest cohort (Cohort 1) consists of individuals born
before 1930. The middle cohort (Cohort 2) is made up of individuals born between 1931 and 1941. Lastly, the youngest (Cohort 3) include
individuals born between 1942 and 1959. Year- and state- fixed effects are included in the regression equations but not reported in the table.
Standard errors are in parentheses. For column 4, the reference category is the earliest cohort--those born before 1930. For column 5, the reference
category is (ARIS index)x(cohort 1). *p<0.1 **p<0.05 ***p<0.01
Source: Author's calculations using the HRS.
RE
106

Table 2.5A. Unretirement Decisions by ARIS, Logit Model
outcome = Unretired (1/0) logit margins logit margins logit margins logit margins logit margins logit margins
ARIS index i 0.067* 0.006** 0.070** 0.006** 0.063** 0.005** -0.015 -0.001 -0.296 -0.003 - -
(0.034) (0.002) (0.034) (0.003) (0.028) (0.002) (0.033) (0.003) (0.186) (0.003) - -
cohort (ref= cohort 1, born before 1930)
1(cohort 2, born in 1931-1941) - - - - - - -0.205** -0.017** -0.225***-0.019*** - -
- - - - - - (0.084) (0.007) (0.085) (0.007) - -
1(cohort 3, born in 1942-1959) - - - - - - 3.043*** 0.305*** 3.023*** 0.303*** - -
- - - - - - (0.110) (0.010) (0.110) (0.010) - -
(ARIS index i ) x 1(cohort 2 i ) - - - - - - - - 0.232 -0.005 - -
- - - - - - - - (0.192) (0.004) - -
(ARIS index i ) x 1(cohort 3 i ) - - - - - - - - 0.356* 0.006 - -
- - - - - - - - (0.191) (0.004) - -
(cognitive abilities it ) - - - - - - - - - - 0.351*** 0.029***
- - - - - - - - - - (0.035) (0.003)
(ARIS index i ) x (cognitive abilities it ) - - - - - - - - - - 0.111*** 0.014***
- - - - - - - - - - (0.036) (0.004)
age -0.562*** -0.046 -0.552*** -0.045 -0.464*** -0.037 - - - - -0.497*** -0.041
(0.038) N/A (0.038) N/A (0.034) N/A - - - - (0.039) N/A
age-squared 0.003*** 0.000 0.003*** 0.000 0.003*** 0.000 - - - - 0.002*** 0.000
(0.000) N/A (0.000) N/A (0.000) N/A - - - - (0.000) N/A
male 0.862*** 0.072 0.877*** 0.073 0.621*** 0.051 0.521*** 0.043 0.522*** 0.043 0.870*** 0.072
(0.051) N/A (0.051) N/A (0.042) N/A (0.049) N/A (0.049) N/A (0.051) N/A
race (ref=1.Caucasian)
2. African-American -0.031 -0.003 -0.080 -0.007 -0.229*** -0.018 -0.046 -0.004 -0.043 -0.003 -0.014 -0.001
(0.070) N/A (0.070) N/A (0.057) N/A (0.068) N/A (0.068) N/A (0.070) N/A
3. Other 0.198* 0.016 0.149 0.012 0.142* 0.012 0.193** 0.016 0.197** 0.016 0.189* 0.016
(0.102) N/A (0.100) N/A (0.083) N/A (0.097) N/A (0.097) N/A (0.100) N/A
marital status (ref=1.single)
2. married/partnered 0.173 0.014 0.064 0.005 0.311*** 0.025 0.068 0.005 0.069 0.006 0.072 0.006
(0.111) N/A (0.109) N/A (0.093) N/A (0.105) N/A (0.105) N/A (0.109) N/A
3. separated/divorced/widowed 0.495*** 0.041 0.419*** 0.034 0.349*** 0.028 0.308*** 0.025 0.309*** 0.025 0.425*** 0.035
(0.114) N/A (0.112) N/A (0.095) N/A (0.108) N/A (0.108) N/A (0.112) N/A
education (1. highschool diploma, GED)
2. college degree 0.456*** 0.038 0.410*** 0.034 0.296*** 0.024 0.243*** 0.020 0.244*** 0.020 0.338*** 0.028
(0.076) N/A (0.075) N/A (0.064) N/A (0.072) N/A (0.072) N/A (0.075) N/A
3. graduate degree 0.407*** 0.034 0.311*** 0.026 0.216*** 0.018 0.151* 0.012 0.158* 0.013 0.222** 0.018
(0.093) N/A (0.093) N/A (0.078) N/A (0.089) N/A (0.089) N/A (0.093) N/A
access to Medicare -1.026*** -0.091 -1.008*** -0.089 -0.682*** -0.059 -1.281*** -0.115 -1.280*** -0.115 -1.153*** -0.103
(0.044) N/A (0.044) N/A (0.043) N/A (0.043) N/A (0.043) N/A (0.047) N/A
health conditions -0.537*** -0.044 -0.501*** -0.041 -0.415*** -0.033 -0.519*** -0.042 -0.518*** -0.042 -0.493*** -0.041
(0.023) N/A (0.023) N/A (0.020) N/A (0.022) N/A (0.022) N/A (0.022) N/A
challenges to ADLs, std -0.392*** -0.032 -0.368*** -0.030 -0.305*** -0.025 -0.361*** -0.029 -0.360*** -0.029 -0.365*** -0.030
(0.023) N/A (0.023) N/A (0.021) N/A (0.022) N/A (0.022) N/A (0.023) N/A
access to Social Security benefits -0.576*** -0.050 -0.580*** -0.050 -0.540*** -0.046 -1.177*** -0.105 -1.177*** -0.105 -0.630*** -0.055
(0.046) N/A (0.046) N/A (0.047) N/A (0.043) N/A (0.043) N/A (0.047) N/A
access to employer-pension 1.449*** 0.124 1.355*** 0.116 0.317*** 0.026 1.558*** 0.133 1.557*** 0.133 1.353*** 0.115
(0.050) N/A (0.050) N/A (0.042) N/A (0.050) N/A (0.050) N/A (0.050) N/A
personality: organized - - -0.280*** -0.023 -0.204*** -0.016 -0.286*** -0.023 -0.285*** -0.023 -0.279*** -0.023
- - (0.037) N/A (0.030) N/A (0.035) N/A (0.035) N/A (0.037) N/A
personality: responsible - - -0.111* -0.009 -0.117** -0.009 -0.137** -0.011 -0.136** -0.011 -0.120** -0.010
- - (0.060) N/A (0.048) N/A (0.056) N/A (0.056) N/A (0.060) N/A
personality: hardworking - - 1.011*** 0.083 0.696*** 0.056 1.033*** 0.084 1.033*** 0.084 1.015*** 0.084
- - (0.045) N/A (0.036) N/A (0.043) N/A (0.043) N/A (0.045) N/A
personality: intelligent - - 0.122*** 0.010 0.119*** 0.010 0.113*** 0.009 0.112*** 0.009 0.103** 0.008
- - (0.044) N/A (0.036) N/A (0.042) N/A (0.042) N/A (0.044) N/A
personality: thorough - - -0.045 -0.004 -0.011 -0.001 -0.039 -0.003 -0.038 -0.003 -0.053 -0.004
- - (0.041) N/A (0.033) N/A (0.039) N/A (0.039) N/A (0.041) N/A
ln(income) - - - - 0.158*** 0.013 - - - - - -
- - - - (0.002) N/A - - - - - -
ln(other HH income) - - - - -0.037*** -0.003 - - - - - -
- - - - (0.004) N/A - - - - - -
totalHH wealth (Winsorized at 0.1) - - - - -0.155*** -0.013 - - - - - -
- - - - (0.013) N/A - - - - - -
constant 22.696*** - 19.861*** - 16.424*** - -3.386*** - -3.374*** - 17.948*** -
(1.244) - (1.252) - (1.121) - (0.259) - (0.259) - (1.272) -
observations 115420 115420 115420 115420 115420 115420 115420 115420 115420 115420 115420 115420
dependent variable mean 0.400 0.400 0.400 0.400 0.400 0.400 0.400 0.400 0.400 0.400 0.400 0.400
Note: ARIS refers to the automation-resistance of individual skills index, for each worker. The oldest cohort (Cohort 1) consists of individuals born before 1930. The middle cohort (Cohort 2) is
made up of individuals born between 1931 and 1941. Lastly, the youngest (Cohort 3) include individuals born between 1942 and 1959. Year- and state- fixed effects are included in the
regression equations but not reported in the table. Standard errors are in parentheses. For column 4, the reference category is the earliest cohort--those born before 1930. For column 5, the
reference category is the interaction term (ARIS index) x (cohort 1). *p<0.1 **p<0.05 ***p<0.01
Source: Author's calculations using the HRS.
(6) (1) (2) (3) (4) (5)
107

Table 2.6A. Unretirement Job Choices by ARIS: Task-Specific Automation Resistance of Unretirees' Jobs
(1) (2) (3) (4) (5) (6)
outcome = AROT index it
ARIS index i 0.772*** 0.763*** 0.755*** 0.761*** 1.118*** 0.759***
(0.020) (0.020) (0.020) (0.020) (0.113) (0.020)
cohort (ref= cohort 1, born before 1930)
1(cohort 2, born in 1931-1941) - - - -0.109 -0.101 -
- - - (0.078) (0.076) -
1(cohort 3, born in 1942-1959) - - - -0.021 -0.010 -
- - - (0.095) (0.094) -
(ARIS index i ) x 1(cohort 2 i ) - - - - -0.413*** -
- - - - (0.118) -
(ARIS index i ) x 1(cohort 3 i ) - - - - -0.345*** -
- - - - (0.114) -
(cognitive abilities it ) - - - - - 0.063**
- - - - - (0.027)
(ARIS index i ) x (cognitive abilities it ) - - - - - 0.080***
- - - - - (0.026)
age -0.035 -0.035 -0.032 - - -0.022
(0.029) (0.029) (0.029) - - (0.030)
age-squared 0.000 0.000 0.000 - - 0.000
(0.000) (0.000) (0.000) - - (0.000)
male -0.202*** -0.170*** -0.123*** -0.157*** -0.155*** -0.171***
(0.030) (0.030) (0.030) (0.029) (0.029) (0.030)
race (ref=1.Caucasian)
2. African-American -0.075* -0.103*** -0.043 -0.106*** -0.106*** -0.092**
(0.039) (0.039) (0.040) (0.039) (0.039) (0.040)
3. Other -0.191*** -0.170*** -0.109** -0.175*** -0.172*** -0.162***
(0.049) (0.049) (0.049) (0.048) (0.049) (0.049)
marital status (ref=1.single)
2. married/partnered 0.107* 0.103 -0.017 0.112* 0.109* 0.105*
(0.063) (0.063) (0.066) (0.063) (0.063) (0.062)
3. separated/divorced/widowed 0.005 -0.006 0.001 0.014 0.011 -0.007
(0.066) (0.066) (0.066) (0.066) (0.066) (0.066)
education (1. highschool diploma, GED)
2. college degree 0.711*** 0.660*** 0.605*** 0.661*** 0.661*** 0.652***
(0.042) (0.042) (0.043) (0.042) (0.042) (0.042)
3. graduate degree 1.197*** 1.123*** 1.051*** 1.123*** 1.119*** 1.117***
(0.056) (0.056) (0.058) (0.056) (0.056) (0.057)
health conditions 0.007 0.007 -0.008 0.004 0.004 0.009
(0.017) (0.016) (0.016) (0.016) (0.016) (0.016)
challenges to ADLs, std 0.032* 0.033* 0.015 0.020 0.021 0.036*
(0.019) (0.019) (0.018) (0.018) (0.018) (0.018)
personality: organized - 0.008 -0.004 0.005 0.006 0.008
- (0.021) (0.021) (0.021) (0.021) (0.021)
personality: responsible - 0.068** 0.053 0.067** 0.068** 0.066**
- (0.033) (0.033) (0.033) (0.033) (0.033)
personality: hardworking - -0.087*** -0.039 -0.072** -0.072** -0.087***
- (0.030) (0.029) (0.029) (0.029) (0.030)
personality: intelligent - 0.152*** 0.155*** 0.153*** 0.150*** 0.148***
- (0.025) (0.025) (0.025) (0.025) (0.025)
personality: thorough - 0.076*** 0.059** 0.074*** 0.075*** 0.074***
- (0.023) (0.023) (0.023) (0.023) (0.023)
ln(income) - - 0.011*** - - -
- - (0.004) - - -
ln(other HH income) - - 0.010*** - - -
- - (0.002) - - -
totalHH wealth (Winsorized at 0.1) - - 0.075*** - - -
- - (0.010) - - -
constant 1.064 0.388 0.455 -0.575*** -0.590*** -0.012
(0.890) (0.898) (0.894) (0.181) (0.180) (0.927)
dependent variable mean 0.008 0.008 0.008 0.008 0.008 0.008
observations 115432 115432 115432 115432 115432 115432
Note: ARIS refers to the automation-resistance of individual skills index, for each worker. AROT refers to the automation-resistance of occupational
tasks, for each job. The oldest cohort (Cohort 1) consists of individuals born before 1930. The middle cohort (Cohort 2) is made up of individuals born
between 1931 and 1941. Lastly, the youngest (Cohort 3) include individuals born between 1942 and 1959. Only the second stage (outcome equation
estimation) is shown. Year- and state- fixed effects are included in the regression equations but not reported in the table. Standard errors are in
parentheses. For column 4, the reference category is the earliest cohort--those born before 1930. For column 5, the reference category is (ARIS
index)x(cohort 1). *p<0.1 **p<0.05 ***p<0.01
Source: Author's calculations using the HRS.
108

