.ﬁmwnmn

If. . .

A
3 . A}; I
a .
awn”: $hﬁ1..3:l&3
.cwl t .-
1... . ‘ .Z
. a. Ru... 1 .3 5..

9.5.». 5......

a... I.
4%»...
an
5....
.
. new
£2.
1 .

5%» 3

kid; a
. Pm».

.. new“; . a.

. ‘1.-
. 5. bi.+t..uk.wo..>aw to.”
304...... . . . . [9.56-3.39 unrivuﬁh 4G”)! . . 4.0.0:. 3
-2 _ . . Lawn“? .n u . . , .525

.o >3: . .
a
if. .ﬁv‘ilx: {viii
.éil’taba‘“.r!iu
it

1‘... SF. I'rrltx.
.. ..
..¢€:L...l«l«n§ i3... .
siéii‘:~nn 'I‘Shbu!
. .u. a... 15.2... 1
13.32.5331... I.

., t. (2..

 

. . 1131A)
.. 1. . . .. .L......u.u... .1. .
..3 1...-.. . . .J r .. ..
3. £25...» :3... 17.5321:

 

 

. A». .1! 2'! - O6!
-1. . {1.3.1. . . 71...... 6 y.
If"?! ierl..: 1. , .. . .

 

Illlllﬂlllllllllllllllllllllllllllllllllllllllllillllllllllll

31293 02048 5896

 

LIBRARY

Michigan State
University

 

 

 

This is to certify that the

dissertation entitled
THREE ESSAYS ON THE ECONOMICS OF AGING
presented by
Scott J. Adams

has been accepted towards fulﬁllment \
of the requirements for

Ph . D . degree in Economics

L

V Major professor

Date ?l/Z j/Zood

MS U is an Afﬁrmative Action/Equal Opportunity Institution 0-12771

 

 

 

PLACE IN RETURN BOX to remove this checkout from your record.
TO AVOID FINES return on or before date due.
MAY BE RECALLED with earlier due date if requested.

 

 

 

DAEDUE DATE DUE DATE DUE
NOV 2 3 wuq
JAN 0212002 07 2004
H

 

 

 

 

 

 

 

 

 

 

 

 

 

mm W.“

THREE ESSAYS ON THE ECONOMICS OF AGING
By

Scott J. Adams

A DISSERTATION
Submitted to
Michigan State University

in partial fulﬁllment of the requirements
for the degree of

DOCTOR OF PHILOSOPHY
Department of Economics

2000

ABSTRACT
THREE ESSAYS ON THE ECONOMICS OF AGING
By

Scott J. Adams

An aging US. population creates many challenges for policy makers, and, thus,
renders research that sheds light on the conditions of older Americans of great
importance. My aim is to contribute to this research by using economic analysis to
address three very important questions: What is the effect of educational attainment on
the health of older Americans? Do ﬁrms that use age as a discriminatory criterion in
promotion decisions adversely affect the labor market outcomes of their older workers?
Finally, does legislation aimed at combating age discrimination change ﬁrm behavior in
employing older workers?

Chapter 1 is an attempt to answer the ﬁrst question. The missing element of
investigations of the education-health relationship is an attempt to identify the causal
effect of education. Studies that have shown the correlation between education and
health fail to account fully for the fact that education and health are jointly determined.
Thus, as long as there are unobserved factors that inﬂuence both the attainment of more
education and the health status of adults, then it is certainly possible that the correlation
between education and health is spurious. This paper uses quarter of birth as a source of
exogenous variation in education to Show that there exists a causal component of the

effect of education on health.

Chapter 2 uses the Health and Retirement Study to look at the effect of one type
of age discrimination-ﬁrm preferences in promotion toward younger workers. While the
effect of preferences in promotion is an interesting question in itself, for the purposes of
this paper it serves as an indicator of ﬁrm discrimination in general. First, the paper
veriﬁes whether individuals reporting age discrimination are paid less than similarly
trained and qualiﬁed workers. Second, the paper investigates how the self-assessed
probabilities of early retirement are affected by ﬁrm discriminatory practices. Finally,
workers are tracked to determine whether they ultimately are more likely to experience
lower wage growth, to separate from their employer, or to retire early, as a function of
their ﬁrm’s propensity to promote younger workers. The ﬁndings show that those
reporting discrimination do in fact experience some negative effects in the labor market.
The most signiﬁcant negative effects are on wage growth. Yet, the evidence also
suggests that discrimination may affect individual assessments of the probability of
retiring early. The latter effect is quite important when viewed in a public policy context.
If the aged are driven into early retirement due to ﬁrm discriminatory practices, this has
implications for social security ﬁnancing.

Chapter 3 concludes the dissertation by studying the effects of state age
discrimination legislation on employment outcomes for older individuals. Speciﬁcally, it
is shown that legislation boosts the probability of employment for older workers. The
probability of being a new hire, however, falls. These effects are largely due to more

older workers being retained at their jobs after the passage of the legislation.

ACKNOWLEGEMENTS

David Neumark, my advisor, was extremely patient as he provided invaluable
guidance over the past several years. The dissertation would not have been possible
without his input. Moreover, the time that I spent assisting him with his research taught
me a great deal.

Other faculty members at Michigan State played important roles in my graduaute
education. John Strauss offered numerous suggestions that helped improve the
dissertation, as well as provided much advice throughout the past several years for which
I am extremely grateful. I also owe Jeff Wooldridge, John Goddeeris, Jeff Biddle, and
Stephen Woodbury special recognition for their contributions. I am also indebted to
many of my fellow graduate students for their support. I would particularly like to thank
Marianne Johnson, Vinit Jagdish, Doug Harris, Brian Hannon, John Francis, and Tom
Davis.

As for my family, the true extent of my gratitude is impossible to express. The
love and encouragement of my parents have provided a constant source of strength. My
brothers and sister have also offered support. Thanks John, Cory, and Ashley.

Finally, I truly owe everything to my wife Michele. Her love and friendship make

every part of my life better.

iv

TABLE OF CONTENTS

LIST OF TABLES ................................................................................................... vii
LIST OF FIGURES ................................................................................................. ix
INTRODUCTION ................................................................................................... 1
CHAPTER 1: EDUCATIONAL ATTAINMENT AND THE
HEALTH OF OLDER AMERICANS ................................................................... 7
1. Conceptual Framework .................................................................. 10
II. The Measurement of Health ........................................................... 14
III. Data ................................................................................................... 18
IV. Empirical Framework ..................................................................... 20
A. OLS Approach ....................................................................... 22
B. Instrumental Variables Approach Using Quarter
of Birth ................................................................................... 24
C. Alternative Instrumental Variables Procedures ..................... 30
V. Results ............................................................................................... 32
A. OLS Results ........................................................................... 32
B. Instrumental Variables Results Using Quarter
of Birth ................................................................................... 34
C. Instrumental Variables Results Using Other
Instruments ............................................................................ 36
VI. Conclusion ........................................................................................ 37
References ..................................................................................................... 51

CHAPTER 2: SELF -REPORTED AGE DISCRIMINATION

AND LABOR MARKET OUTCOMES ................................................................ 55
1. Related Work ................................................................................... 57
A. Pre-ADEA ............................................................................. 57

B. Post-ADEA ............................................................................ 59

11. Data ................................................................................................... 61
m. Empirical Methodology .................................................................. 63
A. Earnings ................................................................................. 63

B. Wage Growth ......................................................................... 66

C. Employer Separation ............................................................. 69

D. Early Retirement .................................................................... 70

IV. Results ............................................................................................... 72
A. Earnings ................................................................................. 73

B. Wage Growth ......................................................................... 74

C. Employer Separation ............................................................. 76

D. Early Retirement .................................................................... 78
V. Conclusion ........................................................................................ 79
References ..................................................................................................... 88

CHAPTER 3: AGE DISCRIMINATION LEGISLATION,
DELAYED PAYMENT CONTRACTS, AND THE HIRING

OF OLDER WORKERS ......................................................................................... 90
I. Background on Age Discrimination and Age
Discrimination Legislation .............................................................. 94
II. Conceptual Framework ................................................................... 97
A. Delayed Payment Contracts and Age
Discrimination Legislation ..................................................... 98
B. Delayed Payment Contracts and Firm
Hiring Practices ...................................................................... 101
C. Age Discrimination Laws and Firm
Hiring Practices ...................................................................... 105
III. The Effect of Age Discrimination Legislation
on the Employment of Older Individuals ...................................... 107
IV. The Effect of Age Discrimination Legislation
on Hiring and Worker Retention ................................................... 111
A. Identifying Newly Hired Workers in the
CPS ........................................................................................ 1 12
B. Empirical Methods and Results ............................................. 114
C. Age Discrimination Laws and Worker
Retention ................................................................................ 1 17
D. Discussion of Results ............................................................. 119
V. Conclusion ........................................................................................ 121
References ..................................................................................................... 133

vi

CHAPTER 1

Table 1:

Table 2:

Table 3:

Table 4:

Table 5:

Table 6:

Table 7:

Table 8:

Table 9:

Table 10:

CHAPTER 2

Table 1:

Table 2:

LIST OF TABLES

Checks on the Validity of Functional Limitation
Measures Using Self-Reported General Health ................................. 39

Descriptive Statistics for Health Outcomes by
Educational Attainment ..................................................................... 40

Descriptive Statistics for Education by Quarter
of Birth ............................................................................................... 41

OLS Estimates of the Effect of Educational
Attainment on Health Outcomes ........................................................ 42

Summary of Estimates of the Effect of Regressors
on Selected Health Measures ............................................................. 44

2SLS Estimates of the Effect of Educational
Attainment on Health Outcomes Using Quarter of
Birth as an Instrument ........................................................................ 45

TSIV Estimates of the Effect of Educational
Attainment on Health ......................................................................... 46

Alternative ZSLS Estimates Using Family

Characteristics and Quarter of Birth as

Instruments, Women .......................................................................... 47
Alternative ZSLS Estimates Using Family

Characteristics and Quarter of Birth as

Instruments, Men ............................................................................... 48

Summary of Selected Estimates of the Effect of
Education on Health ........................................................................... 49

Descriptive Statistics by Whether or Not
Discriminatory Practices are Reported .............................................. 81

Effect of Discrimination on Wages .................................................... 82

vii

Table 3: Effect of Discriminatory Practices on the
Experience-Earnings Proﬁle, Cross-sectional

Results ................................................................................................ 83
Table 4: Effect of Discrimination on the Percentage

Change in Wages Across Periods ...................................................... 84
Table 5: Estimated Marginal Effects on the Probability

of Employer Separation from a Logit Model ..................................... 85
Table 6: Effect of Discrimination on the Self-Assessed

Probability of Working at Ages 62 and 65 ........................................ 86
Table 7: Estimated Marginal Effects on the Probability

of Retirement from a Logit Model ..................................................... 87
CHAPTER 3
Table 1: Summary of the Existence and Coverage of

State Age Discrimination Laws ......................................................... 124

Table 2: The Effect of Age Discrimination Legislation
on the Probability of Employment ..................................................... 126

Table 3: The Effect of Age Discrimination Legislation
on the Probability of Being Newly Hired .......................................... 127

Table 4: The Effect of Age Discrimination Legislation
on the Probability of Being a Newly Hired
Older Worker ..................................................................................... 128

Table 5: The Effect of Age Discrimination Legislation
on the Probability of Losing One’s Job .............................................. 129

Table 6: The Effect of Age Discrimination Legislation
on the Probability of Unemployment and

Retirement .......................................................................................... 130
Table 7: Implications of Estimates for Individuals in a
Hypothetical State .............................................................................. 131

viii

LIST OF FIGURES
CHAPTER 3

Figure l ...................................................................................................................... 132

ix

INTRODUCTION

An aging US. population creates many challenges for policy makers, and, thus,
renders research that sheds light on the conditions of older Americans of great
importance. My aim is to contribute to this research by using economic analysis to
address three very important questions: What is the effect of educational attainment on
the health of older Americans? Do ﬁrms that use age as a discriminatory criterion in
promotion decisions adversely affect the labor market outcomes of their older workers?
Finally, does legislation aimed at combating age discrimination change the propensity of
ﬁrms to hire younger workers?

Chapter 1 is an attempt to answer the ﬁrst question. The missing element of
investigations of the education-health relationship is an attempt to identify the causal
effect of education. Studies that have shown the correlation between education and
health fail to account fully for the fact that education and health are jointly determined.
Thus, as long as there are unobserved factors that inﬂuence both the attainment of more
education and the health status of adults, then it is certainly possible that the correlation
between education and health is spurious. As an illustration of this, consider an
individual’s childhood environment. There are many dimensions of an individual’s
upbringing that may affect both his attainment of education and health status. Some of
these factors are observable and found in most conventional cross-sectional data sets,
such as the education of parents. Many others are not observed, such as the type of

friends one had as a child. Failure to account for all of these factors leads to biases in

estimates of education’s effect on health, as the effect of these factors on health may be
mistaken for an effect of education.

Thus, the aim of Chapter 1 is to present an empirical model of the determination
of both educational attainment and health. Then, identiﬁcation strategies are pursued
such that the consistent estimation of education’s effect on health status may be
determined. The strategy is twofold. First, the Health and Retirement Study (HRS) is
used. This data set contains many variables that can potentially serve as proxy variables
for previously unobserved factors. Second, instrumental variables are used in a two-stage
least squares procedure. The main instrumental variable employed is one’s quarter of
birth, which is thought to affect the level of education achieved due to the way in which
compulsory schools operate in the US. From there, other potential instruments are tested
for the validity of their use in the estimation of the model. The end result is the most
reliable estimates of education’s causal effect on health to date.

The focus on education's effect on the health outcomes of those nearing old age is
particularly important from a public policy perspective. It will reveal much about the
health and level of independence to be expected of the elderly in the near future. This is
critical for Medicare and Social Security policy decisions.

In Chapter 2, attention shifts to the labor market conditions of older Americans.
Although older workers typically earn more than younger workers and usually enjoy a
lower unemployment rate as a group, the treatment of older workers has long been a t0pic
of utmost political interest, and there has been some evidence to support the concern.

The most commonly cited example of perceived ill treatment was the practice by many

ﬁrms of forcing retirement before workers were willing to depart voluntarily from the

workforce. Others included the preferences in promotion to younger workers displayed
by some ﬁrm managers and the lack of labor market opportunities faced by older workers
that desired to either switch jobs or to get a new job after a layoff.

Such perceived ills led Congress to enact the Age Discrimination in Employment
Act (ADEA) in 1967. Recently, the ADEA has been increasingly used as a basis for
lawsuits ﬁled by individuals that think they are treated unfairly based on their age.
Because of this, it is important to determine whether age discrimination actually is a
problem in current labor markets. Chapter 2 uses the HRS to look at the effect of one
type of age discrimination-ﬁrm preferences in promotion toward younger workers.
While the effect of preferences in promotion is an interesting question in itself, for the
purposes of this paper it serves as an indicator of ﬁrm discrimination in general. First,
the paper veriﬁes whether individuals reporting age discrimination are paid less than
similarly trained and qualiﬁed workers. Second, the paper investigates how the self-
assessed probabilities of early retirement are affected by ﬁrm discriminatory practices.
Finally, workers are tracked to determine whether they ultimately are more likely to
experience lower wage growth, to separate from their employer, or to retire early, as a
function of their ﬁrrn’s propensity to promote younger workers. The ﬁndings show that
those reporting discrimination do in fact experience some negative effects in the labor
market. The most signiﬁcant negative effects are on wage growth. Yet, the evidence
also suggests that discrimination may affect individual assessments of the probability of
retiring early. The latter effect is quite important when viewed in a public policy context.
If the aged are driven into early retirement due to ﬁrm discriminatory practices, this has

implications for social security ﬁnancing.

Chapter 3 concludes the dissertation by addressing the effect of age
discrimination legislation on ﬁrm hiring behavior, which is a previously unstudied topic.
The underlying reason for the investigation stems from the relationship between two
often observed labor market phenomena. First, ﬁrms are reluctant to hire older workers
for many types of jobs. Second, lifetime earnings tend to be concentrated toward the end
of the work life. Delaying payments may be a way in which some ﬁrms efﬁciently
reduce worker shirking. The two phenomena may be related because delaying payments
raises the expectations on the part of the worker that his ﬁrm will renege on the
employment relationship. That is, a ﬁrm will exploit the services of the worker at the
beginning of his work life and ﬁre him before the higher wages are realized. In order for
workers to still enter into delayed payment arrangements, ﬁrms must either increase the
payments to workers to compensate for their fear of termination or bear some of the cost
of shirking. In either case, there is a ﬁxed cost associated with hiring new workers.
From a conceptual standpoint, these ﬁxed costs are not unlike training costs. To
maximize the return to training workers, ﬁrms seek to hire younger workers. Likewise,
delayed payment contract schemes lead to preferences in hiring for younger workers over
older workers.

At ﬁrst glance, it appears as if the effect of age discrimination legislation should
be to decrease the propensity to hire younger workers. Assuming the legislation is
enforced, the treatment of older individuals in labor markets should improve. Thus,
hiring of older workers should increase, especially given the fact that the age
discrimination laws of many states explicitly forbid discrimination in hiring. Moreover,

if Lazear (1979) was correct in his prediction that age discrimination legislation should

reduce the use of delayed payment contracts because ﬁrms would no longer be able to
enforce the endpoint of long-term employment relationships, age discrimination
legislation should result in even more of a shift away from ﬁrm preferences toward hiring
younger workers. The problem is that no evidence supports Lazear’s prediction. In fact,
all of the evidence suggests that mandatory retirement provisions are not critical to the
enforcement of delayed payment contracts and other means exist whereby such
arrangements can be enforced. Moreover, the only evidence that exists that directly tests
whether age discrimination legislation reduces the use of delayed payment contracts ﬁnds
evidence consistent with these laws increasing the use of the long-term contracts
(Neumark and Stock 1999).

The ultimate effect on ﬁrm hiring practices is therefore ambiguous and requires
empirical examination. To conduct the investigation, I utilize state variation in age
discrimination legislation from 1964 to 1967. Using a control group of workers in states
without legislation and a treatment group of workers in states with legislation, I estimate
the effect of discrimination laws on employment and hiring. To supplement the analysis,
I incorporate data from 1968 to 1972 and information on federal legislation.

From a public policy perspective, the results of this paper are quite interesting as
well. Age discrimination legislation, while much maligned by economists from the start
for its potential to disrupt the ability of ﬁrms to engage in efﬁcient long-term contracts,
perhaps may be efﬁciency-enhancing. While there may be clear beneﬁts accrued from
the legislation, these may be coming at the expense of certain workers. For example, if

the fact that more ﬁrms are entering into delayed payment contracts means older workers

are having even more trouble getting hired, then the legislation may be beneﬁcial to
society but detrimental to the exact group of workers it was intended to beneﬁt.

The evidence in Chapter 3 suggests that employment for older workers increases
as a result of age discrimination laws. The effect of the legislation on the propensity
toward hiring younger workers is less clear. The estimates suggest that legislation does
result in a decrease in the probability of an older worker being hired. The fact that the
hiring effects on younger workers are negative as well, however, suggests that there may
not be an increased propensity toward hiring workers. This means that the effect on
hiring propensities may be neutral. The fact that fewer workers in total are hired may
simply be due to the fact that more older workers are retaining their jobs after the passage

of legislation and fewer openings exist.

CHAPTER 1: EDUCATIONAL ATTAINMENT AND THE HEALTH OF OLDER
AMERICANS

The correlation between education and health has been studied for decades.
Kitagawa and Hauser (1973), for example, showed that there exists a signiﬁcant negative
relationships between educational attainment and mortality. Later research on mortality
has reﬁned this general approach (Feinstein 1993; Preston and Taubman 1994;
Christenson and Johnson 1995). A negative correlation between education and adverse
health outcomes separate from mortality has also been shown (House, et. al 1990, Ross
and Wu 1995, and Smith and Kington 1997 are recent examples). Moreover, Freedman
and Martin (1999) recently showed that educational attainment is important in terms of
explaining recent trends in functional limitations among older Americans.

Policy implications of such work abound. For instance, Medicare and Social
Security ﬁnancing plans are heavily dependent on information that helps to determine
accurately the health of older Americans in the future. The effect of education on health,
therefore, is especially important, given the rising levels of education of those that are
just now approaching advanced ages. Using data from the March 1990 Current
Population Survey (CPS), the mean level of education for various age groups was
calculated. For individuals in their sixties (or, roughly, born in the 19205), the mean
level of education was 11.59 years. For those in their ﬁfties, the level of education was

12.24 years. For individuals in their forties, it was 13.08 years. Equally staggering are

the differences in high school graduation rates. Among individuals in their sixties, 34%
never received a high school diploma. This is much higher than the 25% of those in their
ﬁfties and the 15% of those in their forties that never graduated high school.

The missing element of investigations of the education-health relationship is an
attempt to identify the causal effect of education. Studies that have shown the correlation
between education and health fail to account fully for the fact that education and health
are jointly determined. Thus, as long as there are unobserved factors that inﬂuence both
the attainment of more education and the health status of adults, then it is certainly
possible that the correlation between education and health is spurious. As an illustration
of this, consider an individual’s childhood environment. There are many dimensions of
an individual’s upbringing that may affect both his attainment of education and health
status. Some of these factors are observable and found in most conventional cross-
sectional data sets, such as the education of parents. Many others are not observed, such
as the type of friends one had as a child. Failure to account for all of these factors leads
to biases in estimates of education’s effect on health, as the effect of these factors on
health may be mistaken for an effect of education.

Thus, the aim of this paper is to present an empirical model of the determination
of both educational attainment and health. Then, identiﬁcation strategies are pursued
such that the consistent estimation of education’s effect on health status may be
determined. The strategy is twofold. First, the Health and Retirement Study (HRS) is
used. This data set contains many variables that can potentially serve as proxy variables
for previously unobserved factors. Second, instrumental variables are used in a two-stage

least squares (ZSLS) procedure. The main instrumental variable employed is one’s

quarter of birth, which is thought to affect the level of education achieved due to the way
in which compulsory schools operate in the US. (Angrist and 'Krueger 1991). From
there, other potential instruments are tested for the validity of their use in the estimation
of the model. The end result is the most reliable estimates of education’s causal effect
on health to date. Moreover, this paper is the ﬁrst to focus on the health of those that are
approaching old age (i.e., those in their ﬁfties). Thus, the answers to these questions
reveal much about the health and level of independence to be expected of the elderly in
the near future. This is critical for Medicare and Social Security policy decisions.