Table 2.7A. Unretirement Job Choices by ARIS: White-Collar Status, Industry, and STEM jobs
(1) (2) (3)
white-collar same industry STEM
ARIS index i 0.042*** -0.011* 0.029***
(0.005) (0.006) (0.003)
age 0.006 0.083*** -0.005
(0.005) (0.008) (0.003)
age-squared -0.000 -0.001*** 0.000*
(0.000) (0.000) (0.000)
male 0.007 -0.040*** 0.028***
(0.006) (0.009) (0.003)
race (ref=1.Caucasian)
2. African-American -0.060*** -0.054*** -0.013***
(0.008) (0.012) (0.004)
3. Other -0.081*** -0.046*** 0.010
(0.011) (0.017) (0.008)
marital status (ref=1.single)
2. married/partnered 0.020 0.037* -0.005
(0.016) (0.020) (0.009)
3. separated/divorced/widowed 0.006 0.000 -0.011
(0.016) (0.021) (0.009)
education (1. highschool diploma, GED)
2. college degree 0.193*** 0.030** 0.038***
(0.011) (0.012) (0.006)
3. graduate degree 0.307*** 0.112*** 0.030***
(0.013) (0.015) (0.008)
health conditions 0.007** -0.008* 0.005***
(0.003) (0.005) (0.002)
challenges to ADLs, std 0.006** -0.004 0.000
(0.003) (0.005) (0.002)
personality: organized 0.005 -0.013** -0.002
(0.004) (0.006) (0.002)
personality: responsible 0.030*** 0.034*** 0.001
(0.006) (0.011) (0.004)
personality: hardworking -0.034*** -0.018** -0.012***
(0.005) (0.009) (0.003)
personality: intelligent 0.021*** -0.013* 0.007**
(0.005) (0.008) (0.003)
personality: thorough 0.014*** -0.001 0.006**
(0.005) (0.007) (0.003)
constant -0.235 -2.321*** 0.164*
(0.157) (0.237) (0.093)
dependent variable mean 0.068 0.436 0.039
observations 115432 115432 112244
Note: ARIS refers to the automation-resistance of individual skills index, for each worker. Outcomes of interest for columns 1, 2, and 3 are whether
or not the job is white-collar, in the same industry as the worker's career job, and in STEM, respectively. Only the second stage results are shown.
Year- and state- fixed effects are included in the regression equations but not reported in the table. Standard errors are in parentheses.
*p<0.1 **p<0.05 ***p<0.01
Source: Author's calculations using the HRS.
109

Table 2.8A. Monetary and Mental Health Returns to Skill-to-Task Matching
(1) (2) (3) (4) (5) (6) (7) (8) (9)
Y
ARIS index i , for individuals 0.359*** 0.349*** -0.003 -0.003 -0.005** 0.005 -0.002* -0.001 -0.002
(0.040) (0.039) (0.254) (0.002) (0.003) (0.010) (0.001) (0.001) (0.004)
AROT index it , for occupations -0.041 -0.059 1.400** 0.003 0.002 -0.018 0.002 0.003 -0.006
(0.089) (0.088) (0.583) (0.008) (0.008) (0.050) (0.004) (0.004) (0.022)
(ARIS i ) x (AROT it ) -0.651***-0.618*** -1.872*** -0.005* -0.007*** -0.018* -0.001 -0.003*** -0.008*
(0.039) (0.038) (0.200) (0.003) (0.003) (0.010) (0.001) (0.001) (0.005)
cohort (ref= cohort 1, born before 1930)
1(cohort 2, born in 1931-1941) - - -2.919*** - 0.062** 0.059** - 0.058*** 0.055***
- - (0.283) - (0.027) (0.028) - (0.013) (0.013)
1(cohort 3, born in 1942-1959) - - -11.003*** - 0.153*** 0.148*** - 0.084*** 0.081***
- - (0.294) - (0.031) (0.032) - (0.014) (0.014)
(AROT index it ) x 1(cohort 2 i ) - - 0.618** - - -0.008 - - 0.003
- - (0.263) - - (0.010) - - (0.004)
(AROT index it ) x 1(cohort 3 i ) - - 0.239 - - -0.013 - - -0.001
- - (0.256) - - (0.010) - - (0.004)
(ARIS index i ) x 1(cohort 2 i ) - - -1.402** - - 0.028 - - -0.001
- - (0.600) - - (0.052) - - (0.023)
(ARIS index i ) x 1(cohort 3 i ) - - -1.539*** - - 0.018 - - 0.015
- - (0.592) - - (0.051) - - (0.022)
(AROT index it ) x (ARIS index i) x 1(cohort 2 i ) - - 0.826*** - - 0.010 - - 0.004
- - (0.212) - - (0.010) - - (0.005)
(AROT index it ) x (ARIS index i) x 1(cohort 3 i ) - - 1.580*** - - 0.013 - - 0.005
- - (0.203) - - (0.010) - - (0.005)
age 0.118 0.088 - -0.026** - - 0.007 - -
-0.112 -0.113 - (0.011) - - (0.005) - -
age-squared 0.004*** 0.004*** - 0.000 - - -0.000*** - -
-0.001 -0.001 - (0.000) - - (0.000) - -
male -1.703***-1.700*** -0.585*** -0.115***-0.140***-0.140***-0.061***-0.076***-0.076***
-0.134 -0.134 -0.135 (0.012) (0.012) (0.012) (0.006) (0.006) (0.006)
race (ref=1.Caucasian)
2. African-American 0.144 0.322* 0.141 0.026 0.065*** 0.064*** -0.092***-0.088***-0.088***
-0.176 -0.178 -0.181 (0.019) (0.018) (0.018) (0.008) (0.008) (0.008)
3. Other -0.841***-0.743*** -1.089*** 0.135*** 0.178*** 0.178*** -0.018* -0.012 -0.012
-0.25 -0.249 -0.254 (0.027) (0.027) (0.027) (0.010) (0.010) (0.010)
marital status (ref=1.single)
2. married/partnered -0.554* -0.274 -0.155 -0.138***-0.201***-0.200***-0.043***-0.037***-0.037***
-0.294 -0.291 -0.291 (0.033) (0.033) (0.033) (0.014) (0.014) (0.014)
3. separated/divorced/widowed -1.150***-0.965*** -0.258 0.038 0.006 0.006 0.009 -0.003 -0.002
-0.308 -0.305 -0.303 (0.034) (0.035) (0.035) (0.015) (0.015) (0.015)
education (1. highschool diploma, GED)
2. college degree -0.205 -0.133 0.105 -0.044***-0.090***-0.090*** 0.035*** 0.030*** 0.030***
-0.196 -0.196 -0.196 (0.016) (0.016) (0.016) (0.009) (0.008) (0.008)
3. graduate degree -0.003 0.18 0.576** -0.055***-0.118***-0.117*** 0.046*** 0.038*** 0.039***
-0.229 -0.231 -0.237 (0.019) (0.019) (0.019) (0.010) (0.010) (0.010)
health conditions 1.479*** 1.361*** 1.613*** 0.174*** 0.172*** 0.172*** 0.167*** 0.165*** 0.165***
-0.07 -0.069 -0.069 (0.008) (0.008) (0.008) (0.004) (0.004) (0.004)
challenges to ADLs, std 1.305*** 1.161*** 1.098*** 0.298*** 0.314*** 0.314*** 0.018*** 0.025*** 0.025***
-0.081 -0.081 -0.08 (0.013) (0.013) (0.013) (0.005) (0.005) (0.005)
personality: organized - 0.519*** 0.602*** -0.042***-0.042***-0.042***-0.014***-0.014***-0.014***
- -0.096 -0.097 (0.010) (0.010) (0.010) (0.005) (0.005) (0.005)
personality: responsible - 0.625*** 0.767*** -0.086***-0.102***-0.102*** -0.013* -0.017** -0.017**
- -0.164 -0.165 (0.018) (0.019) (0.019) (0.008) (0.008) (0.008)
personality: hardworking - -2.211*** -2.414*** -0.012 -0.012 -0.012 -0.011* -0.014** -0.014**
- -0.123 -0.124 (0.014) (0.014) (0.014) (0.006) (0.006) (0.006)
personality: intelligent - -0.318*** -0.351*** -0.090***-0.089***-0.089*** 0.004 0.005 0.005
- -0.122 -0.123 (0.012) (0.012) (0.012) (0.005) (0.006) (0.006)
personality: thorough - 0.103 0.111 -0.044***-0.053***-0.053*** 0.003 0.003 0.003
- -0.109 -0.111 (0.011) (0.011) (0.011) (0.005) (0.005) (0.005)
ln(income) - - - -0.009*** - - -0.002*** - -
- - - (0.001) - - (0.000) - -
ln(other HH income) - - - -0.004*** - - 0.002*** - -
- - - (0.001) - - (0.000) - -
totalHH wealth (Winsorized at 0.1) - - - -0.044*** - - -0.005** - -
- - - (0.004) - - (0.002) - -
constant -4.276 1.161 22.920*** 2.018*** 0.910*** 0.915*** 0.238* 0.254*** 0.256***
-3.517 -3.556 -0.755 (0.339) (0.084) (0.084) (0.144) (0.035) (0.035)
dependent variable mean -2.291 -2.291 -2.291 -0.113 -0.113 -0.113 0.171 0.171 0.171
observations 115432 115432 115432 115432 115432 115432 115432 115432 115432
ln(income), RE CES-D, RE diag. depression, RE
Note: ARIS refers to the automation-resistance of individual skills index, for each worker. AROT refers to the automation-resistance of occupational tasks, for each job.
The oldest cohort (Cohort 1) consists of individuals born before 1930. The middle cohort (Cohort 2) is made up of individuals born between 1931 and 1941. Lastly, the
youngest (Cohort 3) include individuals born between 1942 and 1959. Only the second stage (outcome equation estimation) is shown. Year- and state- fixed effects are
included in the regression equations but not reported in the table. Standard errors are in parentheses. For columns 5 and 8, the reference category is the earliest cohort--
those born before 1930. For columns 3, 6, and 9, the reference category is (ARIS index)x(AROT index)x(cohort 1). *p<0.1 **p<0.05 ***p<0.01
Source: Author's calculations using the HRS.
110

Chapter 3

Lost in Complexity? The Impacts of Health Insurance Literacy on
ACA-Induced Retirement

3.1 Introduction

In the US, there exists a uniquely strong linkage between employment and health
insurance whereby most individuals obtain insurance from their employers until they reach the
Medicare eligibility age of 65 (Anderson et al., 2003; Garber and Skinner, 2008). In 2010, 63.3%
of firms—with more than 200 workers—still provided employer-sponsored insurance (ESI) to
their employees for their active working-years (Claxton et al., 2010). Given the ties between
insurance and employment, any policies altering the current structure of insurance provision is
likely to significantly impact labor supply behaviors.
Among employers who sponsor insurance, only 36.3% additionally offer retirement-
health insurance (RHI), an extension of the ESI coverage during the post-retirement years
(Claxton et al., 2010). The literature has found that employer-sponsored RHIs encourage early
retirement by providing an insurance buffer for workers considering retiring before the Medicare
eligibility age (Blau and Gilleskie, 2006; French and Jones, 2011; Johnson et al., 2003). In
contrast, the majority of workers who only have ESI, and not RHI, are not only discouraged from
retiring early, but may also feel locked in their jobs. In a phenomenon known as job-lock,
individuals who do not have employer-sponsored RHIs or affordable alternatives during the post-
111

retirement years cannot freely leave their jobs in the fear of losing coverage (Blau and Gilleskie,
2006; French and Jones, 2011; Rust and Phelan, 1997).
78

Passed in 2010, the Affordable Care Act (ACA) created a new insurance option through
establishing the health insurance exchange.
79
The health insurance exchange helps individuals
enroll in affordable medical insurance, independently of their employer. A universal application
is used to facilitate access, and individuals whose household income falls within the range of 100
and 400 percent of the federal poverty level (i.e., the 400% is equivalent to $48,500 in 2015 in a
four-member household) can have their insurance premiums subsidized (Shoven and Slavov,
2013).
80
Because the newly introduced exchange creates a viable, cost-effective alternative to
RHI by allowing all individuals under 65 to purchase plans regardless of their employment
statuses, the job-lock theory postulates that the ACA should accelerate retirement—particularly
among employees without employer-sponsored RHI benefits (French and Jones, 2011; Gallen
and Mulligan, 2013; Gustman et al., 2018; Kaiser Family Foundation 2011; Nyce et al., 2013).
However, recent studies examining this relationship have found conflicting results—with
some showing that the ACA holds a significant influence on retirement while others finding null
effects (Ayyagari, 2017; Gustman, Steinmeier, and Tabatabai, 2018; Levy, Buchmueller, and

78
Prior to ACA, the RHIs served as a main source of insurance for early-retirees exiting labor force before the
Medicare eligibility age of 65 (Robinson & Clark 2010). While insurance facilitated through the Consolidated
Omnibus Budget Reconciliation Act (COBRA) was the closest alternative to the RHIs, the job-lock removing effect
of the COBRA was not as significant as that of the RHIs. First, COBRA was offered for up to 18 months—a time
frame too short to give individuals sufficient retirement security to plan for early retirement. Also, retirees on
COBRA had to bear the full amount of the insurance premiums plus a 2% additional administrative fee. Given the
short duration and the high costs of insurance provided through COBRA, the RHIs remained the best option for
aging workers to retain health insurance in their post-retirement years—should they choose to exit the labor force
before age 65 (Shoven and Slavov, 2014).
79
We use the terms “exchange” and “marketplace” interchangeably.
80
This subsidy rule covers the majority of Americans. The FPL in 2013 was $11,490 for singles and $15,510 for
married; the subsidy covering up to the 400 percent of the FLP will extend to $46,879 for singles and $63,280 for
married—assuming a 2% cost of living adjustments in the FPL for 2014. The sliding scale of ACA subsidies for the
silver plan is as follows: 2% for income level up to 133% of FPL; 3-4% for 133-150% of FPL; 4-6.3% for 150-
200% of FPL; 6.3-8.05% for 200-250% of FPL; 8.08-9.5% for 250-300% of FPL; and 9.5% for those earning 300-
400% of the FPL. (Shoven and Slavov, 2013).
112

Nikpay, 2015; Dillender et al., 2016). Those who found no effect of the ACA on either
retirement decisions or retirement plans utilized analysis periods ending in 2014—tracing only 1
to 2 years after the legislation for the exchanges (Gustman et al., 2018; Levy et al., 2015). The
short time-span since the ACA’s enactment likely contributed to a limited ability to observe the
reform’s full impact on labor supply behaviors. Studies that have found an ACA-associated
increase in retirement plans were limited to specific industries—such as retail, accommodation,
and food services—or did not distinguish between the effects of the health insurance exchanges
and other policies initiated by the ACA that can affect retirement decisions (e.g., Medicaid
expansions, individual mandates to purchase health insurance, reforms to benefit design, etc.).
Moreover, one potentially important, yet relatively overlooked, factor in the assessment
of the role played by the exchanges on labor supply is health insurance literacy (Bhargava et al.
2015; Handel, 2013; Abaluck and Gruber, 2016; Myerson 2018). In devising a comprehensive
healthcare reform, policy-makers behind the ACA introduced a myriad of policy changes to
expand insurance coverage, control healthcare costs, and improve healthcare service and product
delivery systems. Given the sheer number of policies and their interactions, it can be daunting for
a lay person to understand all moving parts of the ACA and to act on the newly created
opportunities. Recognizing the complexity in specifically insurance plan selection, policy-
makers mandated that states implement in-state ACA Assister Programs in 2013. The programs
are designed to train Assisters (e.g., Navigators and In-Person Assisters) so that they can help
consumers improve their understanding and usage of the insurance marketplaces—or their
overall insurance literacy.
In this chapter, we shed new light on the impact of the health insurance exchange on
retirement decisions by considering a longer analytic period, distinguishing between different
113

policies introduced by the ACA (i.e., accounting for conflation between the opening of insurance
exchanges and Medicare expansion), and importantly, identifying the role of insurance literacy
measured by the degree of availability of the ACA Assister Programs.
81
Relying on data from the
Health and Retirement Study (HRS) between the years 2006 and 2016, we examine the
retirement plans of individuals between ages 40 and 64 who are currently working and has ESIs.
First, we employ a difference-in-difference (DID) framework where we compare (1) individuals
with ESIs and RHIs through employers to individuals with ESIs but no RHIs (2) before and after
the introduction of the health insurance exchange. Next, using a triple difference (DDD)
regression, we additionally compare individuals with varying levels of health insurance literacy,
which changes differentially across states due to state-specific differences in the amount of
funding injected into the ACA Assister Programs. We assess the robustness of our results by
including a rich set of covariates and state-specific year trends.
Because an individual’s purchase of private insurance depends on not only the
availability of the health insurance exchange but also its generosity, we re-run the DID and DDD
regressions while replacing the second—temporal—difference with the between-state differences
in the exchange generosity based on the benchmark premium rates and subsidies.
82
Steering
away from the simple DDD setup, we also conduct a staggered DDD modeling by replacing the
uniform timing of the first open enrollment period of exchanges (i.e., the source of variations for
the second difference) with the varied timing at which the states enacted the legislation for the

81
“Health insurance literacy measures the degree to which individuals have the knowledge, ability, and confidence
to find and evaluate information about health plans, select the best plan for their financial and health circumstances,
and use the plan once enrolled.” https://www.huffingtonpost.com/bruce-caswell/health-insurance-literacy-is-key-
to-reaching-uninsured-populations-this-open-enrollment-period_b_8524332.html
82
The benchmark premium is calculated for the second lowest silver plan of a 40-year-old whose income level is at
the 210% of poverty.
114

exchanges. Finally, we consider an alternative planned retirement outcome: the expected
probability of continue working at age 65.
We find that, compared to those who had RHIs through employers, those without RHIs
accelerate their expected timing of retirement by 0.31 years. Based on the triple-differences
estimation results, a one-percent increase in insurance literacy (i.e., measured by the financial
support injected to the ACA Assister Programs per uninsured person) further increases the effect
of the exchanges on early retirement for the treated group by 0.41 years. Our findings are
consistent with the predictions of the job-lock theory which postulates that ACA will hasten
retirement for those job-locked by increasing their access to post-retirement insurance
independent of employment. Moreover, we add to a small but growing literature on health
insurance literacy by emphasizing the role of literacy levels in individual decision-making. Our
results indicate that complex policy changes—specifically within insurance provision—ought to
be complemented by efforts to increase the target population’s understanding of the changes.

3.2 Conceptual Framework

We begin by explaining the rationale behind job-lock. We apply the theoretical
framework of insurance-induced job transition behaviors, introduced by Gruber and Madrian
(2002), to the context of retirement decisions. Then we consider how insurance literacy can
affect insurance-induced job lock and retirement decisions.

115

3.2.1 Insurance-Induced Job Lock

To illustrate the role of insurance-induced job-lock on retirement decisions, assume a
simplified world where labor supply decisions of workers involve deciding only between
employment (i.e. ‘working’) and non-employment (i.e., resorting to full and permanent
retirement before the Medicare eligibility age of 65). All firms here are considered identical, thus
creating no incentives for the workers to choose among firms. Consider an individual who is
currently working and trying to decide whether to continue working or retire. For simplicity,
assume that her utility is a function of only wage, leisure, and access to health insurance, and
assume that wage is the sole product of labor. In the absence of insurance, it is straightforward to
show that it is optimal for this individual to retire if her marginal utility of taking full-retirement
is greater than her marginal utility from the marginal product of labor (i.e. wages).
The dynamics can change when insurance is added to the picture: if her valuation of
maintaining the health insurance coverage is sufficiently large—larger than the preference for
retirement—and if she were to lose coverage entirely by retiring without RHIs from her
employer, the insurance ‘locks’ her in her job. The strength of the job lock will depend on the
individuals’ value of health insurance and their ability to access health insurance independent of
employment.
We apply this conceptual framework to the context of the ACA. Workers with both ESI
and RHI already have insurance coverage guaranteed for the post-retirement years, so they can
retire from their current job at any time without a fear of losing coverage. In contrast, workers
with ESI and no RHI will lose coverage if they retire from their current jobs. Thus, in the
absence of other access to post-retirement insurance, individuals with ESI and not RHI will be
locked into their jobs to maintain insurance coverage. The introduction of exchanges in the ACA
116

should have a greater impact on the retirement decisions of workers with only ESI and no RHI,
relative to those with both ESI and RHI.
From this framework, we devise our first hypothesis: For individuals who do not have
access to RHIs through employers, the ACA will have a job-lock removing effect similar to that
of the RHIs—accelerating the timing of retirement.

3.2.2 Insurance Literacy, Insurance Uptake, and Retirement

In contrast to what is theorized above, recent studies that evaluate the impact of the ACA
on retirement decisions have generated puzzling results. Levy, Buchmueller, and Nikpay (2015)
examine discrepancies in actual retirement patterns in 2015 between workers in states that
anticipate the adoption of state-based insurance exchange through ACA and those in non-
anticipating states. Even after controlling for the impact of Medicaid expansion, the authors
found no significant change in actual retirement patterns of individuals in the ACA-participating
states. Considering that workers may not be able to instantaneously act upon the retirement
incentives embedded in the ACA, Gustman, Steinmeier, and Tabatabai (2018) examine changes
in the retirement plans as opposed to actual retirements behaviors, but they also found no
significant evidence that ACA has affected retirement.
While Dillender et al. (2016) and Ayyagri (2017) provide suggestive evidence that the
short analysis period can explain the null findings, there can be other understudied elements that
are in play. In particular, we suspect the impact of ACA can be affected by the insurance literacy
of the target beneficiaries—in other words, the extent to which target beneficiaries actually
understand the changes in insurance provision.
117

A priori the empirical evidence, it is unclear how insurance literacy affects retirement
decisions when insurance access expands. Assuming risk-aversion and Jensen’s Inequality, both
of which ensure the strict concavity of worker utility functions, the expected utility of an
insurance expansions will be lower for the individual with lower literacy.
83
This observation is
independent of an individual’s monthly insurance premium and probability of falling ill, and it
holds true under the assumption that those with low health insurance literacy do not overestimate
the true payout amount of insurance plans (De Meza and Webb, 2001; Myerson, 2018). In this
case, lower-literacy individuals will face greater uncertainty from unknown insurance payouts.
As a result, they are more likely to remain working, fail to take up post-retirement insurance
from the exchange, remaining under the influence of ESI-induced job lock. Because of
uncertainty, those without employer-sponsored RHI and higher literacy will be more likely to
purchase insurance plans through the exchange and free themselves from job-lock, relative to
their counterparts with lower literacy.
Yet at the same time, assuming there is a strong and positive association between ‘health
insurance literacy’ and ‘health literacy,’ it is possible for individuals with low literacy to over-
enroll in the exchanges as they over-estimate their future health risks (Myerson, 2018).
84
The
divergent directionality of the insurance uptake scenarios implies their having opposing effects
on retirement, and the empirical results of the estimation model in this chapter will show the net
effect of the two.

83
Assuming all individuals regardless of their literacy levels are risk averse, their preferences for any insurance F(∙)
can be written using the Jensen’s Inequality in terms of insurance F(∙) and utility u(x):
∫ 𝑢 (𝑥 )𝑑𝐹 (𝑥 ) ≤
∞
−∞
𝑢 (∫ 𝑥𝑑𝐹 (𝑥 )
∞
−∞
), for all F(∙).
84
Title V of the Patient Protection and Affordable Care Act (2010) defines health literacy as the capacity to obtain
and process health information to make accurate health decisions.
118

3.2.3 Background on ACA Health Insurance Exchange and Assister Programs

In this chapter, we will rely on variation across time from the establishment of health
insurance exchanges and variation across states in ACA Assister Program funding. There are
three types of insurance exchanges: state-based, state-federal partnership, and federally-
facilitated.
85
A state that has a state-based exchange is responsible to oversee all exchange
functions including their IT platform (i.e. exchange website) from which consumers can
purchase insurance plans. A state with a state-federal partnership relies on the Department of
Health and Human Services (HHS) to handle all exchange functions except the management of
insurance plans offered on the exchange. Lastly, a state with a federally-facilitated exchange
grants the HHS full authority to oversee all exchange functions.
There is a considerable variation in the timing at which the states passed legislations for
launching the insurance exchanges—from as early as Sept. 2010 in California to July 2013 in
Delaware.
86
Further details on the dates and the title of the legislations are provided in the
Appendix, Table 3.1A. Regardless of the legislation dates, all states were required by federal law
(45 CFR 155.410) to hold their first open enrollment periods for the insurance exchanges
between October 1, 2013 and March 31, 2014. Because our data (discussed in Section 3.3) is
biennial, our main specifications do not rely on this month-to-month variation, and instead, treat
2014 as the first period in which the exchanges launched.

85
Definitions for the types of marketplace are provided below (source: https://www.kff.org/health-reform/state-
indicator/state-health-insurance-marketplace-types). In this study, we equate state-based and state-based
marketplaces that incorporate federally-facilitated Marketplace IT platform.
86
While states with federally-facilitated exchange did not pass legislations, we refer to the timing at which the states
made official announcement of their intention to launch the federally-facilitated exchange, thus granting full
authority to the HHS to oversee their marketplaces.
119

Since the start of the insurance marketplaces in late 2013, all states have been required by
law to implement in-state ACA Assister Programs to consumers in need. The programs are
designed to train assisters so that they can help consumers improve their understanding and
usage of the insurance marketplaces—and as a result, increase their overall insurance literacy.
While all states exchanges maintain Assister Programs, their nature, as well as funding source,
depends on the type of the exchanges.
There are five funding sources are available for the Assister Programs: Navigators, In-
Person Assisters (IPA), Certified Application Counselors (CAC), Federally-Qualified Health
Center Assisters (FQHCA), and Federal Enrollment Assisters (FEA). While the CACs comprise
the largest of the marketplace assister programs in terms of the number of personnel (i.e. 45-65%
between 2014 and 2016), their funding comes from a variety of private sector sponsorships in
lieu of states and thus are particularly difficult to track the funding sources (Pollitz et al., 2014).
Hence, we exclude CAC funds when calculating each state’s total amount of Assister grants
received. We also ignore funding from FEAs since they comprise only 0-1% of the funding
source between 2014 and 2016 (Pollitz et al., 2014). In short, the rest of this chapter refers only
to Navigators, IPAs, and FQHCAs when referring to the Assister Program funding sources.
As shown in Table 3.1, these three types of ACA Assister Programs are used in different
combinations, based on the type of insurance exchange.
87
All states are eligible for and have
received the FQHCA funding. The Navigator funding is given only to states with federally-
facilitated and federal-state partnership marketplaces. Lastly, the IPA funding is given only to
the states with partnership and state-based marketplaces.