The work of Berger and Leigh (1989) comes closest to the approach employed in
this paper. They estimate a model where education and health are jointly determined and
ﬁnd a positive relationship between education and health using the National Longitudinal
Survey of Young Men (NLS) and the National Health and Nutrition Examination Survey
(NHANES).l The measures of health that they use from the NLS are whether health
limits work and whether functional limitations are present. The average age of the
individuals in their NLS sample, however, is 18.2 years. Thus, the effect of educational
attainment on health outcomes when old cannot be inferred from their estimates. The
measure that they use in the N HANES is systolic and diastolic blood pressure. The
NHANES covers a wider age range (25-74 year-olds).

Section I of this paper explores some explanations of why the effect of education
on health may be considered causal, as well as some alternative explanations of the
health-education correlation. Section II discusses some of the problems that exist in the

measurement of health and their implications in estimating a causal effect of schooling.

Section IH summarizes the data used in the paper and veriﬁes that the health-education
correlation exists. Section IV outlines the empirical model and strategies for its

estimation, and Section V presents estimates of education’s effect on health.

1. Conceptual Framework

The paths through which education causes better health later in life are complex.
Some paths are certainly more direct than others. The most obvious among these is that
education enhances an individual’s ability to utilize health care services that exist. This
is the idea of productive efﬁciency, which has at its root Grossman's (1972) model of the
health production function.2 Examples of how education could improve productive
efﬁciency are easy to see. For instance, a more educated person might have a better
understanding of his symptoms and, thus, be more able to explain to a doctor what they
are. The result will be more effective treatment and better health outcomes later in life.

Another way in which education improves health in a fairly direct fashion is
through a better choice of individual health inputs, which is called allocative efﬁciency.
This concept makes sense if one thinks of good health as the outcome of a lifelong
process of providing inputs into what is essentially a production process. For instance,
more educated people may engage in habits that are healthier. Much empirical evidence

supports the allocative efﬁciency hypothesis. Winkleby et a1. (1992) ﬁnd that there is a

 

1 As will be discussed below, their identiﬁcation strategies differ from those employed in this
aper.

gMost empirical work in this area does not attempt to estimate a health production function,

however. The lack of complete information on biological and socioeconomic factors that may

affect health throughout and before one’s life makes this task prohibitively difﬁcult. As will be

discussed later, however, a well-speciﬁed two-equation model of the effect of education on health

is adequate for the purposes of the paper.

10

negative relationship between educational attainment and whether one smokes. Shea et
al. (1991) ﬁnd a similar relationship between education and smoking. They also ﬁnd that
with more educational attainment comes better eating habits and more exercise.
Choosing not to smoke, eating well, and exercising are all examples of positive health
inputs.

A less direct but important way in which education improves health is via the
increased socioeconomic standing that the attainment of more education brings. A prime
example is that more educated individuals obtain better jobs with greater earnings and
beneﬁts (Willis 1986) and, if unemployed, experience shorter unemployment durations
(Moen 1999). This means that not only do more educated individuals have better access
to health care, but they are less likely to be without health insurance for extended periods
of time. In these cases, education is not directly changing individual behaviors that
enhance health, but, without the attainment of education, the increased social standing
would not be reached and health later in life diminished. Thus, this pathway of education
to health cannot be ignored.3

There are other potential causal pathways that are less obvious and direct, but
they are no less critical. Many of these rest in the fact that education enhances an
individual’s cognitive ability, which results in greater health-enhancing skills and
behaviors (Ross and Mirowsky 1999). For example, education enhances an individual’s
ability to socialize (Ross and Mirowsky 1989), which in turn leads to better health

outcomes later in life through increased social support (House et. al. 1988). Also, the

 

3 Some argue that education merely buys individuals credentials, which therefore results in
greater income and social standing (Collins 1979). Ross and Mirowsky (1999) present evidence

11

enhanced cognitive ability that comes with greater education results in a stronger sense of
personal control (Wheaton 1980; Ross and Mirowsky 1989). Such control improves an
individual’s decision-making ability in regard to health matters. Moreover, the sense of
control strengthens the will to engage in healthier habits (Seeman and Seeman 1983).4
While these pathways of causality though which education enhances health
certainly exist, it is possible that education is just reﬂecting other underlying factors that
may be leading to improved health status later in life. That means, certain factors that
determine educational attainment also have an independent effect on health. For
example, educational attainment is largely inﬂuenced by parental input. However,
parental input is also vital to early health resources devoted to children. Thus,
education’s effect on health may merely be reﬂecting differing personal endowments.
Clearly related to this are intergenerational health transmissions, which have been shown
to have a strong effect on the health of older Americans (Smith and Kington 1997).
Thus, how healthy one’s parents were will reﬂect how healthy he or she is. It may also
be a contributing factor to the childhood environment, and, thus affect educational
attainment. Moreover, the decisions made by parents and grandparents concerning where
and how to raise a family, self-image, and attention from others are all family

background characteristics that affect early health and education investments. At the

 

to the contrary, however, showing that the quantity of education received as the much stronger
correlate with good health.

4 It is important to note that all of the above pathways to better health last throughout one’s life.
For example, education improves lifelong learning (Hyman et al. 1976). Thus, one can continue
to enhance allocative and productive efﬁciency for the rest of his or her life. Thus, the estimates
of Berger and Leigh (1989) that are obtained using the NLS cannot reﬂect fully the improvements
to allocative and productive efﬁciency that they claim their evidence supports.

12

very least, ignoring these factors overstates the effect of education (E10 and Preston
1996).

In addition to unobserved childhood environment and family background factors,
many other unobserved factors affect both education and adult health status. For
example, individuals with longer life expectancies will have more time in which to earn
returns to education. Thus, the human capital model suggests that these individuals
invest more in their schooling (Cawley 1998). If one makes the logical assumption that
life expectancy early in life is formed by family health history or one’s general level of
health early in life, then such perception of the length of life is an important unobserved
factor. Another variable that has been mentioned as a potential cause of both higher
educational attainment and better health is the rate of time discount. The basic idea is
that some people place a higher value on the future than others do. These individuals are
more likely to make investments in both their education and health.5

This paper’s objective of determining whether there exists a causal effect of
education on health requires estimates of the effect of education on health to be rid of
biases resulting from the omitted variables described above. At the same time, however,
there must be caution taken so as not to over-control, which would result in failing to
detect the effect of education working through one or more of the above causal pathways.
As an example, consider the tendency of an individual to engage in risky behavior. As

discussed above, this is certainly affected by educational attainment. Although prior

 

5The effect of the rate of time discount on health is the focus of Fuchs (1982) and Farrell and
Fuchs (1982). Both papers provide only mild empirical support for the hypothesis.

13

studies frequently control for such individual behavior, this paper does not so as not to
ignore this important pathway through which education improves health.‘5

Thus, the idea of this paper is to control for as many exogenous factors as
possible that may effect both health and education up until the point where the ultimate
level of education is reached. Then, differing health status later in life can be attributed
to education. Alternatively, this paper also identiﬁes a potential source of exogenous
variation in educational attainment that arises due to compulsory school laws.
Essentially, the latter approach is the closest we can come to randomly assigning
individuals educational levels and then observing health outcomes. Certainly, the paper

represents the best effort to date to estimate a causal effect of schooling.

II. The Measurement of Health

A very relevant problem in estimating education’s effect on health is ﬁnding a
proper way to measure health. There are a wide variety of possibilities, most of which
are self-reported answers to health-related questions and, thus, contain measurement
error. Anderson and Burkhauser (1984) were among the ﬁrst that expressed concern over
the wide usage of self-reported health measures in empirical work. Bound (1991) also
notes that self-reported general health may not be accurate for a variety of reasons. Since

it is subjective, one individual's interpretation of what it means to have good health may

 

‘5 Another case like this concerns the treatment of income. After all, more educated individuals
typically have higher incomes and can afford better health care. Including controls for income
would be a mistake because, as suggested above, one effect of education may be that it affords
greater health consumption possibilities through increased social standing. What is needed is a
measure of the component of a person’s income that has nothing to do with educational
attainment. Clearly, however, measures like own earnings, spousal earnings, and household
assets are all in some way a result of educational attainment.

14

not be the same as another individual’s interpretation. Another problem with self-
reported general health measures is that individuals may have ﬁnancial incentives to
describe themselves as being of poorer health than they actually are. This is because
many social welfare beneﬁts are contingent on one’s inability to work. Moreover, older
men who are not working are more apt to seek ways to justify to themselves and to others
why they are not working. Over-reporting of poor health is one such way.

There are other related pitfalls associated with measuring health. For instance,
individuals who respond with a high number to a question like How many times have you
seen a doctor in the past twelve months? may not be less healthy than individuals who
respond with a low number or even zero. Those who see doctors frequently may be
hypochondriacs or have good health insurance. Wrought with similar problems are
questions like How many times have you been diagnosed with disease X ?.7 Therefore,
questions that pertain to speciﬁc disease diagnoses may not give an accurate overall
depiction of health.

If the ultimate objective is to infer the effect that education has on an individual’s
general health as he or she approaches old age, use of answers to the above questions
may not be the best approach. More importantly, if the reasons for the errors in the
measurement of these variables are systematically related to education, then estimates of
the true effect of education on health may be biased. Unfortunately, the biases of the
estimates go in different directions depending on the reason for the error in reporting.
For instance, if individuals are reporting poorer health simply because they have better

access to health care, estimates of the effect of education on good health may be biased

15

downward since those with better access to health services also tend to be better
educated. On the other hand, if individuals are reporting poor health to justify not
working, then the effect of education on good health may be biased upward since
individuals with more education tend to remain in the labor force longer. Thus, the
ultimate direction of the bias is uncertain (Bound 1991).

Another problem with ﬁnding an adequate measure is that health is multi-
dimensional. For instance, one dimension of health pertains to mortality and morbidity.
Another is embodied in anthropometric measures like body mass index (Thomas and
Strauss 1997). It is quite possible that education may have differing effects on the
various dimensions of health. This raises another question as to the interpretation of the
results in Berger and Leigh (1989) as evidence of education’s causal effect on health.
They argue that diastolic and systolic blood pressure reﬂect an individual’s health. It is
quite possible that they are only picking up one dimension of health.8

The most promising measures of health from an empirical standpoint use the
responses of individuals to questions that ask them to rate on some scale how well they
perform a variety of functions. For instance, one may be asked to rate how easy it is for
him or her to walk several blocks or to sit for two hours. Since almost everyone does
these things each day or at least attempts them every once in a while, most people can

give a dependable account of their ability to perform these functions.

 

7 Strauss et al. (1993) note that there exists systematic bias in self-reported morbidity measures,
thus partially explaining why wealthier individuals commonly report more illness.

8 This is if blood pressure is reﬂecting health at all. Berger and Leigh (1989) themselves note that
blood pressure increases in response to stressful events. In predictive validity tests performed for
this paper, the diagnosis of high blood pressure is among the poorest health measures in the HRS.

l6

There have been efforts to determine the reliability and validity of such measures
of health functioning. Typical reliability tests involve asking a set of health questions of
individuals at one point in time and returning a short time later to ask the same questions,
presumably before health status has changed. In such tests, questions like those above
have been found to be quite reliable as health measures (Strauss et al. 1993). Validity
tests can be done by way of internal consistency checks. That is, if ability to perform
certain functions are good measures of health, then it should be the case that those
reporting high degrees of difﬁculty in physical functioning should have a low opinion of
their own health. Such tests of internal consistency are normally passed by measures of
functional limitation.9

Table 1 of this paper presents some evidence in favor of the validity of the
functional limitation measures used in the HRS. In row one, the ﬁrst column reports the
proportion of individuals that claim to have little or no difﬁculty climbing several ﬂights
of stairs that also rate their own health as excellent. The second column reports the
proportion of this group that rate their own health as very good, the third column reports
the proportion that rate their own health as good, the fourth column reports the proportion
that rate their own health as fair, and the ﬁfth column reports the proportion that rate their
own health as poor. In rows two through ﬁve, similar numbers are reported for some
other functional ability measures. These are ease in stooping, kneeling, or crouching,
ease in walking a block, ease in bathing or showering, and ease in picking up a dime.10

These proportions should be compared to the proportions for the population as a whole

 

9 See Strauss et al. (1993).

17

that appear in the ﬁnal row. The table shows that a higher proportion of those who report
ease in performing certain daily functions also have a higher opinion of their health than
the population as a whole.

The ﬁnal column in each row reports coefﬁcient estimates of regressions of self-
reported good health on each of the functional ability measures. Each row contains a
coefﬁcient estimate from a separate regression. Good health is a dummy variable set
equal to one if the respondent’s rating of his or her own health is good or better. Given
that this is a linear probability model, estimates should be interpreted as the difference in
probability of reporting good or better health between respondents who can perform the
functions with ease and respondents who cannot. For instance, an individual able to
climb several ﬂights of stairs with ease is 42.37% more likely to report his health as good
or better than is an individual who has difﬁculty climbing stairs. As can be seen from the
table, all of the health measures that will used in the rest of the paper appear to be
internally consistent, as each signiﬁcantly increases the probability of an individual
reporting good health. Overall, one’s ease in bathing or showering has the biggest effect
on reporting good health.

Functional ability measures are also a good measure of health because they reﬂect
an individual's level of independence. This is why measures of functional limitation have
been used extensively in studying the health of older Americans. Smith and Kington's
(1997) work is an example. Also, Smith (1998) uses the HRS to determine how wealth

affects the health of older Americans. He also follows individuals across time, and he

 

‘0 Attention is limited to these functions for purposes of exposition. However, estimates for more
than a dozen other functions reveal similar results to those reported throughout the paper and are
available upon request.

18

ﬁnds that those reporting difﬁculties in performing certain functions in one period are

more likely to report ill health in later periods.

III. Data

The data used in this paper come from the HRS. The HRS is a national
longitudinal study, but since one would not expect the education of those in their ﬁfties to
change very much, there is little or nothing to be gained here from using panel data to
solve the econometric problems at hand. Thus, attention is limited to the Wave I public
release of the data set. This release consists of answers to questions posed during 1992
on the health, retirement, and economic status of individuals born between the years 1931
and 1941. The sample is restricted to U.S.-born individuals between the ages of ﬁfty-one
and sixty-one.

Panel A of Table 2 presents means of the health variables of this paper by
educational attainment for females. The differences in the health status between these
education groups are striking. Among those women that have never graduated high
school, 60% rate their own health as good or better. This is lower than the 82% of
women who graduated high school and rate their health as good or better and the 94% of
college graduates who rate their health in such a positive fashion. Similar differences are
found when comparing the proportions rating health as very good or better and excellent
or better. One’s ease in performing daily functions is also correlated with her educational
attainment. For instance, 62% of high school dropouts ﬁnd stooping, kneeling, and

crouching to be easy tasks. However, 75% of high school graduates and 84% of college

19

graduates ﬁnd them easy. Even one’s ease in picking up a dime differs by schooling
group.

Panel B of Table 2 presents the proportions of the same health measures by
educational attainment for men. The differences across education groups appear to be
smaller, but they are still very evident. For men who have never graduated high school,
63% rate their health as good or better. A similar high opinion of one’s own health,
however, is shared by 82% of high school graduates and 92% of college graduates. Also,
a college graduate is almost three times as likely to rate his general health as excellent or
better than is a high school dropout. The measures of functional ability also are
performed with ease by a greater proportion of those with higher levels of education. For
example, a little less than 90% of men who have not graduated from high school can
walk a block with ease. Over 95% of high school graduates and over 98% of college

graduates, however, can walk a block.11

IV. Empirical Framework

 

” A valid concern is that the results in these tables merely reﬂect the strong age-education link of
the cohort. In fact, a simple regression of the highest grade completed on age indicates that men
have about .07 of a grade less education on average than do men a year younger. Similar
numbers exist for women. Also, men and women are about 1% less likely to have graduated high
school than individuals a year younger. These seem like small effects, but across the cohort as a
whole, which spans eleven years, the educational differences of the average sixty-one year old
and the average ﬁfty-one is over three-quarters of a year less schooling and a 10% lower
probability of having graduated high school for the sixty-one year old. Given that health
deteriorates with age (e. g., regression results show that sixty-one year old men are about 6% less
likely to report excellent health than ﬁfty-one year old men and are about 6% less likely to be
able to walk across a room with ease), it is useful to check if the results of Table 2 hold for age
subsets of the cohort. This was done, and the results do hold across age groups. Also, similar
results were found when the population was broken down into whites and non-whites. Thus,
there appears to be an education-health correlation that cuts across races.

20

The basic econometric model used to investigate the effect of education on health

is summarized by the following two equations:

0) E=Xm+2¢+m
and
(2) H, = X43 + 7E, + 8,.

E is the educational attainment of individual i measured as years of schooling completed.
Its coefﬁcient, 7, is the parameter for which this paper seeks an estimate. H, is the health
outcome for individual i. X, denotes the observed variables that are thought to affect
educational attainment and health outcomes. Included are ten year of birth dummy
variables,12 dummy variables for Black and Hispanic race, dummy variables for region of
birth, and dummy variables for being divorced or separated, widowed, and never
married.13 Z; are observed factors affecting education but not health. The error term 11,
represents the factors causing variation in education that are not included in X, or Z.
Likewise, Si is the health error term. These error terms are allowed to be correlated with
each other. This allows for omitted factors, like childhood environment and rate of time

discount, to affect both education and health.

 

'2 Year of birth is one way to control for the fact that both education and health decrease su’ongly
by age in the cohort, as documented earlier. Another way is to detrend the education variable by
subtracting off a moving average term from the highest grade completed term, as was done by
Angrist and Krueger (1991) when studying the effects of education in this cohort. Another way
is to include a linear term in age. Each of these alternatives was tried, with no remarkable change
in the results.

'3 A dummy variable is also included to indicate whether the respondent answered the survey
questions for himself or a proxy answered the questions.

21

The model can be amended to account for the potential of reverse causality. For
instance, Hi could be added to the right side of equation (1). This would not be
appropriate here, however, because the health outcomes of this paper are measured when
old. These are not known when the education decision is being made. The more sensible
approach is to allow factors related to life expectancy to be components of the error terms
in both equation (1) and equation (2). Thus, incorporating the potential of reverse
causality (as it pertains to the analysis of this paper) requires no explicit change to the
above model.

Estimation of the model can proceed in a variety of ways. First, one can estimate
equation (2) only by ordinary least squares (OLS), using proxy variables to control for
those factors causing correlation between the error terms. Alternatively, one can estimate
the two-equation model via a 2SLS approach. This requires ﬁnding a suitable Z; to

identify the model. Both approaches are outlined in the remainder of the section.

A. OLS Approach
For this subsection, equation (1) is ignored except for the fact that it is understood
that there exists a set of factors correlated with educational attainment that are also

correlated with health. In this context, equation (2) can be rewritten

(3) H = XiB + 7E, + ct + v,.

The vector of unobservable variables that may affect health outcomes is denoted ct.

These variables may also be correlated with the observed demographic characteristics

22

and education. It is in ct that one would ﬁnd family background information, the rate of
time discount, and other unobserved factors affecting health.

One approach is to assume that c; is not correlated with the observed variables in
the analysis and to incorporate it into the error term. Then, the parameters in equation (3)
can be estimated consistently through an OLS approach.14 This is a good starting point,
but given the prior discussion of how family background and unobserved individual-
speciﬁc characteristics might affect health and education, the exogeneity of education in
such speciﬁcations is questionable at best. A better approach is to break up the c, in
equation (3) into an intergenerational health transmission component (c.), a childhood
environment component (ch) and an individual component (cg). The equation that will

now be estimated is

(4) Hm = XmgB + yEmg + c. + ch +cg + am.

To estimate equation (4), an assumption can be made that C8 is not correlated with
educational attainment. This means that one must assume that individual-speciﬁc factors
like the rate of time discount either are not correlated with education or do not belong in
eg (i.e., are not correlated with health). Given that the empirical evidence of the effect of
the rate of time discount on health is not completely convincing (Fuchs 1992; Farrel and
Fuchs 1992; Cawley 1998), and the effect of some abstract inherent ability concept on

health is uncertain, this assumption is not at all ludicrous. Thus, if one obtained proxy

 

'4 Technically, given that all health variables are coded to be binary, a linear probability model is
used to obtain the estimates. All standard errors reported are corrected for heteroscedasticity.

23

variables for the family background factors that comprise c. and ch, consistent estimates
of 7 could be found.

Several good proxy variable candidates exist in the HRS. Including in the
speciﬁcation a variable indicating whether or not one’s parents are currently living can
capture the effect of the intergenerational health transmission component of an
individual’s family background. The childhood environment component of family
background can also be captured at least partially by these proxy variables (Smith and
Kington 1997). Other important proxies for childhood environment are mother’s
education, father’s education, and height. More educated parents improve the health of
children for the same reasons that an individual with more education is able to improve
his own health later in life. Education will enhance both a parent’s allocative efﬁciency
in choosing better inputs for the health production of the child and productive efﬁciency
as they can better use the health inputs that exist (Schultz 1984). Height has also been
shown to be a good indicator of early investments in one’s health (Thomas and Su'auss
1997). This is important here because these investments most likely took place before
one’s ultimate educational attainment was reached. Thus, height is a good control for an
individual’s initial health condition. It is also a good proxy variable for the life

expectancy correlates that may affect education.

B. Instrumental Variables Approach Using Quarter of Birth

If one is not willing to assume that individual unobserved factors affecting health

are uncorrelated with educational attainment or that the proxy variables mentioned above

24

are inadequate, then equation (1) and equation (2) must be estimated as a system. A
ZSLS approach will be used to do this.

The ﬁrst step is to ﬁnd a valid Z, (i.e., an instrumental variable or variables). To
be valid, instruments must be correlated with education but free of correlation with the
omitted factors that are causing the correlation between the errors in equation (1) and
equation (2). Berger and Leigh (1989) use family background characteristics, such as
mother’s education, father’s education, and number of siblings, as instruments.
Arguably, these instruments are not appropriate here because they do not meet the latter
criterion mentioned above for instrument validity.15 Speciﬁcally, family background
characteristics are likely correlated with omitted factors affecting health. For this reason,
one could improve on the attempts of ﬁnding the causal effect of education by searching
for better instruments.