87
Definitions for the types of marketplace are provided below (source: https://www.kff.org/health-reform/state-
indicator/state-health-insurance-marketplace-types). In this study, we equate state-based and state-based
marketplaces that incorporate federally-facilitated marketplace IT platform. The names of the ACA Assister
Programs are equivalent to the names of their funding sources.
120

3.3 Study Data and Methods
3.3.1 Analysis Sample
Data is drawn from the Health and Retirement Study (HRS), a biennial survey
representative of older Americans.
88
The HRS contains information on retirement plans,
retirement expectations, employment history, health, health insurance, financial and housing
wealth, income, Social Security and pension, family structure, and basic demographics of age-
eligible respondents and their spouses/partners. For this study, we use the HRS data from 2006 to
2016 and merge the restricted-use geographic and occupational identifiers.
We confine our sample to individuals ages 40 to 64, all of whom are still working and
have ESIs for active-working years. In doing so, we exclude individuals aged 65 and above as
they are eligible for Medicare. For all estimations, our main outcome variable is the age of
planned retirement, defined as the age at which an individual expects to leave her current job.
We include individuals who plan to leave their current occupations to enter full-retirement as
well as those entering partial retirement (i.e. getting part-time jobs and starting self-employed
work) because the removal of the ESI-induced job lock in an individual’s current occupations
can lead to both full- and partial- retirements. While planned retirement may differ from actual
retirement, the planned retirement variable offers two important advantages. First, the short lapse
since the introduction of the exchanges precludes the ability to observe the full effect of the ACA
on actual retirement behaviors, and the planned retirement variable allows us to examine

88
Currently, a total of thirteen HRS waves are available from 1992 to 2014. Throughout the two decades, the HRS
has included six cohorts: HRS cohort born between 1931 and 1941, AHEAD cohort born before 1924, Children of
Depression (CODA) cohort born 1924 to 1930, War Baby (WB) cohort born 1942 to1947, Early Baby Boomer
(EBB) cohort born 1948 and 1953, and lastly, Mid Baby Boomer born between 1954 and 1959.
121

anticipated changes in an individual’s retirement behavior. Second, the planned measure—which
is ascertained in each wave of the survey—allows us to capture within-individual variations in
retirement decisions across time (i.e. individual fixed effects).
89

The HRS offers a rich set of variables that are potential determinants of retirement,
including individual demographic information (i.e. age, marital status, and number of
dependents), a natural log of individual annual earnings, and a natural log of household total net
wealth.
90
In order to control for retirement incentives embedded in employer-sponsored pension
plans, we categorize individuals based on whether they have access to defined-benefit pension
plans, defined-contribution plans, or do not have any coverage at each interview year.
91
We also
identify individuals who collect Social Security income from those who do not. As we include in
our sample only those who are working and under age 65, SS income receipt is a proxy for those
with limited financial means. The HRS respondents’ health status is captured via a health index
that takes into account a number of critical, work-impeding medical conditions; the index is
normalized to have mean of zero and a standard deviation of one, with a higher value
corresponding to better health.
92

Whether an individual purchases private insurance plans at the exchange depends not
only the availability of the exchange but also its generosity. A simple way to estimate the

89
Most individuals who fully retire do not revert to working again, so there is little variation in retirement behavior
within an individual.
90
Individuals who are retired and thus have zero earnings are excluded from the data. The net wealth includes
information on the checking and savings accounts, bonds, stocks, deposits, mutual funds, primary and secondary
housing, and any other savings less debt has also been included as a covariate.
91
Individuals who are covered in the defined-benefit pension plans tend to face stronger early retirement incentives
due to the particular nature of the defined-benefit accrual (Shoven and Slavov, 2014).
92
The conditions include 1) high blood pressure or hypertension; 2) diabetes or high blood sugar; 3) cancer or a
malignant tumor of any kind except skin cancer; 4) chronic lung disease except asthma such as chronic bronchitis or
emphysema; 5) heart attack, coronary heart disease, angina, congestive heart failure, or other heart problems; 6)
stroke or transient ischemic attack (TIA); 7) emotional, nervous, or psychiatric problems; and 8) arthritis or
rheumatism.
122

generosity of exchanges is to compare their benchmark premiums and subsidies. We calculate
the monthly premium for the second-lowest silver plan of a 40-year-old non-smoker earning an
average annual salary at 210% of poverty level (i.e. approximately $50,000) in the largest city of
each state. The post-tax credit subsidy premium rates for the silver plan in many states end up
being similar despite large differences in their pre-tax credit subsidy premium amounts due to the
varying levels of subsidy. Thus, we estimate the exchange generosity by the maximum
percentage of the pre-tax credit subsidy premium subsidized in each state.

3.3.2 Measuring Health Literacy

We measure individuals’ health insurance literacy using state-specific variations in the
provision of ACA Assister Program. In doing so, we take into consideration the fact that the
ACA and its Assister Programs are set to alter not only the availability of insurance but also the
level of insurance literacy over time. Since its launch in 2013, the ACA Assisters provide
outreach and education, technical assistance necessary to navigate the exchanges, and guidance
in making informed plan-choices to individuals who are eligible to purchase plans and yet suffer
from low insurance literacy (Pollitz et al., 2016). Given the purpose of Assisters, we assume that
the increased ACA Assister Program funding implies an increase in the insurance literacy of
their targeted beneficiaries.
Explained in the background section (Section 3.2.), we focus on three major streams of
funding for the Assisters: Navigator-, IPA, and FQHCA grants. The FQHCA funding comes
from the Health Resources and Services Administration (HRSA), which reports total funding
amounts by year and state (MAC, 2015). Even though a small portion of the funds are used for
extraneous tasks as shown in HRSA’s breakdown of funding by program usages, most of the
123

FQHCA funds are implemented to promote education, outreach, and enrollment in the
exchanges.
We accumulate the information on the annual Navigator funds as well as the one-time
IPA grants from the state health marketplace profiles at Kaiser Family Foundation page.
93
The
Navigator funds are distributed annually by the Department of Health and Human Services
(HHS), and all states are required by law to maintain the Navigator programs. The IPA funds are
one-time funds from state-based marketplaces whose lump sums are given in 2013, before the
first enrollment period. Unlike the Navigators, IPA programs are recommended but not
mandatory, and only state-based exchanges are eligible for IPA funding.
94
States with missing
IPA grant information are assumed to have not implemented the program.
The total funding variations that arise across states are not minor and are due to the
different levels of generosity of the three funding sources (i.e. Navigator-, IPA-, and FQHCA-
grants). The IPA and FQHCA funding programs give grants whose amount matches the budgets
proposed by the applicants in the partnership- and state-based exchange states. On the contrary,
the Navigator funding’s grant amounts tend to be more rigid, strictly based on the population size
of each state and a set minimum funding floor. States that are eligible for IPA funding tend to
receive greater amounts (i.e. with an average of $95.66 per uninsured in 2014 and 2015) than
those that are ineligible (i.e. an average of $54.88 per uninsured in 2014 and 2015). While the
Assister funding variations arise due to the types of exchanges available in each state, there is
little to no possibility of a reverse causality where the types of exchanges are dictated by the
Assister funding amounts: the legislation for the exchanges preceded that of the Assister funding

93
See https://www.kff.org/state-health-marketplace-profiles/.
94
For Navigator funding, see 45 CFR § 155.210 at https://www.law.cornell.edu/cfr/text/45/155.210. The IPA
funding is also known as the non-Navigator funding. N(See 45 CRF 155.205 (c), (d), and (e)).
124

programs—indicating that the chances for the states to choose the type of exchange to embrace
based on the Assister funding amounts available in different funding sources.
We sum the amounts of funding granted for Navigators, IPAs, and FQHCA programs by
state and year. To proxy the extent to which each potential marketplace enrollee can benefit (i.e.
have her insurance literacy increased) from the ACA Assisters, we divide the total funding
amount by the number of uninsured individuals per state and year and take a natural log.
95

3.4 Empirical Approach

We begin with a difference-in-differences (DID) estimation that treats the introduction of
the insurance exchanges as an exogenous shock. We compare (1) individuals with ESIs and
RHIs through employers (i.e. control), and (2) individuals with ESIs but no RHIs (i.e. treatment),
before and after the first open enrollment of the exchanges in 2014.
96
The introduction of the
exchanges will affect only our treated individuals because those in the control group already had
access to post-retirement health benefits prior to the ACA. The DID specification is not written
as it can be inferred from the triple-difference model below.
Next, using a third difference stemming from varying levels of Assistor program funding,
we conduct triple-differences (DDD) regressions. For individual i in state s and year t, we
estimate,

95
Uninsured (nonelderly, below age 65) by state: https://www.kff.org/other/state-indicator/nonelderly-0-
64/?currentTimeframe=3&sortModel=%7B%22colId%22:%22Location%22,%22sort%22:%22asc%22%7D.
Individual’s use of Assister funding is unobserved.
96
Aside from access to RHIs, there are no systematic differences between individuals without RHIs (i.e., our
treatment group) and those with RHIs (control group). Once an employer decides on an insurance package—which
might include RHIs—it is provided unilaterally to all full-time employees. In other words, the employer does not
selectively provide RHIs to a subset of employees based on the perceived ‘qualities’ or ‘merits’ (Kramer, 2012).
125

(17) 𝑌 𝑖𝑡𝑠 = 𝛽 0
+ 𝛽 1
1(𝑇𝑟𝑒𝑎𝑡 𝑖𝑡
) + 𝛽 2
1(𝑃𝑜𝑠𝑡 𝑡 ) + 𝛽 3
ln (𝐿𝑖𝑡𝑒𝑟𝑎𝑐𝑦 𝑡𝑠
) +
𝛽 4
1(𝑇𝑟𝑒𝑎𝑡 𝑖𝑡
)1(𝑃𝑜𝑠𝑡 𝑡 ) + 𝛽 5
1(𝑇𝑟𝑒𝑎𝑡 𝑖𝑡
)ln (𝐿𝑖𝑡𝑒𝑟𝑎𝑐𝑦 𝑡𝑠
) + 𝛽 6
1(𝑃𝑜𝑠𝑡 𝑡 )ln (𝐿𝑖𝑡𝑒𝑟𝑎𝑐𝑦 𝑡𝑠
) +
𝛽 7
1(𝑇𝑟𝑒𝑎𝑡 𝑖𝑡
)1(𝑃𝑜𝑠𝑡 𝑡 )ln(𝐿𝑖𝑡𝑒𝑟𝑎𝑐𝑦 𝑡𝑠
) + 𝛽 8
𝑋 𝑖𝑡𝑠 + λ
i
+ 𝛾 𝑠 + δ
t
+ 𝜖 𝑖𝑡𝑠 ,

where Yits is the planned retirement age for individual i living in state s at time t, and 1(𝑇𝑟𝑒𝑎𝑡 𝑖𝑡
)
is a binary indicator that equals one if the individual is in the treatment group. 1(𝑃𝑜𝑠𝑡 𝑡 ) equals
one from 2014 onward. The ln(𝐿𝑖𝑡𝑒𝑟𝑎𝑐𝑦 𝑡𝑠
) equals the natural log of funding from ACA Assister
Programs, which we use a proxy for insurance literacy. 𝑋 𝑖𝑡𝑠 is a vector of time-variant covariates
comprised of demographics and determinants of retirement decisions (e.g., health, income,
wealth, pension access and types, access to Social Security income, and access to other health
insurance plans. The vector also includes an identifier variable that distinguishes Medicaid-
expansion from the non-expansion states so as to avoid potential conflation between the impact
of Medicaid expansion and introduction of exchanges. We include state (𝛾 𝑠 ), year (𝛿 𝑡 ), and
individual (𝜆 𝑖 ,) fixed-effects so that our remaining variation comes from within an individual-
state over time. Of note, the individual fixed effects allow us to control for unobserved
preferences that can affect an individual’s job choice and retirement decisions. 𝜖 𝑖𝑡𝑠 is an
idiosyncratic error term with mean zero. Standard errors are clustered at the individual level to
generate the most conservative estimates. We treat 2014 as the reference year to match the
timing of enrollment in the exchanges.
97

We conduct several robustness checks. First, we perform the DID and DDD regressions
without covariates to test for the models’ sensitivity. Second, we rerun the regressions with not
only the covariates but also an additional variable capturing the state-specific year trends. Third,

97
Since the HRS is conducted biennially, we only have responses for even numbered years.
126

in lieu of introducing a temporally uniform shock (i.e. the first open enrollment of state-based
exchanges) in 2014, we utilize the month-to-month heterogeneity in the timing of legislation for
state-based exchanges (discussed in Section 3.2.3. and shown in Appendix Table 3.1A).
Following Stevenson and Wolfers (2006), we use a staggered DDD approach. For each state, we
set the reference year equal to the year in which the legislation for the exchange passed. Fourth,
we perform non-temporal DID and DDD regressions whereby we replace the Postits variable in
eq. (17) with the generosity indicator. Lastly, we re-run all models while replacing the planned
retirement age outcome with another variable: a self-reported probability (0-100%) of continue
working at age 65.

3.5 Results
3.5.1 Assessing Baseline Trends
Our sample consists of 15,194 person-year observations, spanning over a decade from
2006 to 2016. Summary statistics of the variables used in the analysis are displayed in Table
3.2.
98,99
We use the probability weights to ensure a nationally representative sample.
100
The first
two columns display the pooled average of the variables from all pre-exchange years (i.e. 2006-
2012) for the treatment- and the control- groups, respectively. The last two columns show the

98
Categories for the married are 1. Married/partnered, 2. Divorced/widowed/separated, and 3. Never married. For
pension, 1. has DB, 2. has DC, 3. has other private pensions, and 4. has none.
99
While we have included time-invariant demographic variables in the summary statistics to better show the
characteristics of the sample used in our study, the impact of such variables is eliminated in the fixed-effect
regressions. Education categories are 0 for those with no degrees, 1 with high-school diploma, 2 with college
degrees, 3 with masters, JD, MBA, and 4 with Ph.D. and over. The categories for race is 1 for Caucasians, 2 for
African-Americans, and 3 for others.
100
We use weighted descriptive statistics for sections 3.5.1. However, for the DID and DDD regressions, both
unweighted and weighted estimates reveal no substantive difference in terms of the coefficients of interest and its
significance. Hence, we present the unweighted results in sections 3.5.3. and 3.5.4. so as to provide accurate
standard errors (Winship and Radbill, 1994).
127

pooled averages from post-exchange years (i.e. 2014-2016). The literacy variable refers to the
amount of ACA Assister funding allotted for each uninsured individual—a potential exchange
enrollee—by state and year. The generosity variable refers to the maximum extent to which the
benchmark premium is subsidized in each state’s marketplace.
We follow Imbens and Rubin (2015) in checking for the covariate balance between the
control and treatment groups in the pre-treatment years by calculating the normalized
differences. Results of the balance testing are also shown in Table 3.2. We account for the
imbalance by estimating inverse probability weights to be used for the DID and DDD
estimations as an additional robustness check.
Next, as one of the key assumptions of the DID and DDD frameworks is common pre-
trends between the treatment and control groups, we graphically plot trends in the retirement for
the control- and treatment- groups in the pre-exchange years. Shown in Figure 3.1, the parallel
trends assumption arguably holds. There is a slight narrowing of the difference between the
control and the treatment group in 2010, which could be attributed to the first announcement (i.e.
via the passage of the reform law) of the ACA. To reduce the impact of year 2010 from
weakening the common trends, we supplement our main analyses by limiting the pre-treatment
period to the years (1) before 2010 (i.e. 2006-2008) and (2) after 2010 (i.e. 2010-2014).
We also conduct a formal regression-based test to evaluate the common trends evaluation
for a multi-year treatment used by Autor (2003). With four pre-treatment periods and two post-
treatment periods, we estimate the following equation,

(18) 𝑌 𝑖𝑡
= 𝛼 + 𝜅 𝑘 ∑ (𝑌𝑒𝑎 𝑟 𝑘 )
𝑖𝑡
+ 𝜆 2016
𝑘 =2006
𝑙 ∑ (1(𝑇𝑟𝑒𝑎𝑡𝑒𝑑 ) × 𝑌𝑒𝑎 𝑟 𝑙 )
𝑖𝑡
2016
𝑙 =2006
+ 𝜂 𝑖 + 𝜖 𝑖𝑡
,

128

where 𝑌 𝑖𝑡
is the outcome that is the planned retirement age for individual i at wave t, and 𝜂 𝑖 is the
individual fixed effects. The lambda coefficients are estimated relative to the reference year (i.e.
2006) and they identify whether there are pre-treatment differences between the treatment- and
the control- groups. Standard errors are clustered at the individual level. The results from
Equation (18) are shown in Table 3.3.
The lack of significance associated with the interactions for all remaining pre-treatment
years (i.e. 2008-2012) indicate that the outcome-trends between the treatment and the control are
not significantly different in the pre-treatment period. Such result further validates the common
trends assumption. The formal test also reveals the dynamics of the change in retirement
decisions brought upon by the introduction of exchanges (i.e. how fast the effect unravels).
While the exchange was mandated to open by the end of March 2014, temporal changes in
retirement did not begin in 2014, but rather occurred in our subsequent year of available data
(2016). The lack of significance associated with the interaction term for 2014 suggests that there
was a lag in adjustments to planned retirement.

3.5.2 Relationship between the Insurance Literacy and Insurance Uptake

The key intermediary factor between insurance literacy and modifications of retirement
decisions is changes in insurance uptake. As discussed in section 3.2.2., by theory, a rise in the
insurance literacy can lead to both a rise and a fall of insurance uptake. Using empirical data, we
show that there is a net positive association between the two elements.
In Figure 3.2, we highlight the state-specific heterogeneity in Assister funding (i.e., our
measure of insurance literacy) and associated insurance enrollment into the insurance exchanges.
Plot (a) of Figure 3.2 indicates that in 2014-2016, California, Florida, and Texas having the
129

highest levels of funding and Hawaii, North Dakota, and District of Columbia having the lowest
level of funding, on average. Plot (b) shows that between 2014 and 2016, enrollment into the
exchanges also varied substantially across states. For further clarification, we plot the
relationship between the Assister funding and the insurance uptake in Figure 3.3. The fitted line
has a statistically significant slope of 0.0041, suggesting that there is a positive correlation
between funding levels and insurance uptake. These figures provide suggestive evidence that
funding to increase insurance literacy is positively associated with enrollment, lending empirical
evidence that the low-literacy individuals are more likely to under-enroll (due to the increased
uncertainty) rather than over-enroll (due to an over-estimation of health risks). We provide more
rigorous evidence of this conjecture in the subsequent DDD estimations.

3.5.3 Impact of the Exchanges on Planned Retirement

The main results from our DID regression are summarized in Table 3.4. We provide
estimates of the detailed covariates in Appendix Table 3.2A. Across all DID analyses, we find
that the introduction of the exchanges in ACA is associated with faster timing of expected
retirement. The main analysis in column (1) shows that, compared to workers with both ESIs and
RHIs through employers, those without RHIs (i.e. treatment group) accelerate retirement by 0.31
years with the opening of exchanges in ACA, demonstrating that the treatment group workers are
significantly influenced by the new opportunities for accessing the post-retirement insurance
made available via the exchanges. This result is statistically significant at 5% level, and the
estimate is robust to controlling for covariates (column (2)), our preferred specification. They are
also robust to the inclusion of state-specific year trends (column (3)).
130

Our various robustness checks, shown in columns (4)-(8), also display consistent
results.
101
When we replace the temporal difference with the level of generosity (i.e., the monthly
premium for the second-lowest silver plan as explained in Section 3.3.1), the estimated
coefficient for the non-temporal DID in column (4) suggests that a one percentage point increase
in the level of generosity of the exchange subsidy is associated with an acceleration of (expected)
timing of retirement of the treatment group by 0.52 years. The results show that insurance
uptake, as well as the subsequent adjustments in retirement decisions, depend not only on the
opening of the insurance exchanges but also—and perhaps more so—on the generosity of the
plan premiums and subsidies available at the exchanges.
Restricting the pre-treatment years to 2006 to 2008 in column (5) or 2010 to 2014 in
column (6) portray a similar picture, suggesting that our results are not affected by the inclusion
of the ACA announcement. Finally, we examine an alternative planned retirement outcome (i.e.,
probability of continue working at age 65) in column (8). The results indicate that the treatment
group decreases their probability of working at age 65 by 4.83 percent from the mean of 40.8
percent (i.e. a decrease of 1.97 percentage points) with the introduction of state exchanges.
Lastly, the weighted DID using inverse probability weights (IPW) in column (9) shows, with a
significance at the 5% level, that those in the treatment group accelerate retirement by 0.30 years.
All of the estimated DID-coefficients are in line with the job-lock theory which postulates that
ACA will have a job-lock removing effect by increasing people’s access to post-retirement
insurance.

101
All robustness checks are done to the main analysis specification that includes all covariates as in column (2).
131

3.5.4 The Role of Insurance Literacy on Planned Retirement

Table 3.5 displays the key coefficients from various triple-differences models, with
detailed results shown in Appendix Table 3.3A. Here, the coefficients of interest are associated
with the triple differences: 1(Post) x 1(Treated) x ln(literacy). Based on the main analysis in
column (1), a one-percent increase in insurance literacy (i.e., measured by the ACA Assister’s
financial support per uninsured person) increases the effect of the ACA on early retirement for
the treated group by 0.0036 years. The finding is robust to adding covariates in column (2),
which increases the statistical significance of the triple difference, and it is also robust to the
inclusion of state-specific year trends as is shown in column (3): A one-percent increase in
literacy accelerates retirement by 0.0041 and 0.0045 years, respectively. Restricting the pre-
treatment years in column (5) and (6) portray a similar story.
102

When we re-estimate Equation (1) to account for the generosity of premium subsidies
available in each exchange, we find an even stronger, additive effect of insurance literacy
funding (column (4)). A one-percent increase in literacy funding among those in the treatment
group accelerates retirement by 0.0073 years. The staggered DDD regression in column (7)
consistently shows that a one percent increase in literacy funding among those in the treatment
group further accelerates retirement by 0.0001 years. Lastly, our examination of the expected
‘probability of working at age 65’ in column (8) generate consistent results, as the one-percent
increase in literacy decreases the probability of working at age 65 by 1.64 percent (i.e. a decrease
of 0.0067 percentage points from a mean of 40.8%), respectively. A similar picture is portrayed
in the IPW-weighted DDD in column (9) which shows that those in the treatment group

102
All robustness checks are done to the main analysis specification that includes all covariates as in column (2).
132

accelerate retirement by 0.43 years. However, the DDD coefficients in columns (7) and (8) lack
significance, possibly because of the relatively short time lapse that hampers precision.

3.6 Discussion

The introduction of the insurance marketplace in ACA has made individual insurance
plans more accessible and affordable, particularly to those without jobs or insurance-sponsoring
employers. In theory, individuals who had access to ESI but not RHIs prior to ACA could be
freed from the ESI-induced job-lock through the option of purchasing individual plans from the
exchanges.
Although our study covers only two years since the introduction of exchanges and three
years since the injection of the major ACA Assister grants, we find empirical evidence of that the
exchanges did remove the job-lock effects. Our DID results show that, compared to those who
had both ESIs and RHIs through employers, those without the employer-sponsored RHIs are
more significantly influenced by the introduction of the exchanges in ACA, hastening their
expected timing of retirement by 0.31 years. With additional years of data added to the analytic
period in the future, we may be able to strengthen our estimates of ACA’s impact on workers’
actual (as opposed to planned) retirement decisions.
Additionally, we find that insurance literacy, and hence the extent to which individuals
understand the ACA-induced policy changes, play an important role in determining the
exchange’s impact on retirement plans. Using state-specific variation in ACA Assister Funding
per capita uninsured, we show that individuals with greater access to Assister funding exhibited
the largest changes in retirement decisions post-ACA. These results suggest that increased
133

insurance literacy leads to a better understanding of insurance enrollment opportunities via the
exchanges, a higher likelihood of insurance uptakes, and a further modification of subsequent
work and retirement decisions. Our results highlight the importance of promoting health- and
health insurance literacy to enhance the efficacy of the healthcare reforms.
Given the strong linkage between insurance and employment in the US, the impacts of
ACA on the extent of job-lock and other labor supply behaviors deserve attention. From an
economic perspective, the removal of the job-lock is welfare improving. When a worker is
locked in her job due to her inability to obtain the RHI, an employer—knowing that this
employee will not retire due to her high valuation of ESI—can extract surplus from the employee
and/or discriminate in wages. In this sense, the job-lock thus can generate a loss of welfare. The
misallocation of workers to firms can also result in decreased worker productivity, the retaining
of sicker employees, and reductions in entrepreneurship and innovation. Our results suggest that
the ACA reduced some of these inefficiencies.
However, from the policy perspective, the accelerated retirement can introduce a negative
side-effect: the accelerated retirement counteracting other policy efforts to foster longer working
lives, a commonly proposed strategy to combat the problems of population aging and a rapidly
shrinking tax base of the nation (Maestas and Zissimopoulos, 2010).
103
It is unclear whether the
welfare gains from reduced job-lock dominate the societal loss of early retirement, and future
research can assess the relative trade-offs of these effects.
This study provides timely insights to the current policy discourse on the possible
consequences of a dramatic reduction in the ACA Assister grants: starting in 2017, the Centers
for Medicare and Medicaid Services announced that there has been a reduction in the total ACA

103
The past efforts include the increase in the Social Security eligibility age, the termination of the Social Security
earnings test, a repeal of the mandatory retirement, etc. (Gustman et al., 2016).
134

Assister funding by 43%—from $52.9 million in 2016 to $36.1 million in 2017. A subsequent
reduction followed in July 2018, which allocated just $10 million to Assister grants.
104
While it
is certainly possible that added funding will yield diminishing marginal returns, the marginal
benefits of funding should be assessed based on not only improved insurance enrollment, but
also its spillover effects on the labor market.

104
See https://www.kff.org/health-reform/issue-brief/data-note-further-reductions-in-navigator-funding-for-federal-
marketplace-states/.
135

Figure 3.1. Trends in Planned Retirement

Source: Author’s calculations based on the HRS.

136

Figure 3.2. Correlations between the Amount of Assister Funding
Available in Each State and the Total Number of Enrollments in Exchanges

(a) ACA Assister Funding

(b) Enrollment in Insurance Exchanges

Source: Author’s calculations using KFF’s state indicators of marketplace enrollments and information on the
Assister funding allotted per uninsured person. Both figures are drawn from 2016 data.

137

Figure 3.3. Relationship between the Assister Funding and the Exchange Enrollments

Source: Author’s calculations using KFF’s state indicators of marketplace enrollment and total Assister
funding information between years 2014-2016. The fitted line is given by the equation ‘number of
enrollment= -25.84+0.0041*assister funding.’

Table 3.1. Types of Exchange and Assister Grants
Exchange Type Assister Funding Type
State-based exchange IPA, FQHCA
State-federal partnership exchange Navigator, IPA, FQHCA
Federally-facilitated exchange Navigator, FQHCA
Source: Information obtained from https://www.medicaid.gov/state-resource-
center/mac-learning-collaboratives/downloads/cnsmr-asstnce-mktplce.pdf.
138

Table 3.2. Summary Statistics and Baseline Characteristics
(1) (2) (3) (4)
Treatment Control norm Treatment Control
ESI + No RHI ESI + RHI diff ESI + No RHI ESI + RHI
pre-2014 pre-2014 >0.5 post-2014 post-2014
planned retirement age 65.962 65.204 + 66.322 66.196
probability of working at age 65 50.844 45.693 + 49.862 50.671
post 0 0 1 1
treatment 1 0 1 0
literacy 35.871 11.679 88.058 93.208
generosity 0 0 0.472 0.447
age 58.425 59.020 + 59.213 60.038
educ* 1.541 1.506 + 1.387 1.507
male* 0.630 0.574 + 0.595 0.607
race* 1.367 1.426 1.453 1.384
married 1.532 1.554 1.548 1.501
pension 3.204 2.763 + 3.133 2.713
income 36607.807 61981.783 39575.936 61425.700
wealth 122703.691 128105.408 94368.989 59818.514
health status (standardized) 0.855 0.854 0.780 0.787
access to other insurance 0.849 0.801 0.753 0.795
access to Social Security income 0.007 0.015 0.068 0.064
living in Medicaid expansion states 0.527 0.613 + 0.606 0.558
Source: Author's calculations based on the HRS data.
Notes: The summary statistics show the averages from pooled cross-sections weighted by probability weights. As we conduct fixed-effect
regressions in this paper, eliminated are these time-invariant variables (*). We conducted tests for covariate balance during the
pre-treatment years using normalized differences between columns (1) and (2). + indicates differences greater than 0.5.
139

Table 3.3. Evaluating Pre-Treatment Trends
treated=1 0.311
(0.205)
YEAR
year=2008 0.546***
(0.099)
year=2010 1.281***
(0.132)
year=2012 1.666***
year=2014 2.215***
(0.154)
year=2016 2.755***
(0.185)
TREATED x YEAR
treated=1 x year=2008 -0.187
(0.182)
treated=1 x year=2010 -0.188
(0.213)
treated=1 x year=2012 -0.070
(0.225)
treated=1 x year=2014 -0.228
(0.236)
treated=1 x year=2016 -0.854***
(0.284)
Constant 63.304***
(0.127)
Observations 15194
Adjusted R-squared 0.055
Standard errors in parentheses * p<0.1 ** p<0.05 *** p<0.01
Note: Data are adjusted for individual fixed effects. Standard errors are clustered at the individual level. 2006 is the reference year.
Source: Author's calculation using HRS.
140