The body of research that attempts to estimate education's effect on earnings
serves as a guide. Speciﬁcally, there has been a long-time concern that there exists some
unobserved factor that affects both educational attainment and earnings. This unobserved
factor has been loosely dubbed "ability." That is, something innate in certain individuals
compels them to get more education. This "ability," however, also makes people better
or harder workers later on in life, which leads to higher earnings. If this is true, OLS
estimation procedures incorrectly attribute some of the effect of this unobservable factor
to education. The thought was that this led to upward bias in the estimates of education's

effect on earnings.16

 

'5 This was conﬁrmed through Newey (1985) overidentiﬁcation tests performed for this paper.
’6 As researchers began to tackle the problem of unobserved effects on schooling, however, they
found evidence that OLS did not yield estimates of the effect of education that were biased

25

The work of Angrist and Krueger (1991) appears to stand out in the returns to
schooling literature. It has actually spurred a rnini-literature of its own ﬁlled with
attempts to invalidate and validate their approach. They use an individual’s quarter of
birth as an instrumental variable for schooling and provide a convincing case for its use.
Angrist and Krueger argue that quarter of birth affects one’s educational attainment due
to the way in which compulsory school laws operate in the US. Most states require an
individual to attend school until he reaches the age of sixteen. Others require one to
attend until the age of seventeen or eighteen. It is also the case that individuals are
permitted to start the ﬁrst grade in most states if they turn six by some speciﬁed month,
frequently by the end of the year. Thus, the older members of a class tend to be born
earlier in the year. Moreover, those born earlier in the year reach the age at which they
are allowed to drop out of school ﬁrst, presumably before they have completed the
current grade. Thus, one should expect that among older individuals (for whom dropping
out of high school was still fairly common according to the statistics in the Introduction)
there should be less educational attainment among those born earlier in the year.

Table 3 provides some descriptive statistics for women and men that indicate that
individuals born earlier in the year typically achieve lower levels of educational

attainment than those born later in the year, with educational attainment measured as

 

upward. For example, see Ashenfelter and Krueger (1994) and Card (1993). If anything, the
estimates were biased downward. Griliches (1977) explained that there were sources of bias at
work in estimating the returns to schooling that were different than the unobserved "ability" bias.
For example, schooling is measured with error. The attenuation bias that results pushes the OLS
estimates of the effect of education toward zero. Moreover, efforts to solve the omitted variables
problem by adding an error-ridden measure of ability (e. g., a test score) can exacerbate the biases
already present in the estimates, thus pushing OLS estimates even further toward zero. Certainly,
the returns to schooling literature has far-reaching implications for any work that attempts to
determine the effect of an obviously endogenous regressor. It therefore should have much to say
about estimating the effect of education on health.

26

years of school completed. The proportion of respondents who graduate high school is
lower among those born earlier in the year as well. Quarter of birth appears to have no
effect on the receipt of a college degree, which is not unexpected given that quarter of
birth affects the decision to drop out of high school, not to obtain additional years of
higher education. Verifying the effect of quarter of birth on educational attainment is
important because the impact of a weak ﬁrst stage in 2SLS is potentially spurious
inference. For instance, Angrist and Krueger (1991) report instrumental variables
estimates that are very close to the OLS estimates for U.S.-born men.l7 Since
instrumental variables estimates are consistent,18 and Angrist and Krueger use US.
Census Data with well over one-quarter of a million observations, many might read the
2SLS results of Angrist and Krueger as validation of their OLS results. This might very
well be a mistake. This is due to ﬁnite-sample bias. It has long been known that 2SLS
estimates are biased toward OLS estimates (Nagar 1959) in small samples. Still, popular
belief held that this ﬁnite-sample problem would be unimportant if one obtained a large
enough sample, given the consistency of instrumental variables estimates. Bound et al.
(1995b) and Staiger and Stock (1994) report that even in studies using huge data sets
ﬁnite-sample bias may still be important, and the fact that estimates of OLS and

instrumental variables estimates are close may be largely due to this bias. In fact, Angrist

 

'7 The OLS estimates of the coefﬁcient of education in a log-wage equation from their preferred
speciﬁcation are .063, with a standard error of .0003, for individuals born in the 1930s (the age
group analyzed in this paper) and .07, with a standard error of .0004, for individuals born in the
19205 . The instrumental variables estimates, using three quarter of birth dummy variables
interacted with ten year of birth dummy variables, are .081, with a standard error of .016, for
those born in the 19305 and .069, with a standard error of .015, for individuals born in the 19205
(These estimates were taken from Angrist and Krueger (1995), which summarized results from
their 1991 paper.).

18 This assumes rank and orthogonality conditions are met.

27

and Krueger (1995) themselves, like Bound et al. (1995b), show that replacing quarter of
birth with a random draw from a four-point discrete uniform distribution yielded results
very similar to those in their 1991 paper.

Less controversy has surrounded quarter of birth in terms of the second criteria
for a good instrument. That is, quarter of birth is not expected to have an effect on
unobserved factors related to earnings and, for the purposes of this paper, health.19

With these objections stated up front, this paper proceeds with an instrumental
variables approach similar to that of Angrist and Krueger’s (1991) and is careful to look
for the signs of a weak instrument along the way so as to avoid overstating the results.
The two equations used in the ZSLS speciﬁcation with quarter of birth as a dummy

variable are

6) E=Xm+Qw+m

and equation (2). Equation (5) is just equation (1), with Z replaced by a series of quarter
of birth dummy variables interacted with year of birth dummy variables (Qt). Note that
the series of year of birth dummies is part of X,, the vector of observable characteristics.20

Again, interest lies with the estimate of the parameter 7 in equation (2).

 

'9 Bound and Jaeger (1996), however, do suggest that quarter of birth may be related to some
factors other than just educational attainment, such as schizophrenia. Moreover, there exists the
belief that those who lack the self-discipline to succeed in life also lack the self-constraint to
control heightened sexual passions in the summer, thus leading to more genetically inferior
individuals being born in the ﬁrst quarter of the year than in the others (Huntington 1938).
Empirical evidence on this is certainly mixed (Angrist and Krueger 1992; Lam and Miron 1987).
2° This is close to the speciﬁcation used by Angrist and Krueger (1991). Dummies for quarter of
birth were used by themselves (not interacted) as instruments as well, but they performed more
poorly than the interactions.

28

As mentioned above, there is a potential for ﬁnite-sample bias in the ZSLS
estimates of the health returns to schooling, even if the sample is a large cross-sectional
data set. Therefore, it is certainly to be expected that ﬁnite-sample bias will be a concern
in a relatively small data set like the HRS. If this bias is ignored and a causal effect is
inferred from a 2SLS estimation procedure with a weak ﬁrst stage, there is the potential
for misleading results to be reported. While the above ZSLS procedure would be ideal if
quarter of birth strongly affected educational attainment among HRS respondents,
unfortunately, as will be shown later, it does not. Thus, 2SLS estimates that are close to
OLS estimates may lead to spurious inference.21

Note, however, that quarter of birth may still be a valid instrument. Angrist and
Krueger (1992) present a two-sample instrumental variables (TSIV) estimator. The TSIV
procedure is designed for cases where a set of instruments is available in two samples,

22

but each sample is missing either the dependent variable or the endogenous regressor.

As long as the moments underlying the instrumental variables estimates of the two

 

2’ One solution that allows for consistent estimation when little of the variance of an endogenous
regressor is explained by the instruments is the split-sample instrumental variables (SSIV)
estimation procedure introduced by Angrist and Krueger (1995). They show that a sample may
be randomly divided into two samples. Then, the ﬁrst stage (equation (5) in this paper) can be
estimated using data from the ﬁrst half of the sample, and the estimated parameters can then be
used to construct ﬁtted values for the endogenous variables using data from the second half of the
sample. Then, the second stage (estimation of equation (2)) can be carried out using the predicted
value of the endogenous regressor and data from the second half of the sample. The result will be
an estimate of the effect of the endogenous regressor on the dependent variable. The SSIV
estimates will still be biased, but the direction of the bias will always be toward zero. An
estimate of the bias can be obtained by regressing the endogenous variable on its predicted value.
Scaling the SSIV estimate by the inverse of this estimated attenuation bias yields a consistent
estimate of education’s effect on health as the number of instruments grows. Angrist and Krueger
(1995) call this estimate USSIV, and they present SSIV and USSIV results that are quite close to
their ZSLS estimates, thus supporting the use of quarter of birth as an instrument. There is one
problem with the above procedure as it pertains to the analysis of this paper. Because the sample
is small to begin with, splitting the sample into two will provide a ﬁrst stage that is estimated
using a sample that is so small that it may not be useful for constructing ﬁtted values.

29

samples are independent, the TSIV estimates will be consistent. Thus, what is needed is
a data set that contains quarter of birth and schooling information and enough
observations so as to lead to precise estimates of educational attainment. The 1980 US.
Census 5% microdata sample is a candidate.23 A simple TSIV approach is then applied
where the 1980 Census is used to calculate the expectation of educational attainment
conditional on quarter of birth for individuals born between 1931 and 1941, and the HRS
is used to calculate the conditional expectation of health. Then, a Generalized Least
Squares (GLS) procedure using the conditional expectations is applied. The GLS
weighting matrix is a function of the moments of the two samples.24 The result will be an
efﬁcient and consistent estimate of the effect of education on health if quarter of birth is a

valid instrument.

C. Alternative Instrumental Variables Procedures
As will be elaborated upon in the next section, although conceptually ideal,
quarter of birth by itself is not a stellar performer in the HRS as an instrumental variable
due to its weak correlation with educational attainment. Perhaps another data set is
needed that, using the framework outlined in this paper, could use quarter of birth
information to provide a deﬁnitive estimate of the causal effect of schooling on health.
As a ﬁnal attempt at employing an instrumental variables procedure, this paper

adds some family background characteristics to the basic two-equation system

 

22 The SSIV estimator is a special case of the TSIV estimator, where both samples come from one
larger sample.
23 The 1990 Census would be better, but there is no quarter of birth information in it.

30

summarized in equation (1) and (2) in order to improve on ﬁrst-stage estimation. The

new system can be written:

(6) Bi = Xi“ + Wim + (PZi + Pi
and
(7) Hi = XiB + YEi + Wip + 8i.

W, is a vector containing the intergenerational health transmission proxies (whether or
not one’s mother and father are still alive) and height. Once these controls are added, one
can conduct speciﬁcation tests to determine whether the remaining family background
variables, like parents’ education, also belong in Wt, or whether these variables can serve
as instrumental variables (i.e., belong in Z,). A speciﬁcation test will be used that works
works under the assumption that quarter of birth is a valid instrument, as the instrumental
variables estimator using quarter of birth must serve as the consistent but inefﬁcient
estimator under the null hypothesis. Also under the null, the instrumental variables
estimators using quarter of birth and the family background information will be
consistent and efﬁcient. Under the alternative hypothesis, the instrumental variables
estimators using quarter of birth and the family background information will be
inconsistent. Thus, p-values of greater than .10 indicate a failure to reject that hypothesis
of no misspeciﬁcation when additional family background characteristics are included as

instruments at the .10 level of signiﬁcance. This will be the criterion used to judge the

 

2‘ In the simple linear model of this paper, the weighting matrix is straightforward. Its elements
are the weighted sums (by inverse of sample size) of the cell variances of education in the 1980
Census and the cells variances of health in the HRS.

31

validity of adding additional family background information as instruments. In addition
to parent’s education, the HRS contains some information on siblings. The use of these
variables is also tested.”

This approach may come the closest thus far in the literature at presenting reliable
estimates of the causal effect of education on health. It avoids the problem of using
instruments that either affect health or are correlated with other omitted factors that have
their own unique effects on health. This is because the parental and sibling information
is tested before being added to the vector of instruments. More importantly, instruments
are used that have a higher degree of explanatory power than just quarter of birth alone.
Quarter of birth, along with the exogenous family characteristics, explain much of the

variation in education.

V. Results
A. OLS Results

The paper ﬁrst estimates equation (2) by OLS, without including X, or any proxy
variables. These estimates for women are reported in column (1) of Table 4. Each row
contains estimates of y from separate linear probability models, where the dependent
variable is the binary variable appearing in the left-hand column. The estimates certainly
indicate a positive correlation between education and health. All of the effects on general

health and functional ability are signiﬁcant at the .05 level of signiﬁcance. The estimates

 

2’ Sibling information used in this paper includes a dummy variable for being a middle child, a
dummy variable for being the youngest child, number of siblings, and the proportion of one’s
siblings that are female. Both birth order and sibling sex composition have been shown to be
important determinants of the resources devoted to a child (Rosenzweig and Wolpin 1988).

32

for the speciﬁcations with no controls for men appear in column (5) of Table 4. The
coefﬁcients also indicate that education is positively correlated with good health.

In column (2) and column (6) of Table 4, the vector of individual demographic
controls, X, is added for women and men, respectively. When these controls are
included, the coefﬁcient estimates fall. Next, a proxy variable approach is employed to
attempt to control for c, in equation (3). Equation (4), however, is now the equation to be
estimated, where ct, the vector of unobservables in equation (3), is broken down into an
intergenerational health transmission component (ct), a childhood environment
component (ch), and an individual-speciﬁc component (cg). If the assumption that C8 is
uncorrelated with education is imposed, and suitable proxy variables for c. and ch are
used, then consistent estimates of y will be obtained. Column (3) and column (7) present
estimates of the effect of education on each health measure when the proxy variables for
intergenerational transmissions (whether one’s mother and father are still living) are
added for women and men, respectively. The estimates of the effect of education in
column (4) and column (8) are reported for speciﬁcations that add the proxy variables for
childhood environment (mother’s education, father’s education, and height). For women,
the coefﬁcient estimates tend to fall slightly, but a signiﬁcant, positive effect on health
still holds. For men, the coefﬁcient estimates fall as well. Actually, when all proxy
variables are added, education’s positive effect on ease in bathing and showering is no
longer signiﬁcant, and its effect on ease in picking up a dime is signiﬁcant at the .10 level

of signiﬁcance.” 27

 

26 The results presented thus far do not change markedly if a probit model is used to estimate
equation (2) and equation (4). Moreover, when the measures of general health and functional
ability are redeﬁned to be on a scale of ill health and an ordered probit model is estimated,

33

Table 5 reports the estimated effects of the other variables that are included in the
speciﬁcations in column (4) and column (8) on selected health measures.28 From the
table it can be seen that blacks tend to be less healthy than whites (but not always), and
married individuals are healthier than those who are separated or divorced, widowed, or

have never been married (especially among men). Individuals born in the south are less

 

education continues to have a positive effect on health. Speciﬁcally, general health is given a
value of one to ﬁve. A value of one indicates that an individual rates his or her own health as
excellent. Values of two through ﬁve indicate ratings of very good, good, fair, and poor,
respectively. Likewise, ability to perform a daily function is given a value of one to four, based
on one’s own rating of relative difﬁculty in performing the task. A value of one represents no
difﬁculty while a value of four represents extreme difﬁculty in performing the function. The
effect of education on ill health is negative and highly signiﬁcant for all measures. Even the
effect of education on ease in bathing or showering and easy in picking up a dime, which became
weaker when the proxy variables were added in the OLS speciﬁcations, are positive and
signiﬁcant when ordered probits are used.

27 Thus far in the paper it has been assumed for convenience that education affects health in a
linear fashion, but this need not be the case. In order to test the signiﬁcance of non-linearities in
the data, the following equation is estimated:

(8) Hi = Xiﬂ + Wip + YE, + mm, + azsxi + al'Eﬂ‘tw, + oriEﬂsx, '1' 8;,

where tw, is a dummy variable indicating that an individual has ﬁnished at least the twelfth grade
but not the sixteenth grade (roughly, a high school graduate that has not completed college as
well), and sxi is a dummy variable indicating that the individual has ﬁnished at least the sixteenth
grade (a college graduate). The estimates of the coefﬁcients of these dummy variables, or, and
a2, reﬂect intercept shifts, and the estimates of al’ or a;’ reﬂect slope shifts in education's effect
on health. Estimation of equation (8) for each health measure reveals that there is some evidence
of non-linearities, but, for most of the health measures, there is not a signiﬁcant intercept or slope
change. Thus, while the linear relationship assumed in this paper might not fully reﬂect the
health-schooling relationship for some of the health measures, a linear approximation still seems
adequate for the purposes of this paper.

A potential problem exists, however, if there are implications of the discontinuities and
slope changes that go beyond merely the fact that those who graduate high school have improved
their allocative efﬁciency or productive efﬁciency in a discontinuous way. For instance, there
may be some individual-speciﬁc characteristic that causes an individual to graduate high school
or college that also leads him or her to have better health. Graduation of high school and college
may likely be correlated with individual-speciﬁc effects for which no controls have been used,
like the rate of time discount or cognitive ability. The evidence summarized above suggests that
these effects may be working through the education variable to overstate its effect on health for at
least some of the measures. This just indicates the need to use an estimation procedure to deal
with the problem of omitted individual effects.

28 Not included in the table are the dummy variables for year born, the dummy variable for
whether or not one answered the survey himself or herself, and dummy variables denoting
missing information for a particular variable. Observations for which a value of a variable was
missing were coded to zero, and dummy variables denoting that this was done were constructed.

34

healthy. The coefﬁcient estimates for the proxy variables indicate those with parents that
are still living are healthier. Taller individuals are healthier later in life as well, except

when health is measured by functional ability. Father’s education tends to have an effect
on good health. Mother’s education, on the other hand, has little effect on health holding
the other factors constant. This is a bit surprising, but Smith and Kington (1997) report a

similar result.

B. Instrumental Variables Results Using Quarter of Birth

The instrumental variables approach using quarter of birth as outlined in equation
(5) and equation (2) is now used, with the results for women reported in left panel of
Table 6. In the second column, ZSLS results are presented. The ﬁrst column copies the
OLS estimates that are reported in the second column of Table 4 for easy comparison
with the 2SLS estimates in the adjacent colurrms. The ZSLS estimates are quite close to
the OLS estimates. Education has a positive effect on many of the health measures, and
the p-values of Hausman tests indicate that instrumenting may not be necessary. There
are a few reasons to be skeptical of these results, however. First, the standard errors of
the ZSLS estimates are quite high. Second, the F-statistic for a test of the joint
signiﬁcance of the ﬁrst-stage excluded instruments is near one. Thus, the ZSLS estimates
themselves are strongly biased toward the OLS estimates.

The right panel of Table 6 presents similar results for men. Again, ZSLS
estimates are close to the OLS estimates, which leads to p-values on Hausman tests that

suggest instrumenting to be unnecessary. Still, however, the standard errors are large,

35

and the F-test of the joint signiﬁcance of the quarter of birth instruments indicates that
the ZSLS approach is perhaps even more unsuitable for men.29
To improve on the use of quarter of birth as an instrumental variable in the HRS,
Census data is incorporated into the analysis. Speciﬁcally, expectations conditional on
quarter of birth are calculated for highest grade completed using the 1980 Census and for
the health variables using the HRS. Then, the TSIV procedure outlined earlier in the
paper is performed using year of birth as a regressor. The results for women and men are
reported in Table 7. While the estimated effects for both women and men are imprecise,

for the most part they support the previous ﬁndings that education has a strong positive

effect on many measures of health.

C. Instrumental Variables Results Using Other Instruments

As a ﬁnal empirical strategy, selected parental and sibling characteristics are
added to the vector of instrumental variables, and the system summarized by equation (6)
and (7) is estimated using 2SLS. Remember that instruments are added only if the
estimator including them as instruments passes a speciﬁcation test.

The results appear for women in Table 8. The ﬁrst column reports the 2SLS

estimates using quarter of birth and the family background characteristics that are

 

29 To further investigate the validity of using quarter of birth as an instrumental variable in the
HRS, the SSIV procedure outlined in footnote 21 is used. For both women and men, a mini-
Monte Carlo experiment was conducted, where the SSIV procedure was performed 125 times.
That is, the sample was divided and SSIV and USSIV estimates were obtained. Then, the
procedure was repeated 124 more times. Across the board, the standard deviations of the
estimates were huge, indicating that that the SSIV and USSIV results are quite sensitive to the
splitting of the sample, a very undesirable property of SSIV and USSIV estimates. Also, SSIV
and USSIV estimates appear to be centered around zero for the most part. Finally, the mean of
the estimates of the inverse of the attenuation bias is about .07 for both women and men.. Taking
all of these results together indicates that quarter of birth is not a strong instrument in the HRS.

36

deemed valid as instruments. These estimates indicate that education has a positive
effect on all of the health measures. The effect is signiﬁcant at the .05 level of
signiﬁcance for each of the general health and functional ability measures, except the
rating of one’s own health as very good or better and ease in stooping, kneeling, or
crouching. For several of the measures, the p-value of the Hausman Test in the second
column suggests OLS estimates are, if anything, biased downward.

Table 9 reports the results for men. As with the results for women, these indicate
that the effect of education becomes stronger after one instruments for education. In fact,
among the general health and functional ability measures, only ease in climbing several
ﬂights of stairs is not positively and signiﬁcantly affected by education according to

these estimates.

VI. Conclusion

This paper provides a framework for empirical investigation into the causal effect
of education on the health of older Americans and presents evidence that there is a
positive effect for various measures of health. Such evidence indicates that there
certainly is a causal component to the correlation between education and health that is
often observed.

Prior investigations into this question to date have lacked a full treatment of the
econometric problems inherent in estimating education's effect on health. This paper,
however, uses proxy variables for omitted intergenerational health transmission and

childhood environment, and it instruments for the education variable to purge estimates

37

of biases resulting from omitted individual-speciﬁc error components correlated with
education.

The OLS results indicate that higher levels of educational attainment appear to
positively affect the health of older Americans. This holds even after one controls for the
observable demographic characteristics of the individual and his or her family
background. The effect is especially strong among women when health is measured as
either self-reported general health or the ease in performing basic functions. For men, the
effect of education on such health measures is also signiﬁcant, but it is not quite as
strong.

The ZSLS results of the paper provide further evidence of education’s causal
effect on health for both women and men. Table 10 summarizes estimates from selected
speciﬁcations throughout the paper. As can be seen from the table, most estimates of the
paper indicate a positive and signiﬁcant (at least at the .10 level of signiﬁcance) effect of
education on health. Moreover, the majority of the estimates that are not signiﬁcant are
positive. Whether quarter of birth or quarter of birth and selected family characteristics
are used as instruments, the effect is positive for the vast majority of the estimates and
signiﬁcant for a large number of them. Altogether, the results indicate that even after
correcting for biases in the estimates resulting from omitted variables, education’s effect

on health remains positive for many reliable health measures.