Table 3.4. Impact of Introduction of Exchanges on Planned Retirement: Difference-in-Differences
(1) (2) (3) (4) (5) (6) (7) (8) (9)
level of generosity
in lieu of 1(post)
pre-treatment
(2006-2008)
pre-treatment
(2010-2014)
staggered
DID
Y= prob. of
working at 65
ipw-
weighted
1(post) 1.027*** 1.193 § -0.169 § § 0.093 19.324 1.810
(0.097) (1.949) (.) (0.673) (.) (.) (0.150) (15.451) (2.149)
1(treated) -0.146 0.155 0.175 0.137 -0.161 0.178 0.079 -0.356 0.090
(0.137) (0.137) (0.138) (0.137) (0.263) (0.177) (0.150) (1.127) (0.147)
1(post) x 1(treated) -0.309** -0.306** -0.301** -0.521* 0.159 -0.358** -0.027 -1.970* -0.304**
(0.142) (0.140) (0.142) (0.291) (0.327) (0.141) (0.142) (1.138) (0.152)
covariates? N Y Y Y Y Y Y Y Y
state-specific year trends? N N Y N N N N N N
dependent variable mean 64.886 64.886 64.886 64.886 64.935 65.105 64.886 40.761 64.814
observations 15194 15194 15194 15194 8971 10943 15194 15194 15194
adjusted R-squared 0.023 0.060 0.063 0.059 0.093 0.030 0.059 0.020 0.061
Standard errors in parentheses * p<0.1 ** p<0.05 *** p<0.01
Source: Author's calculation using HRS.
main analysis
Note: Except for column (8), the outcome variable is the planned retirement age. Data are adjusted for state-, year-, and individual fixed effects. Standard errors are clustered at the individual
level. A possible confounding by the Medicaid expansion has been accounted for by adding the expansion identifier as a control. We observe retirement around 2014 when all states had their
141

Table 3.5. Impact of Introduction of Exchanges on Planned Retirement: Triple-Differences
(1) (2) (3) (4) (5) (6) (7) (8) (9)
level of generosity
in lieu of 1(post)
pre-treatment
(2006-2008)
pre-treatment
(2010-2014)
staggered
DDD
Y= prob. of
working at 65
ipw-
weighted
1(post) -0.866 § § -2.846* § § -0.031 § 0.035
(0.597) (.) (.) (1.635) (.) (.) (0.179) (.) (2.380)
1(treated) 0.110 0.209 0.235 0.175 2.819** 0.213 0.079 -0.931 0.169
(0.144) (0.145) (0.147) (0.146) (1.210) (0.177) (0.216) (1.179) (0.155)
ln(literacy) 0.051*** 0.017* 0.019** 0.018** 0.008 0.013 0.024** 0.239*** 0.017*
(0.006) (0.009) (0.009) (0.009) (0.184) (0.011) (0.011) (0.077) (0.010)
1(post) x 1(treated) 1.222 1.437 1.599* 2.573 § 1.270 0.020 1.754 1.494
(0.883) (0.875) (0.887) (1.857) (.) (0.863) (0.221) (7.431) (0.958)
1(post) x ln(literacy) 0.344** 0.304* 0.326 0.626* 0.444 0.299* -0.001 1.066 0.292
(0.137) (0.168) (0.205) (0.347) (0.420) (0.167) (0.019) (1.426) (0.182)
1(treated) x ln(literacy) 0.006 0.009 0.010 0.006 0.215** 0.016 0.001 -0.088 0.013
(0.009) (0.009) (0.009) (0.009) (0.087) (0.010) (0.015) (0.073) (0.009)
1(post) x 1(treated) x ln(literacy) -0.357* -0.413** -0.451** -0.731* -0.857** -0.392* -0.014 -0.671 -0.434*
(0.205) (0.203) (0.206) (0.437) (0.355) (0.200) (0.026) (1.694) (0.223)
covariates? N Y Y Y Y Y Y Y Y
state-specific year trends? N N Y N N N N N N
dependent variable mean 64.886 64.886 64.886 64.886 64.935 65.105 64.886 40.761 64.814
observations 15194 15194 15194 15194 8971 10943 15194 15194 15194
adjusted R-squared 0.041 0.061 0.065 0.060 0.095 0.032 0.059 0.021 0.063
Standard errors in parentheses * p<0.1 ** p<0.05 *** p<0.01
Source: Author's calculation using HRS.
main analysis
Note: Except for column (8), the outcome variable is the planned retirement age. Data are adjusted for state-, year-, and individual fixed effects. Standard errors are clustered at the individual
level. A possible confounding by the Medicaid expansion has been accounted for by adding the expansion identifier as a control. We observe retirement around 2014 when all states had first
142

3.7 Appendix

Table 3.1A. ACA Assister Program Regulations by State
States Marketplace Types Legislation Dates Medicaid Exp.*
Alabama Federally-facilitated Marketplace Official Announcement in Nov 2011 Not Adopted
Alaska Federally-facilitated Marketplace Official Announcement in July 2012 Adopted
Arizona Federally-facilitated Marketplace Official Announcement in Nov 2012 Adopted
Arkansas State-based Marketplace HB1508 in April 2013 Adopted
California State-based Marketplace AB1602, SB900 in Sept 2010 Adopted
Colorado State-based Marketplace SB11-200 in June 2011 Adopted
Connecticut State-based Marketplace SB921, PA11-53 in July 2011 Adopted
Delaware State-Partnership Marketplace Official Announcement in July 2013 Adopted
District of Columbia State-based Marketplace Act 19-269 in Jan 2012 Adopted
Florida Federally-facilitated Marketplace Official Announcement in Dec 2012 Not Adopted
Georgia Federally-facilitated Marketplace Official Announcement in Nov 2012 Not Adopted
Hawaii Federally-facilitated Marketplace SB1348 in July 2011 Adopted
Idaho State-based Marketplace HB248 in March 2013 Considering Exp.
Illinois State-Partnership Marketplace SB 1555 in 2011 Adopted
Indiana Federally-facilitated Marketplace Exec. Order 11-01 in late 2011 Adopted
Iowa State-Partnership Marketplace Official Announcement in Dec 2012 Adopted
Kansas Federally-facilitated Marketplace Official Announcement in Nov 2012 Not Adopted
Kentucky State-based Marketplace Order 578 in July 2012 Adopted
Louisiana Federally-facilitated Marketplace Official Announcement in 2012 Adopted
Maine Federally-facilitated Marketplace Official Announcement in Nov 2012 Adopted
Maryland State-based Marketplace SB 192, HB166 in April 2011 Adopted
Massachusetts State-based Marketplace Healthcare Reform in 2006 Adopted
Michigan State-Partnership Marketplace Official Announcement in March 2013 Adopted
Minnesota State-based Marketplace SF 1 in March 2013 Adopted
Mississippi Federally-facilitated Marketplace Official Announcement in Oct 2011 Not Adopted
Missouri Federally-facilitated Marketplace Official Announcement in Nov 2012 Not Adopted
Montana Federally-facilitated Marketplace Official Announcement in Dec 2012 Adopted
Nebraska Federally-facilitated Marketplace Official Announcement in Nov 2012 Considering Exp.
Nevada State-based Marketplace SB440 in June 2011 Adopted
New Hampshire State-Partnership Marketplace Official Announcement in March 2013 Adopted
New Jersey Federally-facilitated Marketplace Official Announcement in Dec 2012 Adopted
New Mexico State-based Marketplace CFDA 93-525 in March 2013 Adopted
New York State-based Marketplace Exec. Order 42 in April 2012 Adopted
North Carolina Federally-facilitated Marketplace HB126, HB115, SB418 in 2011 Not Adopted
North Dakota Federally-facilitated Marketplace Official Announcement in Nov 2012 Adopted
Ohio Federally-facilitated Marketplace Official Announcement in Nov 2012 Adopted
Oklahoma Federally-facilitated Marketplace Official Announcement in Nov 2012 Not Adopted
Oregon State-based Marketplace SB99 in June 2011 Adopted
Pennsylvania Federally-facilitated Marketplace Official Announcement in Dec 2012 Adopted
Rhode Island State-based Marketplace Exec. Order 11-09 in Sept. 2011 Adopted
South Carolina Federally-facilitated Marketplace Official Announcement in Nov 2012 Not Adopted
South Dakota Federally-facilitated Marketplace Official Announcement in Nov 2012 Not Adopted
Tennessee Federally-facilitated Marketplace Official Announcement in Dec 2012 Not Adopted
Texas Federally-facilitated Marketplace Official Announcement in July 2012 Not Adopted
Utah Federally-facilitated Marketplace Official Announcement in May 2013 Considering Exp.
Vermont State-based Marketplace HB202 in May 2011 Adopted
Virginia Federally-facilitated Marketplace Official Announcement in Dec 2012 Adopted
Washington State-based Marketplace SB5445 in May 2011 Adopted
West Virginia State-Partnership Marketplace SB408 in Feb 2013 Adopted
Wisconsin Federally-facilitated Marketplace Official Announcement in Nov 2012 Not Adopted
Wyoming Federally-facilitated Marketplace Official Announcement in Nov 2012 Not Adopted
Source: Data compiled through reviewing Marketplace documents and communication between the states and CCIIO by KFF.
Note: Data compiled through reviewing Marketplace documents and communication between the states and CCIIO by the Kaiser Family
Foundation. Regardless of the legislation dates, all states were required by federal law to launch the open enrollment from October 1, 2013 to
April 1, 2014.
143

Table 3.2A. Impact of Introduction of Exchanges on Planned Retirement: Difference-in-Differences, Detailed
(1) (2) (3) (4) (5) (6) (7) (8) (9)
Y=planned retirement age
without cov with cov
with cov, state-
specific year
trends
level of generosity
in lieu of 1(post)
pre-treatment
(2006-2008)
pre-treatment
(2010-2014)
staggered DID
Y= prob. of
working at 65
ipw-
weighted
1(post) 1.027*** 1.193 § -0.169 § § 0.093 19.324 1.810
(0.097) (1.949) (.) (0.673) (.) (.) (0.150) (15.451) (2.149)
1(treated) -0.146 0.155 0.175 0.137 -0.161 0.178 0.079 -0.356 0.090
(0.137) (0.137) (0.138) (0.137) (0.263) (0.177) (0.150) (1.127) (0.147)
1(post) x 1(treated) -0.309** -0.306** -0.301** -0.521* 0.159 -0.358** -0.027 -1.970* -0.304**
(0.142) (0.140) (0.142) (0.291) (0.327) (0.141) (0.142) (1.138) (0.152)
age - 0.124 0.136 0.123 0.017 0.154 0.122 -0.765 0.066
- (0.202) (0.201) (0.203) (0.219) (0.200) (0.203) (1.591) (0.223)
marital status (ref=1.married)
2. div/wid/sep - 0.271 0.271 0.275 0.405 0.370* 0.278 2.836 0.183
- (0.189) (0.187) (0.189) (0.330) (0.224) (0.189) (1.840) (0.196)
3. never married - 0.699 0.680 0.717 1.883 1.154* 0.740 -1.813 0.893
- (0.550) (0.552) (0.553) (1.545) (0.608) (0.553) (4.010) (0.589)
pension access (ref=4.none)
1. defined-benefit - -0.177 -0.176 -0.178 0.021 -0.330** -0.177 -0.221 -0.195
- (0.116) (0.115) (0.116) (0.206) (0.149) (0.116) (1.028) (0.124)
2. defined-contribution - -0.045 -0.049 -0.044 -0.058 -0.084 -0.044 -0.569 -0.135
- (0.088) (0.088) (0.088) (0.166) (0.106) (0.088) (0.696) (0.099)
3. other - 0.189 0.187 0.191 0.447 0.251 0.192 -2.293 0.049
- (0.194) (0.194) (0.194) (0.422) (0.220) (0.194) (1.545) (0.208)
ln(income) - 0.004 0.003 0.004 0.014 0.005 0.003 0.081* 0.003
- (0.006) (0.006) (0.006) (0.011) (0.008) (0.006) (0.047) (0.007)
ln(wealth) - 0.008 0.009 0.008 0.008 0.009 0.008 0.002 0.008
- (0.008) (0.008) (0.008) (0.014) (0.011) (0.008) (0.066) (0.009)
health - 0.362** 0.337** 0.362** 0.139 0.484*** 0.365** 2.485** 0.411***
- (0.142) (0.140) (0.142) (0.233) (0.183) (0.142) (1.228) (0.155)
no. of dependents - -0.012 -0.014 -0.013 -0.008 -0.009 -0.012 -0.369* -0.014
- (0.011) (0.011) (0.011) (0.017) (0.009) (0.010) (0.220) (0.016)
access to other insurance - -0.116 -0.107 -0.112 -0.274 -0.166 -0.109 -2.105 -0.176
- (0.237) (0.236) (0.236) (0.393) (0.301) (0.236) (1.880) (0.271)
access to Social Security income - 1.406*** 1.427*** 1.409*** 0.972** 1.005*** 1.404*** -5.371** 1.380***
- (0.297) (0.302) (0.298) (0.468) (0.348) (0.297) (2.138) (0.302)
Medicaid expansion - 6.791*** -813.065** 6.791*** -11.069*** -0.368 4.707** 53.397* 5.165**
- (2.331) (384.088) (2.321) (3.024) (0.816) (1.925) (29.339) (2.040)
constant 64.689*** 57.777*** -17.895 57.857*** 63.630*** 51.452*** 57.984*** 58.612 60.443***
(0.073) (10.672) (404.070) (10.676) (11.678) (11.056) (10.694) (83.704) (11.715)
dependent variable mean 64.886 64.886 64.886 64.886 64.935 65.105 64.886 40.761 64.814
observations 15194 15194 15194 15194 8971 10943 15194 15194 15194
adjusted R-squared 0.023 0.060 0.063 0.059 0.093 0.030 0.059 0.020 0.061
Standard errors in parentheses * p<0.1 ** p<0.05 *** p<0.01
Source: Author's calculation using HRS.
Note: Data are adjusted for state-, year-, and individual fixed effects. Standard errors are clustered at the individual level. A possible confounding by the Medicaid expansion has been accounted for by
adding the expansion identifier as a control. We observe retirement around 2014 when all states had their first open enrollments except in column (7). Cells with § are omitted due to collinearity.
main analysis robustness checks, with cov
144