38

038300 2000 323 3:80.. 20 Echo 200:0; .:0~m_m=oo.§0u30088203 .0033

.8 Bow 8 5.00: .0; .5 m3 808 0:0 050.? mm 0303; E03030 05 2.0 5:38 98:90. 05 E bowoﬁo 05 5 05 0.8 005023 ﬂ 0309?
802.039: 05 2055 £0.58 3:338.” 000:: 8888 SE... 83850 0:30.. 5:200 Rec 05. soon .5 .03 68w 68w b0> Jc0=0ox0 v.0
5—00: 55¢ :05 008 9:0 ~05 5:38 93.30— 05 E .60ng 5—00: :80 95 E of» soar—omen 05 20 05:28 0.»: 30¢ 05 E 00:30”. ”082

 

 

 

 

Ewe. mwmﬁ GEN. 53. min. coca—amen Exam
Awmmos
3%. 88. $2. 5%. «man. was. 2:6 a a: sea
cwmoe
echo. once. ~62. Cwm. wmmm. mew. 030% 8 053
Awe—co
.33. 33. thﬁ omwm. oven. mwmm. 0.83 a 0:95
880g
Ewm. 28. mag. sown. ammm. 33. 50:80 no ._005_ deem
Awwoog
hmmv. 88. mmwo. mmwm. Rem. RR. 83:. .«o «.0sz 98:0
”mmlg 53, MERE—8
05 20 awe 0:3 0.85 “—0
@033 .8 Bow
mm 5.00: :30
808 ”0303? noon Ea voow voow 30> 80:088

E03039 00 5—00..— 0008 an 53.0: 00.8 00 5—00: v.03.— 00 5—00: 88.. 00 .200: «.88
8065000 30 $5 59395 35 :oquo.& ~05 :oEomoi 85 5930.5 :05 BER—Em

 

5.8: .8050 autonomiom male: .2382 805:5.— .oeouoaa... .o 323» 2: ao 9.85 3 03a...

39

Table 2: Descriptive Statistics for Health Outcomes by Educational Attainment

 

 

Entire Dropout High Some College
population school college graduate
graduate
A. Women
Rate own health as:
good or better .7948 .6027 .8198 .8653 .9384
very good or better .5397 .3059 .5553 .6372 .7403
excellent or better .2297 .0948 .2279 .2878 .3589
Can do the following
with ease:
climb ﬂights of stairs .6823 .5369 .6860 .7481 .8170
stoop, kneel, or crouch .7432 .6211 .7452 .8148 .8379
walk a block .9351 .8735 .9431 .9645 .9721
bathe or shower .9777 .9571 .9776 .9899 .9949
pick up a dime .9726 .9464 .9752 .9840 .9918
Number 4577 1241 1846 860 630
B. Men
Rate own health as:
good or better .8093 .6271 .8176 .8663 .9247
very good or better .5303 .3060 .5156 .5892 .7188
excellent or better .2373 .1274 .2065 .2500 .3781
C_an do ﬁe following
with ease:
climb ﬂights of stairs .7881 .6462 .7790 .8158 .9152
stoop, kneel, or crouch .8259 .7396 .8210 .8357 .9085
walk a block .9483 .8914 .9531 .9641 .9827
bathe or shower .9811 .9655 .9845 .9804 .9918
pick up a dime .9725 .9493 .9802 .9658 .9894
Number 4059 1050 l 377 769 863

 

Note: Reported are the proportions in each education category that ﬁt into the health category in
the left-hand column.

40

Table 3: Descriptive Statistics for Education by Quarter of Birth

 

First quarter Second quarter Third quarter Fourth quarter
loam:
highest grade
completed 12.138 12.255 12.093 12.177
high school graduate .725 .723 .732 .736
college graduate .140 .145 .130 .135
sample size 1149 1146 1145 1137
Men;
highest grade
completed 12.353 12.365 12.389 12.410
high school graduate .738 .737 .748 .742
college graduate .218 .214 .204 .216
sample size 996 985 1105 973

 

Note: Reported are the means (for highest grade attended) or proportions (for level completed) of
various education measures for those individuals born in a particular quarter of the year.

41

 

 

383 6.8.8 .28.. 2 83 2 83 8.8.. 8.8.. 8.8..
.8... 8o. 88. :8. 88. 88. 88. 88. as... a a: .o...

8.8.. .888 .88.. .88.. 2.8.. 8.8.. 8.8.. 8.8..
:8. .8... m8... .8... 88. .8... .8... 88. 4.36..m ao ass

8.8.. .288 6.88 .288 :83 .083 .88.. .88..
88. m. 8. a. .o. 8.... o. .o. 88. 8.... 8.... .8... a 5...;

.88.. .88.. .88.. :83 :83 .88.. .88.. :88
RS. 88. as. 88. 88. 88. 88. :8. .258 so .85. .88

683 .888 $83 :83 .88.. 383 .88.. :83
88. m. 8. :8. 88. 88. $8. 88. 88. an: .o 2...... 2:8
”ammo 5:3
WEBB—0.. 0... cc 3

683 $83 $83 .880 883 383 683 $83
$8. 88. 88. 88. 88. $8. 88. 88. .888 52.86

$83 $88 .880 .888 $88 88.: 683 .88..
88. 88. 88. 88. 83. 88. 88. 3.8. .28 .5 8o» be,

688 .88.. .88.. :83 :83 $83 .88.. $83
88. 88. 88. :8. £8. :8. 88. 88. .82. to 8%
”mm 5—mon— ESO 3a”—

.8 E .8 5 .3 5 .8 3
.5803

 

8.53:0 5.8.. no 30555.... 3808...... .o .88. a... .o 8388.. 0.5 3 05a...

42

.005 .0. 0..000. €000. A... 050.00 0w00.5 .3 050.00 .0050.» .0. 0..000. 0000. .0. 050.00 0w00.5 2. 050.00 ....w.0.. 50000000. 000 50005.5...

5.00.5 500000 00.5500 000. .0800 0000 >0 000.0..0 0.0»0. 00500000 05 000.0... 00.00..0> 5.0.0 .00500..>00 00000....0 00505.8... 5.00.... 0. 000 0.00 0.
.00 0.0..» 000.0> 0005 .0 .05.0 .05 5000.0... 00.0000» .500 000 .w0.>.. 0. .050. 0.000 .00 .0 .050...» 5...»: 0. .0505 0.000 .00 .0 .050...» 000.0... 00.0000»
5.0.0 00.00.5005. 5.00.. 000005000305. 5.00.00 .0 0.005... 000.000.. >0».00 05 00.03000 000 .0500? 000 5.3 .0 00.w0. .0580 .0...05 .000. .500

.0 .00.. 000.00. 0.0.0000 0.0005055 .0000500.00 05 0. 08.000. 0.0 0.0..0 0.00.5.0 .00.0.0000.>..0..0000000.0.0... 050.00 0000.00. 05 0. 0.03.00 5.00..

05 0. 0.0 000 .050...» 0. 0.00.0» .00000000 05 0.003 0.0005 50.0000... .00.... 0.0.0000 0. 0.00.0» 00000000 05 .0 0.00.0508 00.05500 0.0 03.003. ”0.02

 

00 > 07. 02 07. 00 > 07. 0 Z 07. .00500..>00
00000....0 .0. 000.0...

00 > 00> 0 Z 07. 00 > 00 > 07. 07. 00.00.5000...
5.00.. 5000500090....

.0. 00.00...

00 > 00 > 00 > 07. 00 > 00 > 00 > 0 Z 0.0....00 00.005055

.0. 00I0..0_ El

.008. a 0.0....

43

Table 5: Summary of Estimates of the Effect of Regressors on Selected Health Measures

 

 

Women Men
Rate health as Can walk a Rate health as Can walk a
very good or block with ease very good or block with ease
better better

Demo hic

characteristics:

black -.0687 .0008 -. 1798 -.0144
(.0237) (.0186) (.0211) (.0181)

Hispanic .0530 -.0248 -. 1477 .0023
(.0467) (.0354) (.0414) (.0337)

never married -.1399 -.0195 -.0459 .0318
(.0670) (.0446) (.0451) (.0258)

separated or

divorced -.1 150 —.05 19 -.0909 -.0649

(.0552) (.0384) (.0213) (.0172)

widowed -.0131 -.l 123 -.0429 -.0601
(.0305) (.0653) (.0261) (.0210)

born in south —.0444 -.0142 -.0748 -.0459
(.0320) (.0202) (.0289) (.0207)

born in midwest .0071 -.0028 -.0679 -.0145
(.0323) (.0198) (.0291) (.0202)

born in northeast .0243 .0037 -.0269 -.0072
(.0341) (.0205) (.0303) (.0210)

Family background

proxies (Health

transmission and

childhood

environment proxies):

mother alive .0161 .0275 .0728 .0361
(.0184) (.0116) (.0161) (.0116)

father alive .0318 .0041 .0287 .0322
(.0231) (.0345) (.0210) (.0141)

mother’s education .0012 .0011 .0042 .0012
(.0034) (.0024) (.0031) (.0025)

father’s education .0067 .0055 .0022 .0050
(.0030) (.0021) (.0028) (.0022)

height .4487 -.0267 .0475 -. 1546
(.1168) (.0762) (.1172) (.0889)

 

Note: Each column represents the results of a separate linear probability model where the
dependent variable is whether one is in the health category described at the top of the column.
Heteroscedasticity-consistent standard errors are reported in the parentheses. Omitted are the
effect of year of birth, the coefﬁcient estimates for whether one answered the survey questions
himself or herself, the coefﬁcient estimates for the dummy variables that denote whether an
observation had missing information on a particular variable, and the estimated effect of
education, which appears in Table 4.

...0 0. 00.0050» 5..0 .0 .0..000 00.5-5.5 0.0 0.05 .000... 00.5500 .00.. 00»0.0
5..» 00.00.05. 00.5500 5..0 .0 .0..000 00.5 0000.00. 00.0050» 30050.00. .0 .00 00... ...00.00 .0 ..005.0 .0500 05 00.0.»000 000 .0500». 000
.5..0 .0 00.w0. .0550 .0....05 .000. .050 .0 .00.. .0. 0.0.58 0.0.000 00000050000 am 000 30 50m .0000 0000 0. 0000 0. .0005 30.00050

 

.005. < 000050000 0. 0.5.0 0.00050 .00.0.0000-....0..0000000.0.00 5..» 03.000. 0.0 0.05.0» 00000000 05 .0 0.00.0.-..000 00.0500m. ”0.02

 

45

 

m. .. NN. . 0.0050500.

0000.000 05 .0 000000.030 0m0.0

.5... 05 .0 .00. 5.0.. 0 .0 00000.0-..
808.. .05.... .38.. 8.8..

00. :8. 008. R. 080. 008. 2:... a 0.. .0...
$08.. 808.. Ammo... 8.8..

00. .80. 38. mm. $8. .80. .0.»000 .0 0500
3.20.. 6.8.. .005. $30..

0... 0.50. v. 5. m0. 008. m2... V.00.0 0 .00.»
$.80 Ammo... 208.. $80..

0... .05. .wS. 0.. 008.- 88. 0000.0 .0 0000.. 000.0
ammo. Ammo... .88.. $80..

9.. mic. 0. mo. vm. No.5.- 0an. 10.0.0 .0 0.0m... 05:0

”M000 5..»-5.30:0. 05 00 00|U
.38.. $80.. $03.. 68....

0m. 0000. ammo. 00. 0.3. 3.8. .0..00 .0 50:08.0
6.8.. .0800 6.8.. 680..

8. .08.- 003. 3.. $8. 0.3. .20... 8 .080 be.
2 .8. .0800 .38.. .080.

8. 008. 008. N0. 003. 008. .0..00 .o 080

”I00IIIIIII5.000 030 0.0...

.00. b.000w000 .00. b.000m000
005000... 005000..
.0. 00.0»-.. 3mm mac .0. 00.0»-.. 3mm wAO
G02 :0803

 

5050.500— 00 00 0...... .0 .0..000d0.0D 8500.0O 5.00: 00 5050.33. 500.5003 .0 .00.... 0... .0 00.0.5.0...— mAmN .0 0.00.0

Table 7: TSIV Estimates of the Effect of Educational Attainment on Health

 

Women Men
OLS TSIV P-value OLS TSIV P-value
for for
Hausman Hausman
exogeneity exogeneity

Rate own health test test

as:

good or better .0380 -.3620 .00 .0336 .2262 .11
(.0025) (.1302) (.0023) (.1305)

very good or

better .0473 .0736 .87 .0435 -.0209 .70
(.0026) (.1558) (.0025) (.1772)

excellent or .0284 .3025 .14 .0252 -.2153 .02

better (.0026) (.1871) (.0024) (.1 149)

Can do the

following with

ﬁe:

climb ﬂights of

stairs .0359 -.3113 .04 .0314 .3244 .01
(.0028) (.1684) (.0023) (.1187)

stoop, kneel, or

crouch .0300 -.1527 .26 .0181 .2130 .10
(.0028) (.1636) (.0022) (.1199)

walk a block .0135 .0054 .94 .0114 .1032 .15
(.0018) (.1003) (.0016) (.0640)

bathe or shower .0051 -.0592 .32 .0025 -.0062 .82
(.0010) (.0646) (.0009) (.0391)

pick up a dime .0053 .0365 .58 .0036 .0228 .72
(.0010) (.0569) (.0012) (.0535)

heart attack -.0062 .1833 .00 -.0081 -.2042 .06
(.0013) (.0567) (.0018) (.1054)

 

Note: Reported are the estimates of the effect of education on health estimated by the two-sample
instrumental variables (TSIV) procedure. The expectation of educational attainment conditional
on quarter of birth is calculated using the 5% microdata sample of the 1980 Census. There were
628,848 women in the Census sample and 597,405 men. The conditional expectations of the
dependent variables (the health measures in the left-hand column) are calculated using the HRS
(sample size is the same as in the prior tables). Weighted least squares is used to obtain the
estimates. Year of birth dummy variables are included in each speciﬁcation.

46

Table 8: Alternative 2SLS Estimates Using Family Characteristics and Quarter of Birth as

 

 

Instruments, Women
Estimated P—value of Additional P-value of
Effect of Hausman instruments (P: overid. test on
Education exogeneity test parents’ excluded
education; instruments
S: sibling
information)
Rate own health as:
good or better .0485 .02 P,S .92
(.0054)
very good or better .0155 .22 P,S .20
(.0202)
excellent or better .0424 .01 P,S .75
(.0061)
Can do the following
with ease:
climb ﬂights of stairs .0407 .33 P,S .17
(.0063)
stoop, kneel, or crouch .0135 .47 S .74
(.0188)
walk a block .0233 .00 P,S .27
(.0034)
bathe or shower .0076 .13 P,S .92
(.0020)
pick up a dime .0063 .55 P,S .56
( .0021)

 

Note: Estimated coefﬁcients of the education variable are reported in the ﬁrst column with
heteroscedasticity-consistent standard errors in parentheses. A linear probability model is used in
each case. Each speciﬁcation includes as regressors year of birth dummy variables, race controls,
marital controls, region of birth controls, a dummy variable indicating whether or not one
answered the survey herself, height, dummy variables indicating whether each parent is alive, and
dummy variables indicating missing information. The set of instrumental variables always
includes quarter of birth. Additional instruments includes P (mother’s and father’s education)
and/or S (number of siblings, percentage of siblings that are female, a dummy variable denoting
that one is the youngest child, and a dummy variable denoting whether one is the middle child).
The F-statistic of a joint test of the ﬁrst-stage signiﬁcance of all of the possible excluded

instruments is 25.21.

47

Table 9: Alternative ZSLS Estimates Using Family Characteristics and Quarter of Birth as

Instruments, Men

 

 

Estimated P-value of Additional P-value of
Effect of Hausman instruments (P: overid. test on
Education exogeneity test parents’ excluded
education; instruments
S: sibling
informatiorg
Rate own health as:
good or better .0442 .01 P,S .57
(.0052)
very good or better .0278 .53 S .15
(.0160)
excellent or better .0320 .12 P,S .52
(.0057)
Can do the following
with ease:
climb ﬂights of stairs -.1119 .00 P,S .28
(.0152)
stoop, kneel, or crouch .0256 .09 P,S .63
(.0051)
walk a block .0171 .02 P,S .50
(.0029)
bathe or shower .0075 .00 P,S .55
(.0019)
pick up a dime .0083 .02 P,S .52
(.0023)

 

Note: Estimated coefﬁcients of the education variable are reported in the ﬁrst column with
heteroscedasticity-consistent standard errors in parentheses. A linear probability model is used in
each case. Each speciﬁcation includes as regressors year of birth dummy variables, race controls,
marital controls, region of birth controls, a dummy variable indicating whether or not one
answered the survey herself, height, dummy variables indicating whether each parent is alive, and
dummy variables indicating missing information. The set of instrumental variables always
includes quarter of birth. Additional instruments includes P (mother’s and father’s education)
and/or S (number of siblings, percentage of siblings that are female, a dummy variable denoting
that one is the youngest child, and a dummy variable denoting whether one is the middle child).
The F—statistic of a joint test of the ﬁrst-stage signiﬁcance of all of the possible excluded

instruments is 20.24.

48

 

88cc 83cc A353 85cc Qumooc Anon: .v €33 A323
:5. NmS. £8. 88. mmmo. vmoo. coco. 9 5. x83 a in?
A389 32 3 $3.09 Ammo“: $33 683 233 A393
88. oﬂm. 35. mm 5. mm 5. ram ﬂ .- 88.. come. 5:80 .8
485— .983.
Gm :3 Ck— 3 ammog Gmoog 383 3&2 .V Gong Ammooc
2 ~ 7 3mm. mic. £8. 83. m m _m.. 88.. 38. E83 no
acme €50
Holman 53,
7.50:8
05 ow 50
Cheqv 83 3 Sauna Gmoog 2093 CB _ .V 3R3 88¢;
88. mﬂmr wove. vane. 33. mmom. Eve. $8. .633
do 28:85
805; at: 333 Ammoog Among Awmﬂ .V 63.3 Ammooo
wnmo. 88... 580.. mums. mm g o. 35. ammo. move. .6on
.8 coow .02,
Ammoog Gem: A m 3o; 689V Genoa: Sam: Avvma ARR:
vao. memm. memo. mama. mwvo. omen; omvo. ammo. 3:3 co uoow
”ﬂag
3%
$58:me 3:2:ng
ma u? .83 30:5me 3 A? .68 .58:me
new 5.5 we 5:0 moBatg new 5.5 an 5.3 ave—nut?
we 55:: we BBS—e baa .6 85:: he horas—6 55a
.23 mqmm 2mg. 55 mama EW .23 mac 55 3mm amp. 5:» wqmm EN .23 30
:02 5:83

 

5.8: 8 838:3. .6 .85 2: ca 83.55 3.83 do bagm x: 2.3.

49

5038088.. .o 030.3. h8 :528 85.3 05 5:: w 0.8:. .8 2:28 5:... 05 808. 5:28 305.3 05 5:: 5:28
5.88 05. 5038088.. K 035. .8 8528 5.5.3 30:: 5:80.. 05 80.30: 05:28 5.3.. 5:: 3:35 05. 50380.38: .0 0.55. me 8528 55.3 5::
5:800 05 808: 8528 5E 5:: 5:80.. 0.3. 50500.38. .v 030,—. 50.: @ 5:28 5:0 Gav 5:28 808. 8528 5.80.. 5:: 8.5 0.5. .082

 

$803 Ammmoo 380.3 3.83 28.5 808.3 9802 2 .83
$8. wane. Zoo. 38. $8. memo. 08¢. 38. 055
a a: 088

8.83 23qu 3:83 8.83 889v Gvooq Amwoog C .83
poo. N08: 380. 3 So. 0:8. 83.. 88. 9.8. 830:0
.8 05.3

3.508 2 033.

50

REFERENCES

Anderson, K. and R. Burkhauser. 1984. “The Importance of the Measure of Health in
Empirical Estimates of the Labor Supply of Older Men.” Economics Letters 16(4): 375-
80.

Angrist, J. and A. Krueger. 1991. “Does Compulsory School Attendance Affect
Schooling and Earnings?” Quarterly Journal of Economics 106(4): 979- 1014.

. 1992. “The Effect of Age at School Entry on Educational Attainment: An
Application of Instrumental Variables with Moments from Two Samples.” Journal of the
American Statistical Association 87: 328-336.

. 1995. “Split-Sample Instrumental Variables Estimates of the Return to
Schooling.” Journal of Business and Economic Statistics 13(2): 225-235.

 

Ashenfelter, O. and A. Krueger. 1994. “Estimates from the Economic Return to
Schooling from a New Sample of Twins.” American Economic Review 84: 1157-1173.

Berger, M. and J. Leigh. 1989. “Schooling, Self-Selection, and Health.” The Journal of
Human Resources 24(3): 433-455.

Bound, J. 1991. “Self-Reported Objective Measures of Health in Retirement Models.”
Journal of Human Resources 26(1): 433-455.

Bound, J ., D. Jaeger, and R. Baker. 1995. “Problems with Instrumental Variables
Estimation When the Correlation Between the Instruments and the Endogenous
Explanatory Variable is Weak.” Journal of the American Statistical Association 90: 443-

50.

Bound, J. and D. Jaeger. 1996. “On the Validity of Season of Birth as an Instrument in
Wage Equations: A Comment on Angrist and Krueger’s ‘Does Compulsory Schooling
Attendance Affect Schooling and Eamings?’” NBER Working Paper No. 5835.

Card, D. 1993. “Using Geographic Variation in College Proximity to Estimate the
Return to Schooling.” NBER Working Paper No. 4483.

Cawley, J. 1998. “The Correlation Between Schooling and Mortality.” rnimeo,
University of Chicago.

Christenson, B. and N. Johnson. 1995. “Educational Inequality in Adult Mortality: An
Assessment with Death Certiﬁcate Data From Michigan.” Demography 32(2): 215-229.

E10, 1. and S. Preston. 1996. “Educational Differences in Mortality: United States, 1979-
1985.” Social Science and Medicine 42(1): 47-57.

51

Farrell, P., and V. Fuchs. 1982. “Schooling and Health: The Cigarette Connection.”
Journal of Health Economics 1: 217-230.

Feinstein, J. 1993. “The Relationship Between Socioeconomic Status and Health: A
Review of the Literature.” The Milbank Quarterly 71(2): 279-322.

Freedman, V., and L. Martin. 1999. “The Role of Education in Explaining and
Forecasting Trends in Functional Limitations Among Older Americans.” Demography
36(4): 461-473.

Fuchs, V. 1982. “Time Preference and Health: An Exploratory Study.” in Economic
Aspects of Health, V. Fuchs, ed. (Chicago: University of Chicago Press).