Table 3.3A. Impact of Introduction of Exchanges on Planned Retirement: Triple-Differences, Detailed
(1) (2) (3) (4) (5) (6) (7) (8) (9)
Y=planned retirement age
without cov with cov
with cov, state-
specific year trends
level of generosity
in lieu of 1(post)
pre-treatment
(2006-2008)
pre-treatment
(2010-2014)
staggered
DID
Y= prob. of
working at 65
ipw-
weighted
1(post) -0.866 § § -2.846* § § -0.031 § 0.035
(0.597) (.) (.) (1.635) (.) (.) (0.179) (.) (2.380)
1(treated) 0.110 0.209 0.235 0.175 2.819** 0.213 0.079 -0.931 0.169
(0.144) (0.145) (0.147) (0.146) (1.210) (0.177) (0.216) (1.179) (0.155)
ln(literacy) 0.051*** 0.017* 0.019** 0.018** 0.008 0.013 0.024** 0.239*** 0.017*
(0.006) (0.009) (0.009) (0.009) (0.184) (0.011) (0.011) (0.077) (0.010)
1(post) x 1(treated) 1.222 1.437 1.599* 2.573 § 1.270 0.020 1.754 1.494
(0.883) (0.875) (0.887) (1.857) (.) (0.863) (0.221) (7.431) (0.958)
1(post) x ln(literacy) 0.344** 0.304* 0.326 0.626* 0.444 0.299* -0.001 1.066 0.292
(0.137) (0.168) (0.205) (0.347) (0.420) (0.167) (0.019) (1.426) (0.182)
1(treated) x ln(literacy) 0.006 0.009 0.010 0.006 0.215** 0.016 0.001 -0.088 0.013
(0.009) (0.009) (0.009) (0.009) (0.087) (0.010) (0.015) (0.073) (0.009)
1(post) x 1(treated) x ln(literacy) -0.357* -0.413** -0.451** -0.731* -0.857** -0.392* -0.014 -0.671 -0.434*
(0.205) (0.203) (0.206) (0.437) (0.355) (0.200) (0.026) (1.694) (0.223)
age - 0.130 0.146 0.124 0.024 0.159 0.127 -0.735 0.075
- (0.202) (0.201) (0.202) (0.218) (0.200) (0.203) (1.591) (0.223)
marital status (ref=1.married)
2. div/wid/sep - 0.266 0.271 0.270 0.394 0.373* 0.274 2.769 0.183
- (0.189) (0.187) (0.189) (0.328) (0.224) (0.189) (1.835) (0.196)
3. never married - 0.688 0.671 0.703 1.836 1.145* 0.728 -1.898 0.872
- (0.549) (0.551) (0.552) (1.539) (0.607) (0.555) (4.017) (0.590)
pension access (ref=4.none)
1. defined-benefit - -0.184 -0.185 -0.183 0.014 -0.343** -0.177 -0.203 -0.200
- (0.116) (0.116) (0.116) (0.206) (0.149) (0.116) (1.027) (0.125)
2. defined-contribution - -0.042 -0.045 -0.042 -0.062 -0.083 -0.040 -0.536 -0.127
- (0.088) (0.088) (0.088) (0.167) (0.106) (0.088) (0.695) (0.099)
3. other - 0.176 0.173 0.180 0.434 0.227 0.188 -2.358 0.040
- (0.194) (0.194) (0.194) (0.421) (0.220) (0.194) (1.540) (0.208)
ln(income) - 0.003 0.003 0.003 0.014 0.004 0.003 0.077 0.002
- (0.006) (0.006) (0.006) (0.011) (0.008) (0.006) (0.047) (0.007)
ln(wealth) - 0.008 0.009 0.008 0.008 0.010 0.008 0.005 0.009
- (0.008) (0.008) (0.008) (0.014) (0.011) (0.008) (0.066) (0.009)
health - 0.363** 0.340** 0.361** 0.136 0.483*** 0.367*** 2.473** 0.417***
- (0.142) (0.140) (0.142) (0.233) (0.182) (0.142) (1.227) (0.156)
no. of dependents - -0.010 -0.012 -0.012 -0.004 -0.007 -0.011 -0.362 -0.011
- (0.011) (0.011) (0.011) (0.018) (0.009) (0.011) (0.220) (0.016)
access to other insurance - -0.102 -0.092 -0.098 -0.269 -0.148 -0.104 -2.109 -0.160
- (0.236) (0.236) (0.236) (0.391) (0.301) (0.237) (1.883) (0.271)
access to Social Security income - 1.394*** 1.414*** 1.395*** 0.977** 0.983*** 1.386*** -5.505** 1.366***
- (0.299) (0.304) (0.300) (0.470) (0.350) (0.298) (2.144) (0.305)
Medicaid expansion - 5.142*** -795.277** 6.812*** -11.201*** -2.415 6.867*** 58.000*** 7.633***
- (1.985) (379.852) (2.319) (3.038) (2.789) (2.316) (15.017) (2.585)
constant 64.875*** 57.836*** 373.015 58.107*** 63.289*** 51.747*** 57.984*** 60.337 60.215***
(0.077) (10.677) (454.417) (10.672) (13.827) (11.073) (10.695) (83.685) (11.732)
dependent variable mean 64.886 64.886 64.886 64.886 64.935 65.105 64.886 40.761 64.814
observations 15194 15194 15194 15194 8971 10943 15194 15194 15194
Adjusted R-squared 0.041 0.061 0.065 0.060 0.095 0.032 0.059 0.021 0.063
Standard errors in parentheses * p<0.1 ** p<0.05 *** p<0.01
Source: Author's calculation using HRS.
main analysis robustness checks, with cov
Note: Data are adjusted for state-, year-, and individual fixed effects. Standard errors are clustered at the individual level. A possible confounding by the Medicaid expansion has been accounted for by
adding the expansion identifier as a control. We observe retirement around 2014 when all states had first open enrollments except in column (7). Cells with § are omitted due to collinearity.
145

Bibliography

Abaluck, Jason, and Jonathan Gruber. "Evolving choice inconsistencies in choice of prescription
drug insurance." American Economic Review 106, 8 (2016): 2145-84.
.
Acemoglu, Daron. "Changes in unemployment and wage inequality: an alternative theory and
some evidence." American Economic Review 89, 5 (1999): 1259-1278
.
Acemoglu, Daron, and David Autor. "Skills, tasks and technologies: Implications for
employment and earnings." Handbook of labor economics 4 (2011): 1043-1171

Acemoglu, Daron, and Pascual Restrepo. "The race between man and machine: Implications of
technology for growth, factor shares, and employment." American Economic Review 108, 6
(2018): 1488-1542

Acemoglu, Daron, and Fabrizio Zilibotti. "Productivity differences." The Quarterly Journal of
Economics 116, 2 (2001): 563-606

Adams, Kathryn B., Sheryl Sanders, and E. A. Auth. "Loneliness and depression in independent
living retirement communities: risk and resilience factors." Aging & Mental Health 8, 6
(2004): 475-485.

Aguila, Emma. "Male labor force participation and social security in Mexico." Journal of
Pension Economics & Finance 13, 2 (2014): 145-171

Allen, Steven G., Robert L. Clark, and Linda S. Ghent. "Phasing into retirement." ILR Review
58, 1 (2004): 112-127

Anderson, Gerard F., and Peter Sotir Hussey. "Population aging: a comparison among
industrialized countries." Health Affairs 19, 3 (2000): 191-203.

146

Anderson, Gerard F., Uwe E. Reinhardt, Peter S. Hussey, and Varduhi Petrosyan. "It’s the prices,
stupid: why the United States is so different from other countries." Health Affairs 22, 3
(2003): 89-105.

Ayyagari, Padmaja. "Health insurance and early retirement plans: Evidence from the Affordable
Care Act." Manuscript, Department of Health Management and Policy, University of Iowa.

Autor, David H. "Outsourcing at will: The contribution of unjust dismissal doctrine to the growth
of employment outsourcing." Journal of Labor Economics 21, 1 (2003): 1-42.

Autor, David H., Frank Levy, and Richard J. Murnane. "The skill content of recent technological
change: An empirical exploration." The Quarterly Journal of Economics 118, 4 (2003):
1279-1333.

Autor, David, Lawrence F. Katz, and M. Kearney. "Trends in US wage inequality: Revising the
revisionists." The Review of Economics and Statistics 90, 2 (2008): 300-323

Autor, David H., Lawrence F. Katz, and Alan B. Krueger. "Computing inequality: have
computers changed the labor market?." The Quarterly Journal of Economics 113, 4 (1998):
1169-1213.

Bartel, Ann P., and Nachum Sicherman. "Technological change and retirement decisions of older
workers." Journal of Labor Economics 11, 1, Part 1 (1993): 162-183.

Berkovec, James, and Steven Stern. "Job exit behavior of older men." Econometrica 59, 1
(1991): 189.

Bhargava, Saurabh, George Loewenstein, and Justin Sydnor. “Do individuals make sensible
health insurance decisions? Evidence from a menu with dominated options.” No. w21160.
National Bureau of Economic Research, 2015.

Bibby, Courtney L. "Should I stay or should I leave? Perceptions of age discrimination,
organizational justice, and employee attitudes on intentions to leave." Journal of Applied
Management and Entrepreneurship 13, no. 2 (2008): 63.

147

Blau, David M., and Donna B. Gilleskie. "Health insurance and retirement of married couples."
Journal of Applied Econometrics 21, 7 (2006): 935-953.

Blazer, Dan G. "Self-efficacy and depression in late life: a primary prevention proposal." Aging
& Mental Health 6, 4 (2002): 315-324.

Blundell, Richard, and Thomas MaCurdy. "Labor supply: A review of alternative approaches."
Handbook of Labor Economics 3 (1999): 1559-1695.

Borghans, Lex, Angela Lee Duckworth, James J. Heckman, and Bas Ter Weel. "The economics
and psychology of personality traits." Journal of human Resources 43, 4 (2008): 972-1059.

Briggs, D. C. “Causal inference and the Heckman model.” Journal of Educational and
Behavioral Statistics, 29, 4 (2004): 397-420.

Brynjolfsson, Erik, and Andrew McAfee. The second machine age: Work, progress, and
prosperity in a time of brilliant technologies. WW Norton & Company, 2014.

Cable, Daniel M., and Timothy A. Judge. "Person–organization fit, job choice decisions, and
organizational entry." Organizational Behavior and Human Decision Processes 67, 3 (1996):
294-311.

Cahill, Kevin E., Michael D. Giandrea, and Joseph F. Quinn. "The Role of Re-entry in the
Retirement Process." US Bureau of Labor Statistics Working Paper 439 (2010). Working
Paper.
———"Reentering the labor force after retirement." Monthly Labor Review 134, 6 (2011): 34-
42.
——— "Retirement patterns and the macroeconomy, 1992–2010: The prevalence and
determinants of bridge jobs, phased retirement, and reentry among three recent cohorts of
older Americans." The Gerontologist 55, 3 (2015): 384-403.

Calvo, Esteban, Kelly Haverstick, and Steven A. Sass. "Gradual retirement, sense of control, and
retirees' happiness." Research on Aging 31, 1 (2009): 112-135.

Caplan, Robert D., and R. Van Harrison. "Person‐environment fit theory: some history, recent
developments, and future directions." Journal of Social Issues 49, 4 (1993): 253-275.

148

Card, David, and John E. DiNardo. "Skill-biased technological change and rising wage
inequality: some problems and puzzles." Journal of Labor Economics 20, 4 (2002): 733-783.

Card, David, and Thomas Lemieux. "Can falling supply explain the rising return to college for
younger men? A cohort-based analysis." The Quarterly Journal of Economics 116, 2 (2001):
705-746.

Claxton, G., B. DiJulio, B. Finder, and J. Lundy. Employer Health Benefits: 2010 Annual
Survey. The Henry J. Kaiser Family Foundation, 2010.

Coile, Courtney C., and Phillip B. Levine. "Bulls, bears, and retirement behavior." ILR Review
59, 3 (2006): 408-429.

Coile, Courtney, and Jonathan Gruber. “Future Social Security Entitlements and the Retirement
Decision.” The Review of Economics and Statistics 89, 2(2007): 234–246.

Costa Jr, Paul T., and Robert R. McCrae. "Longitudinal stability of adult personality." In
Handbook of Personality Psychology, pp. 269-290. Academic Press, 1997.

Costa Jr, Paul T., Jeffrey H. Herbst, Robert R. McCrae, and Ilene C. Siegler. "Personality at
midlife: Stability, intrinsic maturation, and response to life events." Assessment 7, 4 (2000):
365-378.

Cutler, David M., and Richard J. Zeckhauser. "The anatomy of health insurance." In Handbook
of Health Economics vol 1, pp. 563-643. Elsevier, 2000.

Currie, Janet, Brigitte C. Madrian, Orley Ashenfelter, and David Card. "Handbook of Labor
Economics." Handbook of Labor Economics (1999): 3309-3416.

De Meza, David, and David C. Webb. "Advantageous selection in insurance markets." RAND
Journal of Economics (2001): 249-262.

Deming, David J. "The growing importance of social skills in the labor market." The Quarterly
Journal of Economics, 132, 4 (2017): 1593-1640.

149

Dillender, Marcus, Carolyn Heinrich, and Susan N. Houseman. "Effects of the Affordable Care
Act on part-time employment: early evidence." Working Paper (2016).

Ensher, Ellen A., Elisa J. Grant‐Vallone, and Stewart I. Donaldson. "Effects of perceived
discrimination on job satisfaction, organizational commitment, organizational citizenship
behavior, and grievances." Human resource development quarterly 12, no. 1 (2001): 53-72.

Eriksson, Tor, and Nicolai Kristensen. "Wages or fringes? Some evidence on trade-offs and
sorting." Journal of Labor Economics 32, 4 (2014): 899-928.

Feldstein, Martin. "Structural reform of social security." Journal of Economic Perspectives 19, 2
(2005): 33-55.

Fernandez, Roberto M. "Skill-biased technological change and wage inequality: Evidence from a
plant retooling." American Journal of Sociology 107, 2 (2001): 273-320.

Fillenbaum, G. G., D. C. Hughes, A. Heyman, L. K. George, and D. G. Blazer. "Relationship of
health and demographic characteristics to Mini-Mental State Examination score among
community residents." Psychological Medicine 18, 3 (1988): 719-726.

French, Eric, and John Bailey Jones. "The Effects of Health Insurance and Self‐Insurance on
Retirement Behavior." Econometrica 79, 3 (2011): 693-732.

Frey, C. B., & Osborne, M. A. “The future of employment: how susceptible are jobs to
computerisation?” Technological Forecasting and Social Change, 114 (2017): 254-280.

Gallen, Trevor S., and Casey B. Mulligan. Wedges, Labor Market Behavior, and Health
Insurance Coverage under the Affordable Care Act. No. w19770. National Bureau of
Economic Research, 2013.