Griliches, Z. 1977. “Estimating the Returns to Schooling: Some Econometric
Problems.” Econometrica 45(1): 1-21.

Grossman, M. 1972. “On the Concept of Health Capital and the Demand for Health.”
Journal of Political Economy 80(2): 223-255.

House, J ., R. Landis, and D. Umberson. 1988. “Social Relationships and Health.”
Science 241: 840-845

House, J., R. Kessler, A. Herzog, R. Mero, A. Kinney, and M. Breslow. 1990. “Age,
Socioeconomic Status, and Health.” The Milbank Quarterly 68(3): 383-411.

Huntington, E. 1938. Season of Birth: Its Relation to Human Abilities (New York: John
Wiley).

Hyman, H., C. Wright, and J. Reed. 1976. Then Enduring Effects of Education
(Chicago: University of Chicago Press).

Kitagawa, E. and P. Hauser. 1973. Differential Mortality in the United States: A Study
in Socioeconomic Epidemiology (Cambridge, MA: Harvard University Press).

Lam, D., and J. Miron. 1987. “The Seasonality of Births in Human Populations.”
Research Report 87-114, University of Michigan: Population Studies Center.

Moen, E. 1999. “Education, Ranking, and Competition for Jobs.” Journal of Labor
Economics 17(4): 694-723.

Nagar, A. 1990. “The Bias and Moment Matrix of the General k-class Estimators of the
Parameters in Simultaneous Equations.” Econometrica 27: 575-595.

Newey, W. 1985. “Generalized Method of Moments Speciﬁcation Testing.” Journal of
Econometrics 29: 229-256.

52

Perri, T. 1984. “Health Status and Schooling Decisions of Young Men.” Economics Of
Education Review 3: 207-213.

Preston, S. and P. Taubman. 1994. “Socioeconomic Differences in Adult Mortality and
Health Status.” in Demography of Aging, L. Martin and S. Preston, eds. (Washington,
DC: National Academy Press).

Rosenzweig, M., and K. Wolpin. 1988. “Heterogeneity, Intrafamily Distribution, and
Child Health.” Journal of Human Resources 23(4): 437-461.

Ross C., and J. Mirowsky. 1989. “Explaining the Social Patterns of Depression: Control
and Problem-Solving - or Support and Talking.” Journal of Health and Social Behavior
30: 206-219.

. 1999. “Reﬁning the Association Between Education and Health: The Effects of
Quantity, Credential, and Selectivity.” Demography 36(4): 445-460.

Ross C., and C. Wu. 1995. “The Links Between Education and Health.” American
Sociological Review 60: 719-745.

Schultz, T. 1984. “Studying the Impact of Household Economic and Community
Variables on Child Mortality.” Population and Development Review 10: $215-$235.

Seeman, M. and T. Seeman. 1983. “Health Behavior and Personal Autonomy: A
Longitudinal Study of the Sense of Control in Illness.” Journal of Health and Social
Behavior 24: 144-160.

Shea, S., A. Stein, C. Basch, R. Lantigua, C. Maylahn, D. Strogatz, and L. Novick. 1991.
“Independent Associations of Educational Attainment and Ethnicity with Behavioral

Risk Factors for Cardiovascular Disease.” American Journal of Epidemiology 134(6):
567-582.

Smith J ., and R. Kington. 1997. “Demographic and Economic Correlates of Health in
Old Age.” Demography 34(1): 159-170.

Smith, J. 1998. “Socioeconomic Status and Health.” American Economic Review 88:
192-196.

Staiger, D. and J. Stock. 1994. “Instrumental Variables Regressions with Weak
Instruments.” NBER Technical Working Paper 151.

Strauss, J ., P. Gertler, O. Rahman, and K. Fox. 1993. “Gender and Life-Cycle
Differentials in the Patterns and Determinants of Adult Health.” Journal of Human
Resources 28(4): 790-837.

53

Thomas, D. and J. Strauss. 1997. “Health and Wages: Evidence on Men and Women in
Urban Brazil.” Journal of Econometrics 77: 159-185.

Wheaton, B. 1990. “The Sociogenesis of Psychological Disorder: An Attributinal
Theory.” Journal of Health and Social Behavior 21: 100-124 .

Willis, R. 1986. “Wage Determinants: A Survey and Reinterpretation of Human Capital
Earnings Functions.” in Handbook of Labor Economics, Volume I, O. Ashenfelter and R.
Layard, eds. (Amsterdam: North-Holland Publishing Company)

Winkleby, M., D. Jatulis, E. Frank, and S. Fortmann. 1992. “Socioeconomic Status and

Health: How Education, Income, and Occupation Contribute to Risk Factors for
Cardiovascular Disease. ” American Journal of Public Health 82(6): 816-820.

54

CHAPTER 2: SELF-REPORTED AGE DISCRIMINATION AND LABOR
MARKET OUTCOMES

Discrimination based on age is difﬁcult to observe. This is due to the fact that
labor market outcomes for older people are generally positive in comparison with the rest
of the population. For instance, the August 1999 Current Population Survey (CPS)
revealed that the median hourly wage for men ages 50-65 was $16.15, and the median
hourly wage for men ages 25—49 was $14.42. Moreover, the unemployment rate for men
ages 50-65 was 2.7%, while the unemployment rate for men ages 25-49 was 2.9%.
Other groups that claim that they are the victims of discrimination, such as women and
racial minorities, normally can point to such easily observable adverse labor market
conditions as at least a starting point to justify their claims. The aged cannot.

Nevertheless, the treatment of older workers has long been a topic of utmost
political interest, and there has been some evidence to support the concern. The most
commonly cited example of perceived ill treatment was the practice by many ﬁrms of
forcing retirement before workers were willing to depart voluntarily from the workforce.
Others included the preferences in promotion to younger workers displayed by some ﬁrm
managers (Rosen and Jerdee 1977) and the lack of labor market opportunities faced by
older workers that desired to either switch jobs or to get a new job after a layoff.
Symptoms of the latter have been veriﬁed empirically (Hutchens 1986, 1988).

Such perceived ills led Congress to enact the Age Discrimination in Employment

Act (ADEA) in 1967. The ADEA prohibited discrimination based on age for those ages

55

40-65. More speciﬁc concerns were addressed in 1978 when an amendment to the
ADEA increased the minimum allowable mandatory retirement age to 70 and, a year
later, when enforcement of the original ADEA was switched from the Department of
Labor to the Equal Employment Opportunity Commission (EEOC).l Prior to the late
19708, enforcement of the ADEA was limited at best (Johnson and Neumark 1997).
Recently, the ADEA has been used as the basis for many claims of age discrimination.
Between ﬁscal year 1992 and ﬁscal year 1998, the EEOC reported that there were a total
of 123,111 charges ﬁled under the ADEA (1999).2 This should be compared to the
207,191 charges ﬁled alleging race discrimination and the 170,751 charges ﬁled alleging
sex discrimination. While age discrimination has led to the fewest charges ﬁled among
these types of discrimination, there are still a very large number of charges given the fact
that a prima facie case of age discrimination is so difﬁcult to make.

Because the ADEA has become so vigorously enforced and is increasingly being
used as the basis for lawsuits, it is important to determine whether age discrimination
actually is a problem in current labor markets. This paper uses the Health and Retirement
Study (HRS) to look at the effect of one type of age discrimination-ﬁrm preferences in
promotion toward younger workers. While the effect of preferences in promotion is an
interesting question in itself, for the purposes of this paper it serves as an indicator of
ﬁrm discrimination in general. First, the paper veriﬁes whether individuals reporting age
discrimination are paid less than similarly trained and qualiﬁed workers. Second, the
paper investigates how the self-assessed probabilities of early retirement are affected by

ﬁrm discriminatory practices. Finally, workers are tracked to determine whether they

 

' Mandatory retirement was banned completely by another amendment to the ADEA in 1986.

56

ultimately are more likely to experience lower wage growth, to separate from their
employer, or to retire early, as a function of their ﬁrm’s propensity to promote younger
workers. The ﬁndings show that those reporting discrimination do in fact experience
some negative effects in the labor market. The most signiﬁcant negative effects are on
wage growth. Yet, the evidence also suggests that discrimination may affect individual
assessments of the probability of retiring early.3

In Section I, the relevant literature is reviewed. Section H describes the data used.
Section III outlines the various empirical approaches employed. Section IV summarizes

the results.

1. Related Work

As there has been an increase in the number of cases ﬁled with the EEOC using
the ADEA as a basis, the research that is relevant to this paper falls into two
categories-the pre-ADEA enforcement period and post-ADEA enforcement period.
Because enforcement of the ADEA had begun by 1979, the pre-ADEA evidence will
include all studies that use data prior to this year. Breaking the existing literature up in
such a fashion illustrates what little work has been done to address the impact of age

discrimination after the ADEA.

A. Pre-ADEA

The wide-sweeping provisions of the ADEA and its amendments naturally

 

2 Of these claims, 15,546 reached a merit resolution (12.6%). Monetary beneﬁts totaled nearly
$300 milllion (1998 dollars).

57

prompted economists to wrestle with the ideas of what discrimination on the basis of age
actually was, whether or not it led to adverse outcomes in labor markets, and why it may
have existed in the ﬁrst place. Because mandatory retirement was the most visible of
these “discriminatory” practices, it was the most often evaluated. Lazear (1979, 1981)
argued that mandatory retirement provisions were efﬁcient mechanisms through which
ﬁrms combated shirking and malfeasance among workers. This was because ﬁrms
offered workers wages below their marginal product during their initial years with the
company with the promise that good work would lead to wages above their marginal
product later on. Although the promises were frequently implicit, much evidence
supports the fact that ﬁrms behaved in such a fashion (e.g., Kotlikoff and Gokhale 1992;
Lazear and Moore 1984; Medoff and Abraham 1980). Given that workers might not
choose to retire if they were earning a wage greater than their marginal value of leisure,
ﬁrms had to enforce the implicit contract by requiring workers to retire at a certain age.
It was Lazear’s view that banning the use of mandatory retirement provisions would be
beneﬁcial to those who were nearing the retirement age. It would be detrimental,
however, to the long-run interests of the ﬁrms (who could no longer offer the implicit
contacts to reduce shirking) and workers (who would experience losses in lifetime
earnings because they would no longer have the incentive not to shirk).

Less attention has been paid to the effects of other aspects of discrimination that
led to the passage and enforcement of the ADEA. There have been some observations of
symptoms of a problem. The previously cited problems with job mobility (Hutchens

1988), with getting hired (Hutchens 1986), and with managerial preferences toward

 

3 The latter effect is quite important when viewed in a public policy context. If the aged are
driven into early retirement due to ﬁrm discriminatory practices, this has implications for social

58

younger workers (Rosen and Jerdee 1977) are examples.4 These papers, however, fall
short of showing a link between employer discrimination and adverse labor market
outcomes. Johnson and Neumark (1997) attempt to ﬁll this void in the literature by
showing that individuals who reported age discrimination were more likely to separate
from their employer and more likely to become non-employed. Moreover, they show
that when workers were rehired, they earned less than they did at their prior job. The
upshot of these results is that there appears to be some justiﬁcation for the legislation that
aimed to combat age discrimination and the subsequent amendments that strengthened its

enforcement.

B. Post-ADEA

As for investigations of age discrimination in the post-ADEA enforcement
period, the only concrete evidence that exists seems to assess the impact of banning
mandatory retirement. There has been considerable attention devoted to whether ﬁrms
have continued to behave in a fashion that is consistent with Lazear contracts.
Burkhauser and Quinn (1983) present evidence from the 1970s that retirement ages were
by no means solely determined by mandatory retirement provisions, as they show that the
impact of such provisions is cut in half when other factors, such as social security and
pension incentives, are taken into account. Others show that ﬁrm incentives for
retirement at certain ages have remained an important determinant of the retirement

decision (e.g., Kotlikoff and Wise 1989 and Gustman and Steinmeier 1986). Thus, it

 

security ﬁnancing.

4 Both Hutchens and Rosen and Jerdee use data primarily from before 1979. Hutchens uses the
National Longitudinal Survey of Older Men (NLSOM) and Rosen and Jerdee use their own
survey of ﬁrms.

59

certainly seems possible that ﬁrms have continued to offer incentive-laden contracts that
pay workers more than their marginal product in the post-ADEA period. Why would
ﬁrms continue to enter into such contracts if there is no mechanism in place to remove
older workers at a certain age? Some argue that the ADEA actually serves to strengthen
the bond between ﬁrm and worker, largely due to the fact that ﬁrms are prohibited from
laying off older workers before the age when retirement beneﬁts would be realized.
Thus, workers may be even more likely to enter into a long-term incentive contract.
Neumark and Stock (1999) present evidence using variation in state legislation limiting
age discrimination prior to the federal legislation that shows employment for older
workers increasing after the passage of the laws, an expected result of age discrimination
legislation. They also show that earnings proﬁles become more steep following the
passage of ADEA-type provisions, thus indicating that the ADEA might actually serve to
increase labor market efﬁciency by promoting the formation of Lazear contracts.5

Thus, there is little evidence of the major negative effect predicted to be the result
of age discrimination legislation (i.e., ﬁrms have not been discouraged from entering into
long-term incentive contracts). Given this, is there any evidence that the ADEA has had
any of its predicted good effects? Moreover, is there evidence that the increasing number
of claims ﬁled alleging age discrimination a response to real labor market problems faced

by older Americans? This paper is the ﬁrst attempt to answer these questions.

 

5 The alternative view of Lazear was that the elimination of involuntary layoffs would result in
ﬂatter earnings proﬁles since ﬁrms would not have the ﬁnancial incentive to enter into long-term

60

11. Data

As noted above, this paper uses the HRS, which is a national longitudinal survey
that asks respondents a very detailed list of questions on employment, health, retirement,
income, assets, and wealth, along with basic demographic and family information. Two
waves of the data will be utilized for the purposes of this paper. The ﬁrst wave,
conducted in 1992, asks questions of individuals born between the years 1931 and 1941
and their spouses.‘5 The second wave was conducted in 1994.

In the ﬁrst wave, individuals were asked a series of questions about their current
job. For one such question, respondents were asked whether they strongly agree, agree

somewhat, disagree somewhat, or strongly disagree with the following statement:

In decisions about promotion, my employer gives
younger people preference over older people.

The answer to this question is used in this paper as an indicator of age discrimination.7
The second wave is used to track individuals to see if they are more likely to experience
lower wage growth, to separate from their employer, or to retire early following a report
of discrimination.

There are advantages to using the question of discrimination posed above. First,
allowing respondents to agree or disagree with an actual workplace experience may be

more concrete than just asking a blanket question like Have you experienced age

 

incentive contracts.

6 Spouses of respondents who answered the survey but were out of the age range are excluded
from the sample, however.

7 This is a binary indicator that is set to one if an individual somewhat agrees or strongly agrees
with the statement.

61

discrimination ?. It provides workers an example of something in the workplace that they
may have observed. Second, the speciﬁcity of the managerial preference in promotion
question in the HRS allows a unique ability to determine whether this commonly cited
example of age discrimination actually leads to negative effects in the workplace.

There are disadvantages stemming from determining discrimination from the
above question. It is self-reported, so estimates of the effect of discrimination derived
from these self-reports may be biased due to systematic ways in which individuals report
discrimination (Kuhn 1987, 1990). Some of the approaches to deal with this problem,
which will be outlined in later sections, involve the use of HRS questions that reﬂect job
satisfaction and underlying labor force attachment. Also, it should be noted that no
where in the preference promotion statement above does the word "discrimination"
appear. Individuals are simply asked if they agree with a statement made about practices
within their company. Thus, some of the heterogeneity in reporting of discrimination
should be partially combated by this phraseology.

A second problem with the discrimination question in the HRS is that it may be a
bit narrow in scope. For instance, the NLSOM, which was used by Johnson and
Neumark (1997) to study the effect of self-reported discrimination in the pre-ADEA
enforcement period, allowed individuals to report age discrimination on the basis of
hiring, layoffs, or general company age discrimination, in addition to discrimination on
the basis of promotion and assignment.

The ﬁrst column of Table 1 presents descriptive statistics of the variables used in
the analysis of the paper. The sample is restricted to men working for a wage between

one and one hundred dollars in the ﬁrst wave of the HRS. In addition to workers born

62

outside the 1931 to 1941 period, the self-employed, agricultural workers, and those

lacking valid information in both waves are excluded from the sample.

III. Empirical Methodology
A. Earnings

There appears to be no study that has looked at whether those who have been ill
treated in the workplace based on their age have experienced lower earnings than other
workers. This is due to a number of factors; primary among these is that the methods for
ascertaining the existence of this kind of discrimination cannot be applied to older
individuals. Typical studies of sex (and race) discrimination rest on the assumption that
factors that affect productivity can be observed. Therefore, once controls are in place for

these factors, an equation like the following can be estimated:

(1) an=XB+yF+a

Individual subscripts are omitted. W denotes wages. X is a vector of basic
individual and job characteristics, which include factors related to productivity, such as
education and experience. F is a dummy variable for women (or minority group). The
estimate of y is typically interpreted as the estimated effect of discrimination on earnings.
Intuitively, this is the difference in earnings that persists between men and women once
factors that affect productivity have been taken into account. 8 is an error term.

The obvious but incorrect method to look at the effect of age discrimination on

earnings would be to replace F in equation (1) with a dummy variable for old age.

63

Johnson and Neumark (1997) note that this a mistake because of the way in which the
process of aging itself is a determinant of productivity. Productivity increases with
experience, but it then declines as physical and mental skills deteriorate (Hellerstein et al.
1999).

Another approach may be more suitable to investigating age discrimination. The
basic idea is to limit the sample to older individuals. An HRS sample of 51-61 year olds
seems appropriate. Then, identify among this group who is susceptible to discrimination.
To identify the susceptible, this paper uses whether or not an HRS respondent agrees with
the statement about his ﬁrm’s preference toward younger workers in promotion
decisions. Applying self-reports in such a fashion was done by Neumark and McLennan
(1995) to study discrimination on the basis of sex and by Johnson and Neumark (1997) to
study age discrimination in the pre-ADEA enforcement period. This paper picks up on
the general approach by estimating the following equation for a group of older

individuals:

(2) an = XB + yD + e.

X again consists of controls for basic individual and job characteristics, many of which
attempt to capture productivity differences among the older individuals that may account
for wage differentials. The remaining difference in wages is attributed to discrimination
(D).

Reading a negative estimate of y as a causal effect of discrimination is premature,

however, for a number of reasons. First, the reporting of age discrimination is

endogenous. Speciﬁcally, the fact that individuals have low wages may prompt them to
attribute it to some perceived ill treatment in the workplace, such as a company practice
of promoting younger workers over older workers. Second, as previously mentioned,
individuals have different propensities to report discrimination (Kuhn 1987). This may
be problematic if such propensities are systematically related to factors that determine
wages. For example, individuals who are very hard workers and devote much time to
work may be less likely to report discriminatory practices on the part of their ﬁrm. They
are more likely, however, to earn higher wages. Thus, this may bias the estimated effect
of discrimination toward ﬁnding a more negative effect on wages than is actually the
case.8'9 A ﬁnal problem applicable here, which was also noted previously, is that the
measure of discrimination used is quite narrow in scope. This may lead to
underestimating the effect of discrimination as a whole, since certain wage differentials

that are perhaps resulting from forms of discrimination not captured by the measure of

 

8 This need not be the direction of the bias, however. For instance, if more able individuals are
better equipped to detect discrimination at a ﬁrm, as well as earn more, omitting this ability factor
in the analysis could lead to a bias in the evidence in the direction of a positive effect of
discrimination on wages. This possibility is addressed by Kuhn (1990) in response to the
comment by Barbezat and Hughes on his 1987 paper. He ﬁnds that young, educated females are
more perceptive of statistical evidence of sex discrimination. Another important note to keep in
mind when estimating age discrimination’s effect on earnings is that the presence of Lazear
contracts may affect the interpretation of the results as well. This is because ﬁrms that enter into
such long-term relationships are probably more likely to promote workers that are younger than
those in the 51-61 age range. After all, the older workers should already be in the reward phase
of their worklife, and further promotion of such workers would only prompt them to remain with
the ﬁrm longer. Moreover, relatively high wages for older workers is a feature of delayed
payment contracts (Lazear 1981). For these reasons, the negative effects of discrimination on
wages that are estimated later may be understated.

9 Given that this paper follows a similar approach to Johnson and Neumark (1997), it is worth
noting that they combat the above problems mainly through the use of multiple self-reports of
discrimination. Johnson and Neumark (1997) also had to address two other problems that this
paper does not face. First, their discrimination question asked about any ill treatment based on
age in the prior ﬁve years, thus leaving it unclear whether the question referred to the current
employer. The HRS question used in this paper is speciﬁc to the current employer. Second, they
needed to worry about workers who reported discrimination merely due to the existence of

65

discrimination in this paper are attributed to the random error in equation (2), not as an
effect of discrimination. Thus, the results of the paper should be interpreted with this in
mind.

There are ways to mitigate the problem of the differential in the propensity to
report discrimination. First, HRS respondents provide answers to questions that serve to
reveal (at least partially) their propensities to report discrimination. These can be added
to the X vector. Also, the HRS asks respondents whether they strongly agree, agree

somewhat, disagree somewhat, or strongly disagree with the following statement:

My pay is fair considering what other people in this

line of work are paid.
The answer to this question could go far in attempts to control for individual
heterogeneity in reporting discrimination. Unfortunately, adding these extra controls may
also further narrow how much workplace discrimination the promotion question is
picking up. Thus, adding these controls may further bias the estimates away form ﬁnding

negative effects of discrimination.

B. Wage Growth

Another potential negative effect of discrimination is that those who experience it
may have stunted wage growth in their latter years in the workforce. This was implied by
the work of Rosen and Jerdee (1977), who argued that managerial preferences toward

younger workers were decreasing the opportunities of older workers. Moreover, Johnson

 

mandatory retirement provisions. Mandatory retirement was banned on the federal level in 1986
and is not an issue here.

66

and Neumark (1997) ﬁnd that those who separated from their employer due to
discrimination experienced lower wage growth than other workers. Those who stayed
with the same employer, however, experienced no such ill effect. Finally, although their
sample consists of young workers, Pergamit and Veum (1999) show that there are many
positive consequences to promotion, such as an increase in supervisory responsibilities
and the receipt of training. These consequences may certainly translate into greater wage
growth.

A variety of approaches are used in this section to test whether age discrimination
affects wage growth. The simplest approach is to divide the sample into two groups-
those who report discriminatory practices and those who do not. Then, the returns to

experience are estimated for each group via the following equation:'0

(3) an = x3 +8EXP + 1.1.

Differences in the estimates between those reporting discrimination and those not
reporting discrimination will be evidence of the effect of discrimination on wage growth
for older workers in the population.“

This approach is simple, but it is not without problems. The main drawback is the
rather restrictive assumption that there is no systematic difference between the

determination of wages across age groups that is correlated with the decision to report

 

'0 This approach is similar to that employed by Lazear and Moore (1984) to compare slopes of
earnings proﬁles of self-employed workers and non-self-employed workers as a means to test for
the existence of Lazear contracts.

1' Experience is measured as potential years of experience. Speciﬁcally, this is age - years of
completed schooling - 5. Obviously, this is just a rough approximation of actual experience.

67

discriminatory practices at the ﬁrm. Speciﬁcally, cohort effects may exist for which
there exists no means to control. Thus, the best that can be done is to include as part of X
controls for job satisfaction or other factors that may change across the cohorts. The
HRS does afford an opportunity to do this, as it contains information that reﬂects job
satisfaction.12

Another approach can be used that does not present the same difﬁculties with
cohort effects. Utilizing the longitudinal feature of the HRS, workers that report

discrimination are followed into the second wave. The following equation is estimated:

(4) PCW = XB + yD + a.

PCW is the percentage change in an individual’s hourly wage from the ﬁrst wave to the
second. As with estimating the effect of age discrimination on earnings, there may be a
problem here with workers reporting discriminatory practices due to dissatisfaction with
their job. If the low wage growth is related to the reason for that dissatisfaction, then D is
not exogenous in the above speciﬁcation.13 Proxy variables are also used here to control

for general job satisfaction and workforce attachment.

 

However, for the purposes of this exercise, an estimate of the wage proﬁle in the latter stages of
the worklife is what is desired. Thus, the rough approximation is appropriate.

'2 The presence of Lazear contracts may again be relevant here. If those workers engaged in such
implicit contracts are more likely to perceive ﬁrm practices as discriminatory, the lower wage
growth in old age that is associated with Lazear contracts may be unduly attributed to a
discriminatory effect.

'3 Biases may go in both directions, however. It is possible that the dissatisfaction with one’s job
may be due to lower current wages. Then, the bias could be in the direction of ﬁnding positive
wage growth effects of discrimination. One approach that may help to deal partially with this
problem is to control for as many factors as possible that may affect an individual’s wage in the
ﬁrst wave. Thus, dummy variables denoting an individual’s ﬁnancial situation are also included
in the speciﬁcation. Speciﬁcally, the HRS contains extensive information about the wealth of
individuals, including detailed information on asset holdings. Thus, a series of dummy variables

68

C. Employer Separation

Employer separation is of concern because, as noted earlier in the paper, older
individuals have historically had more trouble ﬁnding work than younger workers
(Hutchens 1986, 1988). Also, the consequences for older workers that separate from
their employer, especially involuntarily, have worsened over the past several decades.
These consequences include a failure to ﬁnd employment elsewhere and, for the others
who are lucky enough to ﬁnd employment, a failure to obtain a wage equal to that earned
prior to the job separation (Polsky 1999). This paper seeks evidence on whether ﬁrm
discriminatory promotion practices are leading to these job separations.

As with determining the effect of age discrimination on wage growth, determining
its effect on employer separation requires that the problem of heterogeneity in reporting
of discrimination be addressed. The approach here is similar to the methodology used by
Johnson and Neumark (1997) to determine whether employer separation, as well as some
other changes in job status, can be the result of discriminatory practices on the part of
ﬁrms. The idea is that each individual, at any given time, has a probability of separating
from his employer. Obviously, for some individuals that are very attached to their
employer, this probability is almost zero. Yet, one can imagine a simple two-period

latent variable model,14 which is reﬂected in the equation:

(5) 82* = X13 + yD1+ e.

 

are constructed that indicate various levels of an individual’s household’s total level of liquid
assets, and these are added to the speciﬁcation. Results do not change substantially if level of
household net worth is used instead.

69

82* is an individual’s propensity to leave his or her employer in the second period. X. is
a vector of ﬁrst-period individual characteristics and employer characteristics that may
affect the second-period propensity to separate. D1 is whether an individual has reported
discriminatory practices at his ﬁrm. 8 is an error term. In the HRS, as in most data sets,
there is no way to observe 82*. Instead, only whether or not an individual actually leaves
his employer in the second period is observed. Thus, I estimate the parameters in
equation (5) imposing the assumptions of a standard logit model.

For the same reasons as in the prior subsections, the estimate of 7 should not be
read as a causal effect of discrimination due to differing propensities to report
discrimination. Inclusion of job satisfaction and workforce attachment controls will

again be used to deal with these problems.

D. Early Retirement

To assess the effect of discrimination on the retirement decision, the paper uses a
two-fold approach. In the ﬁrst wave of the HRS, an individual is asked to rate on a scale
of zero to ten the chances that he will be working at age 65. The question is repeated
with age 62 as the reference age. On the scale, zero denotes absolute certainty that the
respondent will not be working at the age of reference, while ten denotes absolute
certainty that he will be working. The responses are used as a measure of the self-
assessed chance of remaining in the workforce until age 65 (or age 62). They serve as the

dependent variable (R) in separate estimates of the following equation:

 

” This paper will track individuals into the second wave, so the interpretation of the second
period can be taken to mean anything that is reported after the ﬁrst-wave interview date.

70

(6) R=XB+yD+e.

X is still the vector of individual and job characteristics that has been used throughout the
paper. D is the same discrimination measure. 7 is interpreted as the estimated effect of
discrimination on the self-assessed chance of working at 65 (or 62).” Thus, a negative
coefﬁcient estimate shall be interpreted as a decrease in the probability of working at the
age of reference. It can also be interpreted as an increase in the probability of early
retirement.16

There are several drawbacks to this approach. First, R is a self-reported
assessment of the future. Thus, it does not reﬂect events that have actually occurred.
Interpretation of the results should keep this in mind. Nevertheless, the estimates reveal
more information than was previously known about the effect of ﬁrm discriminatory
practices on the retirement decision. Second, unobserved factors related to the propensity
to report discrimination may be related to the subjective probability of working at age 65 .

For instance, the previously discussed problem of the disgruntled worker that is

 

'5 The equation (6) parameter estimates presented in the paper come from an OLS model, with
heteroscedasticity-corrected standard errors.

‘6 Alternatively, an ordered logit model was employed and the results were similar. Moreover,
given the fact that interpretation of the numbers on the scale may differ across respondents, a
variety of approaches using categorical variables derived from the scale were used. For example,
the 0-10 scale was broken down into 0—3, 4-6, and 7-10 ranges. This was done because a large
number of individuals responded with O, 5, or 10. This may indicate that many were unsure of
the exact spot on the scale they fell, so they just chose one of the extremes or the middle of the
scale. Thus, to determine the sensitivity of the results to this possibility, many different versions
of the scale were constructed and used as the dependent variable in an ordered logit estimation
procedure. The estimates obtained from these sensitivity analyses were not substantially different
than those obtained from the OLS approach.

71

experiencing negative labor market outcomes and blaming it on discrimination is also
applicable here.

In addition to estimating equation (6), an approach to estimating the effect of
discrimination on early retirement like the one used to look at employer separation is
used. Speciﬁcally, individuals are tracked into the second period to see if they retire
early after a report of discrimination. A logit model is then used to estimate the effect of
discrimination on the probability of retirement. This approach has as a drawback the fact
that individuals are only followed for one period (i.e., into the second wave). Thus, all of

the respondents are younger than 65.

IV. Results

Table 1 of the paper presents some descriptive statistics for the sample broken
down by whether or not an individual reports discrimination by his ﬁrm. There are some
entries that warrant mention. First, those reporting discriminatory practices report higher
tenures. Attributing great meaning to this would be purely speculative, but it still reﬂects
that there is a correlation between tenure and whether one thinks that his company
engages in discriminatory practices.17 In addition to being related to the reporting of
discriminatory practices, tenure is expected to affect the labor market outcomes that are
analyzed in this paper, such as earnings (Topel 1991) and job separation (Idson and
Valletta 1996) . Thus, it appears critical to include this variable in any analysis of the

effect of company policies toward older workers.

 

‘7 Two possible explanations immediately come to mind. First, those with higher tenures have
been with the ﬁrm long enough to be able to observe company practices or to have been passed
over at some point for promotions themselves. Second, those with higher tenures are more likely

72

Second, the table indicates that workplace satisfaction is strongly associated with
the reporting of discriminatory practices. For example, individuals reporting
discrimination are less likely to report working in a friendly environment and more likely
to report being paid unfairly. Thus, these seem to be important proxy variables for the
propensity to report discriminatory practices.

Finally, some job characteristics seem to be more common among those reporting
discrimination. For example, those who claim that their ﬁrm shows preferences in
promotion toward younger workers also are enrolled in pension plans in greater numbers.
This seems consistent with the point discussed earlier concerning Lazear contracts.
Pension plans are perhaps one tool used by ﬁrms to continue the use of implicit delayed
payment contracts with workers. Thus, workers who are enrolled in pension plans may
also observe other practices at their ﬁrm consistent with these implicit contacts, such as
preferences in promotion shown toward younger workers. Thus, inclusion of the pension
variable in the speciﬁcations may help control for unobserved factors related to

employer-employee relationships.

A. Earnings

Using the ﬁrst wave of the HRS, equation (2) is estimated. The results are
reported in Table 2. First, only the basic job and individual controls are used. These
include all of the variables in Table 1 that are not listed in the left-hand column of Table
2. Also included is a quadratic in age measured in years. Column (1) contains the

results. The estimated effect of a ﬁrm promoting younger workers over older is a 6.2%

 

to work at ﬁrms where Lazear contracts are employed, thus making it more likely for them to
report discriminatory practices.

73

reduction in wages. The effect is signiﬁcant at the .10 level of signiﬁcance. Note,
however, that important controls for job satisfaction and workforce attachment are
omitted.

Equation (2) is next estimated with characteristics of one’s job and labor force
attachment added in order to capture the effect of some factors on wages that may also be
a determinant of one’s propensity to report discrimination. The coefﬁcient estimates of
these variables are also included in column (2). The estimated effect of a ﬁrm promoting
young workers ﬁrst is now a .1% reduction in wages, but this negative effect is no longer
signiﬁcant.18

Finally, in column (3), industry and occupation controls are added to capture
differences in propensities to report discrimination across industry and occupation that
may be related to differences in wages. Estimated effects remain negative but not

signiﬁcant.19

B. Wage Growth

Table 3 presents the results of the estimation of equation (3). The ﬁrst column
gives the estimate of the returns to experience (8), with the standard error in parentheses,

for the group of respondents reporting no discrimination. The second column reports the

 

'8 It is worth mentioning that the estimated effects of the other variables in the table, which are
interesting questions in their own right and perhaps a subject for further investigation, are in the
expected direction. For instance, thinking that one is paid unfairly has a negative effect on
wages. The experience and schooling required for a job affect wages positively. The fact that a
person’s health may limit his ability to work in the future has a negative and signiﬁcant effect on
wages.

'9 Given that one was considered to have reported a discriminatory practice if he somewhat
agreed or strongly agreed with the statement about his ﬁrm's discrimination, the results in Table 2
were estimated when the discrimination indicators required strong responses to the question

74

estimates for workers reporting that their ﬁrm promotes younger workers ﬁrst. The ﬁnal
column reports the p-value of a Wald test of the restriction that the estimated returns are
equal. The various rows add control variables as indicated.

The estimated returns to experience suggest that those workers in ﬁrms that
discriminate based on age experience lower wage growth in the latter stages of their
worklife than workers in ﬁrms that do not discriminate. The difference is signiﬁcant (p-
value = .061) when only basic individual and job controls are included. As attempts are
made to control further for differences in the job and workforce attachment that may
affect the propensity to discriminate, the difference in the returns to experience becomes
somewhat smaller.

When the full set of controls is included, which includes industry, occupation, and
asset level controls, the estimated experience-eamings proﬁles still reveal lower wage
growth for workers in companies reporting ﬁrm discriminatory practices, and the
difference in the slopes is signiﬁcant at the .10 level.

While this is an interesting ﬁrst look at discrimination’s effect on wage growth,
there may be cohort effects that bias the estimates. Thus, utilizing both waves, equation
(4) is estimated to avoid this problem. In Table 4, the estimates are reported for those
working for an hourly wage between one and one hundred dollars in both waves. A
further sample restriction is added to cut down on the inﬂuence of outliers. All

individuals whose wage is more than double what it was in the ﬁrst period are excluded.20

 

posed. The results are not remarkably different, if anything suggesting a stronger negative effect
on wages. Still, the results cannot be judged statistically different from zero from those estimates.
20 Failure to impose this restriction results in negative effects of discrimination on wage growth
that are too large to be sensible. Removing several outliers results in estimates that are far more
reasonable. Given that those receiving huge wage gains (and, thus, removed from the sample)

75

With only individual controls included, the estimated effect on wage growth is negative
and signiﬁcant at the .05 level of signiﬁcance. Speciﬁcally, the promoting of younger
workers over older workers is estimated to reduce the percentage change in wages across
periods by 3.9 percentage points. Moreover, as controls for labor force attachment, job
satisfaction, wage satisfaction controls, and industry and occupation controls are added in
turn, the effect of discrimination remains roughly the same.

Again, controls for the level of household liquid assets are added. When this is
done, the effect of age discrimination on wage growth becomes smaller, but it is still
statistically signiﬁcant at the .10 level. When the full set of controls is added (including
industry, occupation, and asset controls), the promotion of younger workers over older
workers at an individual’s ﬁrm reduces the growth in wages across periods by about 4

percentage points.21

C. Employer Separation

Next, the effect of discrimination on employer separation is explored. A logit
model is used, and the results are reported in Table 5. An individual is considered to
have separated from his employer if, at any point during the period between wave 1 and
wave 2, the individual did not work with the employer with whom he worked in the ﬁrst
wave. Thus, individuals that returned to the same employer, switched employers, or were

not employed at all as of the second wave interview are considered to have separated

 

also tend to report discriminatory practices by their ﬁrm, however, the results reported in the table
may be bias toward ﬁnding no effect.

2’ When the effect of discrimination on wage growth was reestimated with the discrimination
indicator requiring a strong response to the promotion question, the cross-sectional results
revealed perhaps even a stronger negative effect of discrimination on wage growth and the results
using the two waves revealed a weaker effect.

76

from their employer.

When none of the job satisfaction and workplace attachment proxy variables are
included, the effect of discrimination on employer separation is positive, but it is not
nearly signiﬁcant. As additional controls are added, the table shows that the effect of
ﬁrm preferences in promotion remains essentially zero.

Remember, however, that Johnson and Neumark (1997) found that age
discrimination positively affected the probability of job separation prior to the
strengthening of the enforcement of ADEA provisions. The difference in the results may
be for a variety of reasons. First, the measure of discrimination used in this paper is
based on promotions. Thus, the absence of an effect seems consistent with the work of
Pergamit and Veum (1999), who show that promotions have little direct effect on job
attachment. Johnson and Neumark’s deﬁnition of age discrimination went far beyond
just ﬁrm discriminatory practices in promotion, as it pertained to hiring, layoffs, and
general age discrimination, as well as promotion. Thus, these other aspects of age
discrimination may be what are leading to the job separations. Second, only two waves
of the HRS are used in the analysis of this paper. Thus, the period of time in which
individuals can respond to discrimination by separating from their employer is limited to
just two years. Johnson and Neumark observed individuals over a much longer period of
time. A ﬁnal possibility may be that this paper uses data from the post-ADEA
enforcement period, as opposed to the data of Johnson and Neumark that comes from
primarily before 1979. Thus, this paper may be providing the ﬁrst bit of evidence that
suggests the ADEA may be combating age discrimination in the workplace through the

reduction of the employer separations that were observed by Johnson and Neumark. In

77

the future, subsequent waves of the HRS will be available that will allow a more
thorough investigation into these alternatives, as individuals will be tracked for many

years.

D. Early Retirement

Next, the two approaches outlined in the prior section of the paper are used to
obtain evidence on the effect of discrimination on early retirement. In Table 6, the results
from the estimation of equation (6) are presented. The top panel of the table presents the
results for the question that pertains to the probability of working at age 62. Remember,
individuals are asked to respond on a scale of O to 10. The results then suggest that ﬁrm
practices of promoting younger workers reduces an individual’s self-assessed chance of
working at age 62 by .4523 of a point on the scale when only basic controls are
included.22 The effect is almost signiﬁcant at the .10 level. When the full set of proxy
variables, industry and occupation, and asset level controls are added, the effect is a .5877
reduction on the scale. This effect is signiﬁcant at the .05 level. These results suggest
that discrimination in promotion enhances the chance of early retirement.

Similar results are reported in the bottom panel of Table 6 for a reference age of
65. Here, the estimated reduction in the chance of working at age 65 on the ten-point
scale ranges from .3796 of a point to .4782 of a point. When all controls are added, the
effect is signiﬁcant at the .05 level.

Finally, in Table 7, the results using both waves of the HRS to estimate a logit

model are presented. The estimated effects of discrimination on the probability of being

 

22 Another way to state this is that for workers whose company discriminates, there is about a 4.5
percentage point decline in the chance that they will be working at age 62.

78

retired in the second period appear in the ﬁrst row. Regardless of the controls used in the
speciﬁcation, the results indicate that the effect of discrimination on the probability of
being retired in the next period is essentially zero. Note, however, that the same
problems with interpreting the effect of age discrimination on employer separation are
applicable here.

Although Johnson and Neumark (1997) follow individuals for a much longer
period of time and well into the age range when individuals normally retire (age 65 and
older), they report very similar results to those reported in Table 7. The estimated effects
of ﬁrm discriminatory practices on the self-assessed chance of early retirement reported
in Table 6, however, reveal that there may be some effect of discrimination on early
retirement. More work needs to be done to determine adequately how well these self-
assessed probability scales serve as indicators of early retirement. The ability to follow
individuals beyond the age of retirement will be a necessity. Unfortunately, the data do
not allow such an investigation at this time. Nevertheless, the approach of this paper
serves as a good starting point for future investigations into the effect of age

discrinrination on retirement.

V. Conclusion

For quite some time, there have been stories advanced that suggest individuals are
being passed over for promotion on their job because they are too old (Rosen and Jerdee
1977). Next to mandatory retirement, this has probably been the most visual form of
alleged age discrimination. Passage of the ADEA in 1967 marked the beginning of a

several decades-long legislative and judicial process that has sought to reduce the practice

79

of age discrimination in the workplace. An outlet has been afforded to those that think
that they have been wronged on the basis of their age. Despite the fact that age
discrimination is difﬁcult to detect and a prima facie case of it is so difﬁcult to make,
recent experience suggests that these outlets are being utilized almost as much as those
claiming race and sex discrimination are using similar outlets.

All of this attention was the motivating factor behind this paper. In the paper,

individuals whose employer engages in one form of age discrimination-promotion of

younger workers before older workers—are observed. First, the paper veriﬁes whether
individuals reporting discrimination are paid less than similarly qualiﬁed workers.
Second, the paper investigates how the self-assessed probabilities of early retirement are
affected by ﬁrm discriminatory practices. Finally, workers are tracked to determine
whether they ultimately are more likely to experience lower wage growth, to separate
from their employer, or to retire early.

The results are mixed, but they certainly do indicate that there are at least some
negative effects experienced by those who claim that their ﬁrm engages in age
discrimination. The most signiﬁcant negative effects are on wage growth. Yet, the
evidence also suggests that discrimination may drive some workers to retire early, as
revealed by the effect of age discrimination on their self-assessed chances of being in the
workforce at advanced stages of their worklife. This paper appears to indicate a
continuing need for enforcement (and perhaps even more vigorous enforcement) of the

ADEA in light of adverse labor market outcomes experienced by older individuals.

80

Table 1: Descriptive Statistics by Whether or Not Discriminatory Practices are Reported

 

All workers Firm does not Firm promotes
promote younger
younger workers ﬁrst

workers ﬁrst

 

Job/workplace characteristics:

Enrolled in a pension plan .694 .681 .759
Experience required for job (in years) 3.55 3.54 3.59
Pay is unfair .238 .224 .311
Schooling required for job (in years) 11.96 12.06 11.43
Tenure (in years) 13.11 12.54 16.06
Union member .288 .268 .390
Workplace environment is friendly .884 .895 .827

Individual characteristics:

Age 55.91 55.94 55.76
Black .088 .086 .099
Hispanic .032 .030 .041
Never married .042 .041 .044
Separated or divorced .110 .112 .100
Widowed .018 .018 .017
High school graduate .340 .346 .310
Some college .205 .197 .245
College graduate .249 .246 .263
Health will limit work in the future .501 .490 .554
Sample size 1550 1293 257

 

Note: Reported are the mean proportion or mean level of the variable in the left-hand column,
conditional on whether the respondent ﬁts into the category at the top of the column. The means
of the variables for which there is missing information are calculated using fewer observations.
All means are weighted for non-response.

81

Table 2: Effect of Discrimination on Wages

 

(1) (2) (3)
Discriminatory effect -.O621 -.0009 -.0199
(.0341) (.0291) (.0394)
Believes that one is paid unfairly -. 1751 -.1396
(.0274) (.0246)
Workplace environment is friendly .0682 .0488
(.0351) (.0310)
Experience required for one’s job (years) .0350 .0229
(.0048) (.0039)
Schooling required for one’s job (years) .0325 .0206
(.0032) (.0033)
Health will likely limit work in the future -.0400 -.0284
(.0249) (.0229)
Industry and occupation controls included No No Yes

 

Note: Reported are coefﬁcient estimates for the variable in the left-hand column, with standard
errors reported in parentheses. The dependent variable is log wages. Also included in each
speciﬁcation are controls for the individual and job characteristics listed in Table 1 not appearing
in the above table, as well as a dummy variable indicating a proxy answered the survey and a
series of dummy variables for cases where missing data were coded to zero for a particular
variable. Age in years enters the speciﬁcations as a quadratic and tenure enters as four dummy
variables indicating fewer than four years, four to ten years, ten to twenty years, and more than
twenty years of tenure. All results are weighted for non-response. The number of observations is

1550.