Galor, Oded, and Omer Moav. "Ability-biased technological transition, wage inequality, and
economic growth." The Quarterly Journal of Economics 115, 2 (2000): 469-497.

Garber, Alan M., and Jonathan Skinner. "Is American health care uniquely inefficient?." Journal
of Economic Perspectives 22, 4 (2008): 27-50.

150

Goldin, C., and L. F. Katz. “The Origins of Technology-Skill Complementarity.” The Quarterly
Journal of Economics 113, 3 (1998): 693–732.
—"The race between technology and education." Cambridge, MA: Harvard (2008).

Graebner, W. A history of retirement: The meaning and function of an American institution,
1885-1978. New Haven, Conn.: Yale University Press, 1980.

Greene, William H. "Sample selection bias as a specification error: A comment." Econometrica:
Journal of the Econometric Society (1981): 795-798.

Griffin, Barbara, Piers Bayl-Smith, and Beryl Hesketh. "The longitudinal effects of perceived
age discrimination on the job satisfaction and work withdrawal of older employees." Work,
Aging and Retirement 2, no. 4 (2016): 415-427.

Gruber, Jonathan. "Health insurance and the labor market." Handbook of Health Economics, vol.
1, pp. 645-706. Elsevier, 2000.

Gruber, Jonathan, and Brigitte C. Madrian. Health insurance, labor supply, and job mobility: a
critical review of the literature. No. w8817. National Bureau of Economic Research, 2002.

Gruber, Jonathan; Wise, David A., editors. Social Security and Retirement around the World.
NBER, The University of Chicago Press; 1999.

Gustman, Alan, Thomas Steinmeier, and Nahid Tabatabai. Distributional Effects of Means
Testing Social Security: Income Versus Wealth. No. w22424. National Bureau of Economic
Research, 2016. Working Paper.
—"The Affordable Care Act as retiree health insurance: implications for retirement and Social
Security claiming." Journal of Pension Economics & Finance 18, 3 (2019): 415-449.

Handel, Benjamin R. "Adverse selection and inertia in health insurance markets: When nudging
hurts." American Economic Review 103, 7 (2013): 2643-82.

Hayward, Mark D., Melissa A. Hardy, and Mei-Chun Liu. "Work after retirement: The
experiences of older men in the United States." Social Science Research 23, 1 (1994): 82-
107.

151

Heckman, James J., and Tim Kautz. "Hard evidence on soft skills." Labour Economics 19, 4
(2012): 451-464.

Imbens, Guido W., and Donald B. Rubin. Causal inference in statistics, social, and biomedical
sciences. Cambridge University Press, 2015.

Johnson, Richard W., Amy J. Davidoff, and Kevin Perese. "Health insurance costs and early
retirement decisions." ILR Review 56, 4 (2003): 716-729.

Katz, Lawrence F., and Kevin M. Murphy. "Changes in relative wages, 1963–1987: supply and
demand factors." The Quarterly Journal of Economics 107, 1 (1992): 35-78.

Kramer, William E. "Large employers see scenarios under which they could move workers and
retirees to exchanges." Health Affairs 31, 2 (2012): 299-305.

LaLonde, Robert J. "Evaluating the econometric evaluations of training programs with
experimental data." The American economic review (1986): 604-620.

Lee, Ronald D. The fiscal impacts of population change." Center for the Economics and
Demography of Aging (2001). Working Paper.

Lee, Ronald, and Jonathan Skinner. "Will aging baby boomers bust the federal budget?." Journal
of Economic Perspectives 13, 1 (1999): 117-140.

Levy, Helen, Thomas Buchmueller, and Sayeh Nikpay. "The Effect of Health Reform on
Retirement." Ann Arbor, MI. University of Michigan Retirement Research Center (MRRC)
Working Paper, WP 329 (2015): 4.

Lewbel, Arthur. "Simple endogenous binary choice and selection panel model estimators."
Working Papers in Economics (2005): 70.

Lindenberger, Ulman, Ulrich Mayr, and Reinhold Kliegl. "Speed and intelligence in old age."
Psychology and Aging 8, 2 (1993): 207.

152

Lusardi, Annamaria, and Olivia S. Mitchell. "Baby boomer retirement security: The roles of
planning, financial literacy, and housing wealth." Journal of Monetary Economics 54, 1
(2007): 205-224.

MacCrory, Frank, George Westerman, Yousef Alhammadi, and Erik Brynjolfsson. "Racing with
and against the machine: changes in occupational skill composition in an era of rapid
technological advance." (2014). Working Paper.

Macdonald, Jamie L., and Sheri R. Levy. "Ageism in the workplace: The role of psychosocial
factors in predicting job satisfaction, commitment, and engagement." Journal of Social Issues
72, no. 1 (2016): 169-190.

Madrian, Brigitte C. "Employment-based health insurance and job mobility: Is there evidence of
job-lock?." The Quarterly Journal of Economics 109, 1 (1994): 27-54.

Maestas, N. “Back to work expectations and realizations of work after retirement.” Journal of
Human Resources, 45, 3(2010): 718-748.

Maestas, Nicole, and Julie Zissimopoulos. "How longer work lives ease the crunch of population
aging." Journal of Economic Perspectives 24, 1 (2010): 139-60.

McCrae, Robert R., and Oliver P. John. "An introduction to the five‐factor model and its
applications." Journal of Personality 60, 2 (1992): 175-215.

McKinsey & Company. McKinsey Global Institute: A Future that Works: Automation,
Employment, and Productivity. January 2017.

Medicaid and CHIP (MAC). “Marketplace consumer assistance: approaches to program
sustainability.” In Federally Facilitated Marketplaces Eligibility and Enrollment Learning
Collaborative, January 2015. Retrieved from https://www.medicaid.gov/state-resource-
center/mac-learning-collaboratives/downloads/cnsmr-asstnce-mktplce.pdf

Myck, Michal; Reed, Howard (2006) : A review of static and dynamic models of labour supply
and labour market transitions, IFS Working Papers, Institute for Fiscal Studies (IFS).

153

Myerson, Rebecca. "Health Insurance Literacy and Health Insurance Choices: Evidence from
Affordable Care Act Navigator Programs." Working Paper (2018).

Moses, Hamilton, David HM Matheson, E. Ray Dorsey, Benjamin P. George, David Sadoff, and
Satoshi Yoshimura. "The anatomy of health care in the United States." JAMA 310, 18 (2013):
1947-1964.

Munnell, Alicia H., and Matthew S. Rutledge. "The effects of the Great Recession on the
retirement security of older workers." The ANNALS of the American Academy of Political
and Social Science 650, 1 (2013): 124-142.

Nyce, S., Schieber, S. J., Shoven, J. B., Slavov, S. N., & Wise, D. A. (2013). Does retiree health
insurance encourage early retirement?. Journal of Public Economics, 104: 40-51.

Petkoska, Jasmina, and Joanne K. Earl. "Understanding the influence of demographic and
psychological variables on retirement planning." Psychology and aging 24, no. 1 (2009): 245.

Pierce, Brooks. "Compensation inequality." Quarterly Journal of Economics 116, 4 (2001):
1493-1525.

Pollitz, K. J. Tolbert, and R. Ma. Survey of health insurance marketplace assister programs.
Kaiser Family Foundation, 2014.
—Survey of health insurance marketplace assister programs and brokers. Kaiser Family
Foundation, 2015.

Pollitz, K., J. Tolbert, and A. Semanskee. Survey of health insurance marketplace assister
programs and brokers. Kaiser Family Foundation, 2016.

Radloff, Lenore Sawyer. "The CES-D scale: a self-report depression scale for research in the
general population." Applied Psychological Measurement 1, 3 (1977): 385-401.

Reitzes, Donald C., Elizabeth J. Mutran, and Maria E. Fernandez. "Does retirement hurt weil-
being? Factors influencing self-esteem and depression among retires and workers." The
Gerontologist 36, 5 (1996): 649-656.

154

Robinson, C., & Clark, R. “Retiree health insurance and disengagement from a career job.”
Journal of Labor Research 31, 3 (2010): 247-262.

Rosen, Sherwin. "The theory of equalizing differences." Handbook of Labor Economics, vol 1
(1986): 641-692.

Rotman, David. (2013). "How technology is destroying jobs." Technology Review 16(4), 28-35.

Rust, John, and Christopher Phelan. "How social security and medicare affect retirement
behavior in a world of incomplete markets." Econometrica: Journal of the Econometric
Society (1997): 781-831.

Salthouse, T.A. (1999). Pressing issues in cognitive aging. In N. Schwarz, D.C. Park, B.
Knauper, & S. Sudman (Eds.), Cognition, aging, and self-reports. Philadelphia: Psychology
Press.

Sharit, Joseph, and Sara J. Czaja. "Ageing, computer-based task performance, and stress: issues
and challenges." Ergonomics 37, 4 (1994): 559-577.

Segel-Karpas, Dikla, Liat Ayalon, and Margie E. Lachman. "Loneliness and depressive
symptoms: The moderating role of the transition into retirement." Aging & mental health 22,
no. 1 (2018): 135-140.

Semykina, Anastasia, and Jeffrey M. Wooldridge. "Estimating panel data models in the presence
of endogeneity and selection." Journal of Econometrics 157, 2 (2010): 375-380.

Shoven, John B., and Sita Nataraj Slavov. (2014). “The role of retiree health insurance in the
early retirement of public sector employees.” Journal of Health Economics, 38: 99-108.
—"Does it pay to delay social security?." Journal of Pension Economics & Finance 13, 2
(2014): 121-144.

Sood, Neeraj, Arkadipta Ghosh, and José J. Escarce. "Employer‐Sponsored Insurance, Health
Care Cost Growth, and the Economic Performance of US Industries." Health Services
Research 44, no. 5p1 (2009): 1449-1464.

155

Stevenson, Betsey, and Justin Wolfers. "Bargaining in the shadow of the law: Divorce laws and
family distress." The Quarterly Journal of Economics 121, 1 (2006): 267-288.

Van Solinge, Hanna. "Health change in retirement: a longitudinal study among older workers in
the Netherlands." Research on Aging 29, 3 (2007): 225-256.

Vella, Francis. "Estimating models with sample selection bias: a survey." Journal of Human
Resources (1998): 127-169.

Welch, Finis. "Black-white differences in returns to schooling." The American Economic Review
(1973): 893-907.

Winship, Christopher, and Larry Radbill. "Sampling weights and regression analysis."
Sociological Methods & Research 23, 2 (1994): 230-257.

Abstract (if available)

Abstract In a world where most nations are undergoing rapid population aging and experiencing a shrinking tax base to provide retirees with a sufficient level of public health insurance- and pension- plans, one of the feasible solutions to deal the lack of financial base for retirees is to promote longer working-lives. Using data from the United States, this dissertation explores how the automation of labor and a recent healthcare reform changing the system of health insurance provision pose opportunities and challenges to promoting longer working-lives by influencing workers’ timing and form of retirement. The first chapter asks what encourages aging workers to continue working and seeks for answers from workers’ heterogeneous experiences of the process of automation. Drawing data from the Health and Retirement Study (HRS) and O*NET, we assess the consequence of automation on the retirement behaviors using various panel estimations including fixed-effect, logit, and multinomial logit models. Using the same data for years 1992-2014, the second chapter explores retirees’ late-career job sorting behaviors as they engage in unretirement. Specifically, we assess whether and how much individuals’ automation-resistance plays a role in determining unretirement likelihood—and if unretiring—what types of jobs the individuals choose, and whether the act of ‘working in old age’ can make the them financially and psychologically better off. Lastly, in light of the strong linkage between health insurance provision and employment in the US healthcare system, the third chapter explores whether the healthcare policy reform alters aging individuals’ retirement timing. In doing so, we study the effect of the ACA on workers retirement decisions while accounting for their heterogeneous levels of insurance literacy. After finding that ACA accelerates retirement by removing job-locks induced by employer-sponsored health insurance, we close by discussing that, from the policy perspective, the ACA-induced hastening of retirement could potentially counteract past policy efforts to foster longer working lives.

Linked assets

University of Southern California Dissertations and Theses

Conceptually similar

PDF

Three essays on health, aging, and retirement

PDF

More knowledge, better plans: a study of heterogeneity in dementia prevalence and mid-life cognitive changes

PDF

Essays on the use of microsimulation for health and economic policy analysis

PDF

Health care utilization and spending of the U.S. aging population

PDF

Three essays on estimating the effects of government programs and policies on health care among disadvantaged population

PDF

Essays on the politics of the American welfare state

PDF

Three essays on the evaluation of long-term care insurance policies

PDF

The role of public policy in the decisions of parents and caregivers: an examination of work, fertility, and informal caregiving

PDF

Essays on Japanese macroeconomy

PDF

Essays on the U.S. market for substance use treatment and the impact of Medicaid policy reform

PDF

Essays on improving human interactions with humans, algorithms, and technologies for better healthcare outcomes

PDF

Impacts of caregiving on wellbeing among older adults and their spousal caregivers in the United States

PDF

Essays on wellbeing disparities in the United States and their social determinants

PDF

Essays on information design for online retailers and social networks

Asset Metadata

Creator Lee, Zeewan (author)

Core Title Essays on work, retirement, and fostering longer working lives

School School of Policy, Planning and Development

Degree Doctor of Philosophy

Degree Program Public Policy and Management

Publication Date 04/20/2020

Defense Date 04/17/2020

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

Tag Affordable Care Act,aging,automation,health insurance,insurance literacy,labor supply,OAI-PMH Harvest,Retirement,skills,tasks,technological changes

Language English

Contributor Electronically uploaded by the author (provenance)

Advisor Aguila, Emma (committee chair), Chen, Alice (committee member), Sood, Neeraj (committee member), Zissimopoulos, Julie (committee member)

Creator Email lee.zeewan@gmail.com,zeewanle@usc.edu

Permanent Link (DOI) https://doi.org/10.25549/usctheses-c89-281425

Unique identifier UC11674977

Identifier etd-LeeZeewan-8264.pdf (filename),usctheses-c89-281425 (legacy record id)

Legacy Identifier etd-LeeZeewan-8264.pdf

Dmrecord 281425

Document Type Dissertation

Rights Lee, Zeewan

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