82

02-0... .38 08:8... 8.03.8 2.8.V .05....
s. N 2.5 a. 8.8 80 .805 5.... 8.8.0 0.0 .805 u .85 08.8 .. .800 .880 u .8.... .08.... n .80 .800 .. .8. .8... .. . .o
A0880 2. 0.08. .080 8.30:0. 0... 83.00.02 850.00.. 5.500 250 .0 8.80 0 8 82.00.3080. 0... 08.8 .5520 .080 0.0.... 0:... 00.00.02
8 00000 0.0 0.0.200 .88...:0..0.. 2 0.0.8 0:00:08 .23 Am. 8000.00 2 80.0508 00889.0 0... .0 8.0.2.8 0... 0.0 00:08”. ”0.07.

 

 

.08.. .88.. 000202 0.0.200 .080

8.. 88.- 38. 0:0 02.00800 0:0 5.80:.
.88.. .0m8g 000202 0880

mwm. ..8. 88. 02.... .0 .5550 .0.. 20:80
88.. 8m8.. 000202

8. 8.- 28. 0.00:8 8009.000 0:0 5.80:.
000202 802.808

.88.. 88.. 00.0.2.0? 0:0 8.80.8000

mm .. 8. .N8. 00.. .0.. 8.00..0> 5.0.x.
.88.. .88.. :8

.8. 88.- 0N8. 883.80% 2 20:80 0.00m.

.0810 0.0 0.5.8.. 0.02.2820 0.02.2820 .0:

00.0.5.8 0... .0... 8.8.0.8..
0... .0 .8. 0.03 0 .0 020>-n.

800 .0... 2:... 0 2 8....03 0 .0
008.898 0. 0.5.8.. 00.0.2.8.

800 .0... 5.... 0 2 80.83 0 .0
008.898 0. 0:88.. 00.0200”.

 

0.58m .0.—0.800-880 .080...— mw:.:..0m.-00:0.8&.m 0... :0 80.80.... b30552029 .0 80......— .n 030,—.

83

Table 4: Effect of Discrimination on the Percentage Change in Wages Across Periods

 

(1) (2) (3) (4) (5)
Discriminatory effect -3.949 -4.181 -4.461 -3.700 -3.975
(1.901) (1.981) (2.060) (1.966) (2.029)
Believes that one is paid unfairly 2.986 3.237 2.688 2.934
(1.894) (1.898) (1.911) (1.915)
Workplace environment is friendly -1.1 17 -1.122 -1.284 -1.l99
(2.925) (2.939) (2.929) (2.961)
Experience required for one’s job (years) 0.185 0.128 0.187 0.129
(0.202) (0.218) (0.198) (0.213)
Schooling required for one’s job (years) 0.159 0.119 0.149 0.117
(0.191) (0.225) (0.195) (0.224)
Health will likely limit work in the future 0.547 1.167 0.854 1.527
(1.689) (1.743) (1.722) (1.779)
Additional controls:
Industry and occupation N o No Yes No Yes
Liquid asset levels No No No Yes Yes

 

Note: Reported are the estimated percentage point changes in wage growth across periods. The

sample size is 1166. See notes for Tables 2 and 3 for further details.

84

Table 5: Estimated Marginal Effects on the Probability of Employer Separation from a

 

 

 

Logit Model
(1) (2) (3) (4) (5)
Discriminatory effect .0033 -.0000 .0005 .0007 .0017
(.0300) (.0304) (.0306) (.0305) (.0307)
Believes that one is paid unfairly -.0379 -.0300 -.0352 -.0282
(.0262) (.0264) (.0263) (.0264)
Workplace environment is friendly -.0312 -.0259 -.0332 -.0274
(.0335) (.0335) (.0336) (.0335)
Experience required for one’s job (years) .0005 .0004 -.0000 .0001
(.0028) (.0029) (.0029) (.0030)
Schooling required for one’s job (years) -.0008 .0001 -.0011 .0002
(.0027) (.0030) (.0027) (.0030)
Health will likely limit work in the future .0279 .0230 .0285 .0236
(.0223) (.0225) (.0224) (.0226)
Additiomontrols:
Industry and occupation N o No Yes N 0 Yes
Liquid asset levels No No No Yes Yes

 

Note: Reported are estimates of the marginal effects (evaluated at sample means) for the variable
in the left-hand column, with standard errors reported in parentheses. See notes for Tables 2 and

3 for further details.

85

Table 6: Effect of Discrimination on the Self-Assessed Probability of Working at Ages 62

 

 

and 65
(1) (2) (3) (4) (5)
At age 62:
Discriminatory effect -.4523 -.4572 -.4751 -.5558 -.5877
(.2898) (.2897) (.2934) (.2916) (.2960)
Believes that one is paid unfairly .3626 .4356 .3394 .4085
(.2499) (.2527) (.2512) (.2532)
Workplace environment is friendly -.2437 -.2l78 -.l732 -.l468
(.3171) (.3216) (.3195) (.3243)
Experience required for one’s job (years) .0182 -.0018 .0268 .0047
(.0285) (.0286) (.0303) (.0305)
Schooling required for one’s job (years) .0690 .0480 .0786 .0537
(.0283) (.0313) (.0285) (.0313)
Health will likely limit work in the future .2742 .3272 .2054 .2545
(.2142) (.2165) (.2137) (.2143)
At age 65:
Discriminatory effect -.3979 -.3796 -.3886 -.4544 -.4782
(.2277) (.2297) (.2308) (.2290) (.2313)
Believes that one is paid unfairly .2597 .3259 .2271 .3958
(.2121) (.2123) (.2138) (.2137)
Workplace environment is friendly -. 1349 -.1561 -.0738 -.0962
(.2768) (.2807) (.2808) (.2855)
Experience required for one’s job (years) .0177 .0016 .0253 .0064
(.0217) (.0223) (.0232) (.0238)
Schooling required for one’s job (years) .0658 .0396 .0784 .0476
(.0229) (.0253) (.0234) (.0257)
Health will likely limit work in the future .4802 .5482 .4387 .5063
(.1774) (.1820) (.1779) (.1817)
Additional controls:
Industry and occupation No No Yes No Yes
Liquid asset levels No No No Yes Yes

 

Note: The dependent variable is the self-assessed chance of working at a particular age,

measured on a scale of 0 to ten. See notes for Tables 2 and 3 for further details.

86

Table 7: Estimated Margin—a] Effects on the Probability of Retirement from a Logit Model

 

 

(1) (2) (3) (4) (5)
Discriminatory effect -.0049 -.0034 -.0038 .0030 .0029
(.0149) (.0144) (.0133) (.0132) (.0120)
Believes that one is paid unfairly -.0354 -.0307 -.0306 -.0260
(.0154) (.0142) (.0141) (.0128)
Workplace environment is friendly -.0014 -.0011 -.0014 -.0040
(.0176) (.0161) (.0161) (.0145)
Experience required for one’s job (years) .0013 .0010 .0011 .0010
(.0012) (.0011) (.0012) (.001)
Schooling required for one’s job (years) .0000 -.0006 -.0010 -.0010
(.0015) (.0015) (.0013) (.0013)
Health will likely limit work in the future .0145 .0172 .0145 .0168
(.0111) (.0103) (.0102) (.0092)
Additionﬂontrols:
Industry and occupation No No Yes No Yes
Liquid asset levels No No No Yes Yes

 

Note: See notes for Tables 2, 3, and 5 for further details.

87

REFERENCES

Burkhauser, Richard and Joseph Quinn. 1983. “Is Mandatory Retirement Overrated?
Evidence from the 19708” Journal of Human Resources 18(3): 337-358.

EEOC Ofﬁce of Research, Information, and Planning. 1999. Age Discrimination in
Employment Act Charges (Washington: EEOC).

Gustman, Alan and Thomas Steinmeier. 1986. “A Structural Retirement Model”
Econometrica 54(3): 555-584.

Hellerstein, Judith, David Neumark, and Kenneth Troske. 1999. "Wages, Productivity,
and Worker Characteristics: Evidence from Plant-Level Production Functions and Wage
Equations," Journal of Labor Economics 17(3): 409-446.

Hutchens, Robert. 1986. “Delayed Payment Contracts and a Firm’s Propensity to Hire
Older Workers” Journal of Labor Economics 4(4): 439-457.

. 1988. “Do Job Opportunities Decline with Age?” Industrial and Labor Relations
Review 42(1): 89-99.

Idson, Todd and Robert Valletta. 1996. “Seniority, Sectoral Decline, and Employee
Retention: An Analysis of Layoff Unemployment Spells.” Journal of Labor Economics
14(4): 654-676.

Johnson, Richard and David Neumark. 1997. “Age Discrimination, Job Separations, and
Employment Status of Older Workers: Evidence form Self-Reports” Journal of Human
Resources 32(4): 779—81 1.

Kotlikoff, Laurence and J agadeesh Gokhale. 1992. “Estimating a Firm’s Age-
Productivity Proﬁle Using the Present Value of Workers’ Earnings” Quarterly Journal
of Economics 108(4): 1215-1242.

Kotlikoff, Laurence and David Wise. 1989. The Wage Carrot and the Pension Stick
(Kalamazoo: Upjohn Institute).

Kuhn, Peter. 1987. “Sex Discrimination in Labor Markets: The Role of Statistical
Evidence” American Economic Review 77(4): 567-583.

. 1990. “Sex Discrimination in Labor Markets: The Role of Statistical Evidence:
Reply” American Economic Review 80(1): 290-297.

Lazear, Edward. 1979. “Why is there Mandatory Retirement?” Journal of Political
Economy 87(6): 1261-1284.

88

. 1981. “Agency, Earnings Proﬁles, Productivity, and Hours Restrictions”
American Economic Review 71(4): 606-620.

Lazear, Edward and Robert Moore. 1984. “Incentives, Productivity, and Labor
Contacts” Quarterly Journal of Economics 99(2): 275-296.

Medoff, James and Katharine Abraham. 1980. “Experience, Performance, and
Earnings” Quarterly Journal of Economics 95(4): 703-736.

Neumark, David and Wendy Stock. 1999. “Age Discrimination Laws and Labor Market
Efﬁciency” Journal of Political Economy 107(5): 1081-1125.

Neumark, David, and Michele McLennan. 1995. "Sex Discrimination and Women's
Labor Market Interruptions." Journal of Human Resources, 30(4): 713-740.

Pergamit, Michael and Jonathan Veum. 1999. “What is a Promotion?” Industrial and
Labor Relations Review 52(4): 581-601.

Polsky, Daniel. 1999. “Changing Consequences of Job Separation in the United States.”
?” Industrial and Labor Relations Review 52(4): 565-580.

Rosen, B. and T. Jerdee. 1977. “Too Old or Not Too Old?” Harvard Business Review
55(6): 97-106.

Topel, Robert. 1991. “Speciﬁc Capital, Mobility, and Wages: Wages Rise with Job
Seniority.” Journal of Political Economy 99(1): 145-176.

89

CHAPTER 3: AGE DISCRIMINATION LEGISLATION, DELAYED PAYMENT
CONTRACTS, AND THE HIRING OF OLDER WORKERS

Two phenomena of the labor market for older workers have been observed for
quite some time. First, ﬁrms are reluctant to hire older workers for many types of jobs.
This is true even if there are a high number of them already employed in these jobs
(Hutchens 1986,1988). This may explain why older workers experience longer spells of
unemployment than younger workers (Hutchens 1986). Second, lifetime earnings tend to
be concentrated toward the end of the work life in the form of higher wages or pensions.
Moreover, evidence shows that age-earnings proﬁles tend to be steeper than age-
productivity proﬁles (e.g., Medoff and Abraham 1980; Lazear and Moore 1984;
Hellerstein, et al. 1999). Prior to the Age Discrimination in Employment Act (ADEA)
amendments in 1978 and 1986, these delayed payment contracts often coincided with
mandatory retirement provisions in what amounted to implicit contracts between ﬁrms
and workers. The reason for this type of delayed payment arrangement has been
addressed often. Lazear (1979, 1981) argued that these implicit contracts were efﬁcient
mechanisms through which ﬁrms combated shirking and malfeasance among workers,
and the passage of the ADEA would serve to prevent ﬁrms from enforcing such contracts
through the use of mandatory retirement.

Hutchens (1986) was the ﬁrst to suggest that these two phenomena are related
directly. He writes that delayed payment contract schemes introduce a ﬁxed cost for the

labor input of ﬁrms. This is because delaying payments raises the expectations on the

90

part of the worker that his ﬁrm will cheat.l That is, a ﬁrm will exploit the services of the
worker at the beginning of his work life and ﬁre him before the higher wages are
realized. In order for workers to still enter into delayed payment contract schemes, ﬁrms
must either increase the payments to workers to compensate for employee fears of
termination or bear some of the cost of shirking. In either case, there is a ﬁxed cost
associated with hiring new workers. From a conceptual standpoint, these ﬁxed costs are
not unlike those that are described by Oi (1962), who showed that employers who incur
training costs when hiring a new employee will hire workers who have the longest
expected tenure. Certainly, younger individuals have longer expected tenures than older
individuals. Thus, delayed payment contract schemes lead to preferences in hiring for
younger workers over older workers. This activity of ﬁrms is thought to resemble age
discrimination. Hutchens offers some empirical evidence to support the theory.

This paper seeks to look at the effect of age discrimination legislation on the
hiring of older workers. At ﬁrst glance, it appears as if the effect should be positive.
Assuming the legislation is enforced, the treatment of older individuals in labor markets
should improve. Thus, hiring of older workers should increase. Moreover, if Lazear was
correct in his prediction that age discrimination legislation should reduce the use of
delayed payment contracts, this would remove one reason that ﬁrms have to prefer hiring
younger workers.

The problem is that the empirical evidence tends not to support Lazear’s

prediction. In fact, most of the evidence suggests that mandatory retirement provisions

 

1 It is assumed that there are information asymmetries present between ﬁrm and worker.
Otherwise, the existence of reputation effects would discourage ﬁrms from cheating, as reneging
would send a signal to workers that they should avoid entering into long-term contracts with the
employer.

91

are not critical to the enforcement of delayed payment contracts and other means exist
whereby such arrangements can be enforced.2 Moreover, the only evidence that exists
that directly tests whether age discrimination legislation reduces the use of delayed
payment contracts ﬁnds evidence consistent with these laws increasing the use of the
long-term contracts (Neumark and Stock 1999). There does appear to be a theoretical
justiﬁcation for this ﬁnding. If mandatory retirement is not essential to the enforcement
of delayed payment contracts, then the abolition of age-based terminations should not be
expected to reduce the use of the contracts. If workers fear that ﬁrms will cheat on
contracts in the pre-legislation environment, age discrimination laws could allay these
fears and reduce the ﬁxed costs associated with compensating workers for this fear. The
reduced cost of delayed payment contracts will allow more ﬁrms to utilize them.

The ultimate effect on ﬁrm hiring practices is therefore ambiguous. First, the
effect of the legislation on the hiring of older workers will be negative if either older
workers are retained longer in ﬁrms that use delayed payment contracts (thus, fewer
openings exist and fewer older workers are looking for work) or the use of the contracts
leads to preferences in hiring for younger workers. Alternatively, the effect of the
legislation may be positive if it succeeds in reducing discrimination in hiring.3
Moreover, the model of Hutchens (1986) presents another case where age discrimination
legislation might increase the hiring of older workers. Because the ﬁxed cost associated
with using delayed payment contracts arises due to worker fears that ﬁrms will renege,

age discrimination legislation can reduce but need not eliminate this ﬁxed cost. The new

 

2 See, for example, Gustman and Steinmeier 1986, Kotlikoff and Wise 1987, Burkhauser and
Quinn 1983, and Hurd 1990.

92

ﬁrms utilizing delayed payment contracts, therefore, will continue to prefer young
workers due to the presence of ﬁxed hiring costs. If age discrimination legislation is
strong enough to eliminate worker fears that their ﬁrm will renege, then the ﬁxed costs
could disappear, leaving ﬁrms indifferent between hiring older and younger workers.

From a public policy perspective, the results of this paper are quite interesting.
Age discrimination legislation, while much maligned by economists from the start for its
potential to disrupt the ability of ﬁrms to engage in efﬁcient long-term contracts, may be
efﬁciency-enhancing. While there may be clear beneﬁts accrued from the legislation,
these may be conring at the expense of certain workers. For example, if the fact that
more ﬁrms are entering into delayed payment contracts means older workers are having
even more trouble getting hired, then the legislation may be beneﬁcial to society but
detrimental to the exact group of workers it was intended to beneﬁt.

To conduct the investigation, I utilize state variation in age discrimination
legislation from 1964 to 1967. Using a control group of workers in states without
legislation and a treatment group of workers in states with legislation, I estimate the
effect of discrimination laws on employment and hiring.

Consistent with Neumark and Stock (1999), the evidence of the paper suggests
that the employment of older workers increases as a result of age discrimination law. The
effect of the legislation on the hiring of older workers tends to be negative (although not
signiﬁcant). The reason for this is not clear, but there is no evidence that suggests an
increased propensity to hire younger workers from any of my estimates. In addition to

the increase in the probability of employment for older workers that I estimate, I ﬁnd a

 

3 Most state age discrimination laws explicitly forbid discrimination in hiring. It is most likely true,
however, that discrimination in hiring is difﬁcult to prove, especially when compared to discrimination in

93

substantial reduction in the probability of retirement for workers in states with age
discrimination laws. This suggests that the lower hiring may be due to ﬁrms retaining
more of their older workers either due to the increased use of delayed payment contracts
or the fear of being sued for wrongful terminations on the basis of age.

Section I of this paper describes the public policy context within which this
research should be evaluated, as well as the related literature that addresses age
discrimination and the legislation aimed at combating it. Section II of this paper outlines
a basic theoretical framework where ﬁrms use delayed payment contracts. It is shown
that these ﬁrms will prefer to hire workers with longer expected tenures. Using the
framework, the expected effects of introducing age discrimination legislation are
discussed. Section III presents some estimates of the effect of age discrimination
legislation on employment. Section IV presents the estimates of legislation on the

probability of being hired.

I. Background on Age Discrimination and Age Discrimination Legislation
Since it was ﬁrst introduced, the ADEA has met with criticism. Lazear (1979,
1981) suggested that mandatory retirement, which was banned partially by an amendment
to the ADEA in 1978 and completely by another amendment in 1986, was a key
mechanism by which ﬁrms enforced delayed payment contracts. He argued that the
abolition of its use would result in a one-time beneﬁt to those who were nearing the
retirement age. It would be detrimental, however, to the long-run interests of the ﬁrms
(who could no longer offer the implicit contacts to reduce shirking) and workers (who

would experience losses in lifetime earnings because they would no longer have the

 

terms of discharge.

94

incentive not to shirk).

This potential disadvantage is enough for skeptics to condemn the legislation
altogether, especially in light of the fact that many in the public do not view age
discrimination as an important problem. In fact, many believe that age may even be a
necessary decision criterion given imperfect information about worker abilities. This was
no more evident than in the recent majority opinion of the Supreme Court in Kimel v.
Florida Board of Regents,4 written by Sandra Day O’Connor. She writes, “a State may
rely on age as a proxy for other qualities, abilities, or characteristics that are relevant to
the State’s legitimate interests. The Constitution does not preclude reliance on such
generalizations. That age proves to be an inaccurate proxy in any individual case is
irrelevant.”

The opinion that age discrimination is not a serious problem is consistent with
some observable facts. For instance, from the August 1999 Current Population Survey
(CPS) it was found that the median hourly wage for men ages 50-65 was $16.15, and the
median hourly wage for men ages 25-49 was $14.42. Moreover, the unemployment rate
for men ages 50-65 was 2.7%, while the unemployment rate for men ages 25-49 was
2.9%. Other groups that claim that they are the victims of discrimination, such as
women and racial minorities, normally can point to such easily observable adverse labor
market conditions as at least a starting point to justify their claims.

Nevertheless, there are some reasons to believe that age discrimination may in
fact be a problem. An example of ill treatment of older workers is the preferences in

promotion given to younger workers in some ﬁrms (Rosen and Jerdee 1977). While

 

’ The ruling prevents workers from suing their state’s government in federal court on the basis of
age discrimination.

95

Rosen and Jerdee do not present evidence of a negative outcome that results from this
practice, Pergamit and Veum (1999) show that there are many positive consequences to
promotion, such as an increase in supervisory responsibilities and the receipt of
additional training. Other work has shown symptoms of age discrimination in labor
markets. Hutchens (1988) shows that job mobility for older workers is comparatively
low. Moreover, as previously noted, Hutchens (1986) also shows that ﬁrms employing
older workers tend not to hire them. Johnson and Neumark (1997) show that individuals
who reported age discrimination in the National Longitudinal Survey of Older Men
(NLSOM) were more likely to separate from their employer and more likely to become
non-employed. The evidence did not show that age discrimination resulted in longer
spells of non-employment or early retirement, but it did indicate that, when rehired, the
workers reporting discrimination experienced lost wage growth.

To combat some of the potential problems associated with age discrimination,
many states passed age discrimination legislation throughout the twentieth century. The
earliest known legislation was a Colorado law in 1903 that prohibited discriminatory
dischargess Table 1 provides details for each state, including the year in which the
legislation went into effect and the speciﬁc age groups that the legislation protects. In
1967, Congress followed the lead of many of the states and passed the ADEA. The
ADEA prohibited discrimination based on age for those ages 40-65. More speciﬁc
concerns were addressed in 1978 when an amendment to the ADEA increased the
minimum allowable mandatory retirement age to 70 and, a year later, when enforcement
of the original ADEA was switched from the Department of Labor to the EEOC. In

1986, mandatory retirement was banned altogether. Recently, the ADEA has been used

96

as the basis for many claims of age discrimination. Between ﬁscal year 1992 and ﬁscal
year 1998, the EEOC reported that there were a total of 123,111 charges ﬁled under the
ADEA (1999). This should be compared to the 207,191 charges ﬁled alleging race
discrimination and the 170,751 charges ﬁled alleging sex discrimination.”
An important question that has not been explored to a great extent in the literature
is the effect of this legislation. Neumark and Stock (1999), using variation in state age
discrimination legislation, show employment for older workers increasing after the
passage of the laws, an expected result of age discrimination legislation. They also show
that earnings proﬁles become more steep following the passage of ADEA-type
provisions, thus indicating that the ADEA might actually serve to increase labor market
efﬁciency by promoting the formation of delayed payment contracts. This may be the
case because workers are less fearful that ﬁrms will renege on these contracts. That is,
the specter of potential age discrimination suits ensures that ﬁrms will make good on the

implicit promise to workers that their good work will result in wages above their

marginal product in the latter stages of their work life.

11. Conceptual Framework
To provide a theoretical context for the empirical results of the paper, this section
undertakes three tasks. First, a simple framework is proposed that borrows heavily from

Lazear (1979, 1981) and Hutchens (1986) and maintains all of the basic features of a

 

5 The Colorado law, however, had no enforcement provisions (Northrup 1980).

6 Chapter 2 evaluated the effect of age discrimination in the post-ADEA enforcement period. It
showed that older workers who report that their employer promotes on the basis of age
experience comparatively low wage growth in the latter stages of their work life and rate the
chances of retiring early as greater than the ratings given by those not experiencing
discrimination.

97

model where ﬁrms use delayed payment contracts. It explains why age discrimination
legislation may increase the use of delayed payment contracts. Second, it is shown how
the use of delayed payment contracts may compel ﬁrms to hire younger workers. Third, I
explain how age discrimination legislation may change hiring practices under certain

conditions.

A. Delayed Payment Contracts and Age Discrimination Legislation

Many ﬁrms are endowed with a technology that prevents them from costlessly
observing worker output, and workers gain some beneﬁt from shirking. These two
assumptions, which certainly have some basis in reality, serve as the basic assumptions
for a model of delayed payment contracts.

Let Hbe the gain to a worker from shirking. Let w, be a wage that is paid to a
worker in period t by a ﬁrm that cannot costlessly observe worker output and faces a

downward sloping demand curve. The utility function for a worker in each period is
(1) u, = w, + 6.
Assume that discount rates between ﬁrms and workers are the same (zero, for simplicity).

Also assume that workers are identical in all respects except age. Each worker works

from period 0 to period T.7 Let V be a constant marginal product for all workers. There

 

7 This is clearly not a formal presentation of Lazear’s model, as he assumed that ﬁrms choose T
(the age at which ﬁrms retire) to enforce the end of delayed payment contracts. Empirical
evidence suggests that retirement ages are determined outside of ﬁrms imposing mandatory
retirement on its workers, however (e.g., Burkhauser and Quinn (1983) and Hurd (1990)). Thus,
T is assumed to be exogenous in this paper for simplicity. Hutchens (1986) makes a similar
assumption.

98

exists a competitive sector in which all workers could seek alternative employment, but
shirking is not a possibility there because ﬁrms can costlessly detect it. In the
competitive sector, workers are paid V throughout their worklife. Finally, let q be the
rate at which ﬁrms catch workers shirking and ﬁre them, which is between 0 and l in the
sector where output is not costlessly observed. In the competitive sector, q = I. All
workers are risk-neutral.

In this framework, ﬁrms cannot simply pay workers their marginal product, as the
incentive for workers to Shirk always will exist. Firms could invest in technology to
increase q, but this is costly and more efﬁcient means to reduce worker shirking exist.
Speciﬁcally, ﬁrms can offer one of the wage proﬁles illustrated in Figure 1. The ﬁrst
choice, which is indicated by the darkened line, starts with workers posting a bond equal
to 11% at the time they are hired. They then are paid a wage equal to their marginal
product throughout their worklife. At the time they retire, they receive an end payment

equal to BVT. In the second choice, workers are paid the wage path wowr. If

(2) AV0 -Bv, a 211m}.— w,.) a 0,

i=0

then the worker is indifferent between either of the payment schemes in Figure 1 or one
that pays a constant V throughout his worklife.
In addition to worker indifference between alternative wage paths, the payment

schemes must be set to deter shirking. Assume that the ﬁrm has decided to choose

99

AVoVTB as its wage path.8 Thinking ﬁrst about period T, a ﬁrm sets VTB just high enough
to deter shirking (i.e., the return to a worker from shirking equals the return from not

shirking). Speciﬁcally,

(3) 6+(1-quTB) = (VTB)

is satisﬁed. This simpliﬁes to VTB = 0/ q. Provided that the worker will not shirk in
period T, he will not shirk in period T-I, T-2, etc.

Lazear’s model originally assumed that ﬁrms were bound to the wage path that it
offered by reputation effects. In practice, however, information asymmetries exist.
These may arise for a variety of reasons. Just as ﬁrms cannot costlessly observe worker
output, neither can fellow workers. Thus, when a worker is ﬁred, it is not always the case
that workers can tell if the ﬁring was justiﬁed. Moreover, ﬁrms may have to lay off
workers in response to shifts in the demand curve for their product. Thus, there is a
possibility that ﬁrms can cheat on the contracts that they form and get away with it.
Figure I certainly indicates that ﬁrms have the incentive to renege on contracts. In wage
path wowT, ﬁrms will wish to renege at t’. In wage path AVOVTB, ﬁrms will renege at T.9
In such a case, equation (2) is no longer sufﬁcient for workers to agree to delayed
payment contracts if workers fear that ﬁrms will renege. Therefore, ﬁrms must
compensate workers for the expected loss in lifetime earnings from potential ﬁrm

cheating. While the exact structure of this compensation in regard to the above

 

8 The following logic works for the alternative wage path wowT if BVTequals the area of triangle
EV-er. For simplicity, the rest of the section will focus on a wage path where a worker posts a
bond and the ﬁrm pays back the bond at the end of the ﬁnal period assuming the worker does not
shirk in period T.

100

framework will be discussed in the next subsection, sufﬁcient information exists to state
that bonding is no longer costless. The introduction of the possibility of ﬁrm cheating in
this framework decreases efﬁciency.

Bringing age discrimination legislation into the picture, however, changes the
situation drastically. Because these laws prohibit age-based employment terminations,
worker fears that their ﬁrm will renege on delayed payment contracts could fall. This
will lower the costs of bonding. The costs may fall enough for some ﬁrms to ﬁnd it in
their interest to offer delayed payment contracts. Neumark and Stock (1999) ﬁnd
evidence consistent with this. First, employment rates for older workers increased with
legislation, thus suggesting that workers are retained for a longer term. Second, slopes of
age-earnings proﬁles steepened, thus suggesting that more ﬁrms offered wage paths like

those in Figure 1 after the passage of the legislation.

B. Delayed Payment Contracts and Firm Hiring Practices

In this subsection, 1 summarize the basic model presented by Hutchens (1986)
that explains why ﬁrms using delayed payment contracts tend to prefer hiring younger
workers. His whole argument hinges on the fact that the costs mentioned above that stem
from ﬁrms compensating workers for the probability of ﬁrm cheating contain a
component that is ﬁxed with respect to contract length. That means that every time a
ﬁrm hires a worker, whether it is for four or forty years, they incur a cost. While part of
the cost is related to the length of the contract, there will be a component that is time-
invariant. Given this, ﬁrms will seek to minimize the amount of times it hires a worker

and, therefore, employ workers for longer periods of time.

 

9 The fact that ﬁrms behave in this fashion has been veriﬁed empirically (Orr 1998).

101

To see under what conditions this would be true, let the following be the notation:

W = 2w, (i.e., sum of the wages paid to a worker over his lifetime)

q = probability that a firm catches a worker shirking and ﬁres him

6. = gain to worker i from shirking (ﬁrms only know the distribution
ﬂ 6) across all workers)

b = bond paid by a worker at t = 0 (e.g., AVo in Figure l)

p = end payment or pension paid to a worker the instant after period T

(e.g., BVT in Figure 1)

S = cost to ﬁrm from a worker shirking (e. g., lost credibility with
customers for delaying a shipment; S is always greater than 6)

y = cost to ﬁrm for reneging on a contract (e. g., increased costs
of attracting workers in future—in the next subsection, this will
include the penalties imposed by age discrimination laws; workers
only know the distribution gm across ﬁrms).

To begin, ﬁrst consider the costs to a ﬁrm that arise from employing each

identical worker, which are
(4) C a (l- F(qp))(w + p —b) + F(qpxs +W + (1- q>p —b>.

This is the sum of the expected cost of a worker not shirking and the expected cost of a

worker shirking. F ( qp) is the probability of worker shirking and must be equal

to I f (t9)d6 to deter shirking. Equation (4) simpliﬁes to
W

102

(5) CEW+p+F(qp)(S-qp)-b.

T 0 further simplify matters, I still assume that ﬁrms offer a wage path similar to AVoVrB
in Figure 1.10 Thus, in addition to choosing the number of workers to employ, ﬁrms need
only choose a b and a p to offer each worker. The proﬁt-maximizing b and p, which I
will label b* and p*, must be such that the expected lifetime earnings the worker receives
is equal to his alternative lifetime earnings in the competitive sector (TV). This condition

is
(6) W+(1—G(p»p-b 21v,

where G( p) is the probability that a ﬁrm will cheat on a contract. Speciﬁcally, this must
P

be equal to I g(y)dy.
0

To determine what the optimal cost per worker would be, ﬁrst rearrange equation
(6) in terms of b and substitute for b in equation (5). Then, just solve for the optimal p*
that minimizes the cost of the marginal worker in equation (5).11 The optimal costs per

worker will be as follows after simplifying

(7) C* 5 TV - qp * F (6112*) + F(qp*)S + p * G(p*).

 

‘0 Actually, Lazear shows that this path dominates all others in equilibrium.

“ The optimal p* solves the ﬁrst-order condition qp *ﬂqp *)q + F(qp *)q - p*g(p *) - G(p *) -

ﬂqp *)qS = 0. Note that p* could be completely written in terms of q, g( 0), ﬂ 0), and S, which are
exogenous and time-invariant.

103

Thus far, it has been established that a ﬁrm endowed with a technology that
prohibits costless monitoring of worker output incurs the costs in equation (7) for each
worker it offers a delayed payment contract. TV clearly increases with the length of the
contract. Due to the fear of ﬁrm reneging expressed in G( p *), however, and the positive
probability of worker shirking F(qp *), ﬁrms face a time-invariant cost that is equal to the
sum of the second, third, and fourth terms in equation (7).12 These three terms together
(hereafter HC for hiring costs) represents a ﬁxed cost to the ﬁrm. This cost consists of
the exogenous time-invariant factors p* (which is completely determined by time-
invariant factors), the worker’s perceived distribution of ﬁrm reneging (g(y)), S, and the
distribution of f( 19).13

It is now possible to see why ﬁrm hiring practices are affected by these contracts.
In the absence of a positive HC, a ﬁrm would be indifferent between hiring an older
worker and a younger worker. Because HC is incurred each time a worker is hired,
however, it means that it is more proﬁtable for ﬁrms to minimize the number of times
that it hires workers. If it hires an older worker, it not only pays HC now but will pay it

again shortly because that worker will retire at T. In the above model, therefore, ﬁrms

 

'2 If f(qp*) > 0, then S > qp since S is assumed to be greater than all Band F(qp*) is the

probability of worker shirking, which is deﬁned as F(qp*) = I f (9)d 0.

W.
13 The assumptions stated in the model appear on face to be restrictive. Hutchens (1986),
however, shows that more general formulations still result in positive ﬁxed hiring costs. For
example, the implicit assumption that S, q, ﬂ 0), and g( 0) do not depend on Tcan be relaxed, as
long as the loss to ﬁrms from worker shirking(S) is greater than Has T approaches 0. Also, in this
case, S must be greater than a ﬁrm’s penalty from reneging (y) as T approaches 0. The zero
discount rate assumption can also be relaxed, as can the assumption of a constant V.

104

will always choose workers with the longest expected tenures. Because the workers are
identical in all respects except age, they will hire young workers before older workers.”
As an empirical test of the above model, Hutchens constructs an index of the
propensity of ﬁrms to hire older workers. This is a ratio of the proportion of recently
hired workers that are over the age of 55 to the proportion of total employees over the age
of 55 in industry-occupation cells of the 1970 census. Then, using the NLSOM, he looks
at whether workers in jobs with a higher index are less likely to report that their ﬁrms
engage in practices consistent with the use of delayed payment contracts (e.g., pension
offerings, mandatory retirement). He ﬁnds evidence consistent with the predictions of his

model.15

C. Age Discrimination Laws and Firm Hiring Practices

It was noted in the introduction that age discrimination legislation could affect
hiring outcomes for a variety of reasons. First, it bans discrinrination in hiring, so this
could improve hiring outcomes for older workers. Second, it could increase the use of
delayed payment contracts, which according to the theoretical result in subsection B,
would lower the hiring for older workers. Moreover, one would expect ﬁrms using
delayed payment contracts to retain workers for longer tenures. Thus, fewer openings

exist and less hiring of older workers would occur.

 

'4 The assumptions that are needed to generate a positive ﬁxed cost when model parameters
depend on T (Footnote 12) are sufﬁcient to generate a component of HC that does not depend on
T. Thus, even if certain parameters, such as ﬁrm losses from reneging or worker gains from
shirking, are allowed to vary by the length of the contract, ﬁrms will still prefer to hire younger
workers.

‘5 Although they do not test Hutchens’ theory explicitly, recent ﬁndings by Hirsch, et al. (2000)
are certainly consistent with the theory of delayed payment contracts inﬂuencing ﬁrm hiring
decisions.

105

While it is true that the model of Hutchens suggests that delayed payment
contracts might increase the propensity of ﬁrms to hire younger workers, it allows for one
case in which the propensity to hire younger workers will fall with age discrimination
legislation, even if the use of delayed payment contracts increases. This occurs when the
penalties associated with breaking an age discrimination law become so great that
workers perceive the probability of a ﬁrm reneging to be zero. In such a case, G( p) = 0
for all p.

The ﬁrm problem is still to minimize the cost of employing the marginal worker
given that his expected lifetime earnings must equal his earnings in alternative
employment. The no-shirking condition, however, now becomes determinant of p*
because the worker need not take into account the ﬁrm probability of reneging when
deciding whether or not to shirk. Thus, a ﬁrm need only set p* = ﬂ6)/q, as this is
essentially the same situation outlined in subsection A of this section. In that case,
equation (2) and equation (3) were sufﬁcient conditions for ﬁrms engaging in the types of
contracts in Figure 1. Because there is no ﬁxed cost to hiring workers, ﬁrms can redesign
any of the Figure 1 contracts to suit a young or old worker. Thus, under the conditions
that workers perceive the penalties of breaking age discrimination laws to be severe
enough to prevent ﬁrms from reneging, age discrimination legislation could increase the
use of delayed payment contracts but move ﬁrms from preferring to hire younger workers
to leaving them indifferent in their hiring preferences.16 The direction of the effect of age
discrimination laws on ﬁrm hiring practices is therefore ambiguous even using the model

proposed by Hutchens. This renders the effect of age discrimination laws on hiring an

 

'6 This is not to say that ﬁrms will not still prefer to hire younger workers for other reasons, such
as those that arise from ﬁxed costs associated with training workers.

106

empirical question. The rest of the paper uses state variation in these laws to attempt to

answer the question.

III. The Effect of Age Discrimination Legislation on the Employment of Older
Individuals

Before launching into tests of the effects of age discrimination laws on hiring
practices, I ﬁrst look at the effect of age discrimination legislation on employment for
several reasons. First, it is an interesting question in its own right. Age discrimination
legislation is expected to improve labor market conditions for older workers, and, if
effective, should certainly result in higher employment rates for the aged. Second, the
effectiveness of the legislation is a necessary condition for changes in the use of delayed
payment contracts. Thus, the estimates in this section serve as a starting point for looking
at the effect of age discrimination legislation on outcomes of interest. Finally, Neumark
and Stock (1999) ﬁnd employment effects of age discrimination legislation using census
data before they look at the effects of the legislation on age-earnings proﬁles. In this
sense, the estimates of this section can be compared to their evidence to assess the
validity of my data.

The data that I use for this and the following sections come from the annual
demographic ﬁles of the Current Population Survey (CPS) for the years 1964-1967. This
is a period in which there was much variation in age discrimination laws across states.
As shown in Table 1, some states ﬁrst adopted laws during this time. Moreover, during

this four-year period, all states can be identiﬁed in the CPS, something that is not true of

107

the data from 1968-1977.” Finally, federal legislation took effect in 1967, but
enforcement of the legislation was limited at best (Rosen and Jerdee 1977). Thus, for this
analysis, I look only at those states with legislation that was enforced compared to states
with no laws at all.

I apply some simple restrictions to the data in order to cut down on the
confounding effects that may bias the estimates of the effect of age discrimination
legislation. Speciﬁcally, I exclude women and non-whites to avoid unduly attributing the
effect of laws protecting these groups to laws protecting the aged. This is very
reasonable given the fact that the existence of race and gender discrimination laws are
likely to be correlated with the existence of age discrimination laws. I also limit attention
to private sector workers, as the underlying theory that is relevant to my estimates most
applies to these individuals. Finally, only those between the ages of 18 and 70 make my
sample. I merge to this data the information on state age discrimination laws in Table
1.13

To facilitate comparison of the estimated employment effects in this paper with
those obtained by Neumark and Stock (1999), I use their basic approach. Speciﬁcally,

the following equation is estimated for data pooled from 1964 to 1967:

 

'7 Although data exists for the years 1962 and 1963, these years are not used in the sample
because important information related to the analyses in each year is missing. In 1962, I cannot
identify all states. In 1963, some demographic information is missing.

‘8 Colorado and North Dakota residents are removed from the sample, as their inclusion may
contaminate the control group due the existence of non-enforced age discrimination laws. No
such action is taken for Texas or Illinois residents, even though laws there are not enforced as
well. The law in Texas only applies to public sector employees, who are not included in my
sample in the ﬁrst place. Illinois residents are not excluded as their legislation took effect after
the end of the sample period (March 1967). Residents of Rhode Island and Indiana are excluded
as well. Although enforcement bodies were entrusted with executing the laws in these states, no
penalties for breaking the law are speciﬁed.

108

(8) E = X01 + [32COV*AD + B;UNCOV*AD + 8.

Individual subscripts are omitted. E is a dummy variable indicating whether a worker is
employed. X is a vector of individual characteristics.” COV is a dummy variable set
equal to one if the respondent is in the age group that is covered by the legislation in his
state (See Table 1.). UNCOV is a dummy variable set equal to one if the individual is not
in the covered group.20 The estimate of the coefﬁcient B; is the effect of age
discrimination legislation on those in the covered age ranges and the estimate of Ba is the
effect on those in age ranges that are not covered. [33 is a parameter of interest as well
because the effect of legislation on those that fall outside of protected age ranges is
uncertain.21

Table 2 presents estimates from linear probability models of the parameters in
equation (8).22 In column (1), the numbers in the ﬁrst row are the point estimate and
standard error of [52. In the sixth row, estimates for [33 are presented. For the
speciﬁcation used to obtain the estimates in column (1), dummy variables for each year

and each state are included in the X vector. The table shows that the effect of age

 

'9 These include SMSA status, marital status (dummies for separated or divorced, never married,
and widowed), education level (dummies for high school graduate, some college, and college
graduate), and year dummy variables.

Included in X are a series of dummy variables that denote that an individual is in a particular
age group (speciﬁcally, 18-24, 25-29, 30-34, 35-39, 40-44, 45-49, 50-54, 55-59, 60-64, and older
than 65).

2' While independence across states and years in my sample can be assumed, observations within
state-year cells are most likely not independent. Thus, estimating equation (8) parameters by a
linear probability model may result in standard errors that are incorrect. For this reason, I
estimate standard errors that relax the assumption of independence within state-year cells and
report these in the tables.

22 A probit model was also used to estimate equation (8), as well as other speciﬁcations in the
paper where a dummy variable was the dependent variable. Because the results are not notably

109

discrimination legislation on the rate of employment is positive for those that are covered
by the legislation. The estimate suggests that the probability of employment is increased
by about 1.45% for those workers in states with age discrimination legislation, with the
comparison group being workers of similar age and demographic characteristics in states
without legislation. The effect on those that are not in the covered age ranges is not
signiﬁcantly different from zero. The estimates are very close to those presented by
Neumark and Stock (1999) for a similar speciﬁcation using census data. They ﬁnd that
legislation boosts employment by about 1.64% for individuals in covered age ranges.

In column (2), a series of age-year interactions is added to the speciﬁcation. The
aim of this is to account for changes across time in employment rates of workers of
different age groups that are common across states and not captured by the other
explanatory variables. The changes in the estimates that result from this alteration to the
speciﬁcation are very small. Given that I am using a sample of just four years, this is not
surprising.

Frequently, legislation offers protection to those that can hardly be considered
old. As shown in Table 1, the legislation of several states covers workers that are quite
young (e.g., Idaho and Maine). Every state covers workers in their forties and early
ﬁfties. Given that individuals in the highest end of the age distribution of covered
workers are the most likely to beneﬁt from legislation, estimating employment effects for
these workers may be more appropriate than estimating effects for all covered workers.

Thus, keeping with the methodology of Neumark and Stock (1999), I break up the

effect of the legislation into an effect on older covered workers and younger covered

 

different than those reported for the linear probability model, I only report the estimates obtained
from linear probability models.

110

workers. Older covered workers are ﬁrst deﬁned as any worker aged 60 years or more.
Also, I reestimate with workers aged 50 years or more considered old. These results are
reported in Table 2 as well. In column (1’), the effect on older covered workers is
positive and signiﬁcant. In fact, the effect is quite large, suggesting that these older
workers experience about a 5.6% boost in the probability of employment due to age
discrimination legislation. The effect again is very close to the effect estimated by
Neumark and Stock (1999). The effect on younger covered workers is positive as well
but not signiﬁcant. In column (1"), when old is deﬁned as ﬁfty, the effect is smaller on
older workers but still positive.

The estimates of this section reveal that age discrimination laws are having the
expected effect on employment under the condition that the legislation is effective. As
with the sample of censuses that was used by Neumark and Stock (1999), the sample
constructed in this paper reveals that older workers in states with age discrimination
legislation are experiencing employment increases. It is important to note, however, that
the increased employment of older workers may be due to increased hiring of older

workers or an increase in the probability that older workers are retained by employers.

IV. The Effect of Age Discrimination Legislation on Hiring and Worker
Retention

Although the results of the prior section are reassuring in that they are similar to
those reported by Neumark and Stock (1999), my paper is more concerned with the
previously unexplored question of whether these employment increases are due changes
in hiring practices. In this section, I begin the examination of this question by looking at

the effects of age discrimination legislation on the probability of being a new hire using a

111

sample of individuals. I also look at whether the hiring effects estimated are just a

reﬂection of more older workers being retained at their job.

A. Identifying Newly Hired Workers in the CPS

Although the underlying question of this paper is concerned with the behavior of
employers, the data that is available to answer the question is at the individual level.
Thus, instead of observing ﬁrms, I view individual outcomes. The effects of age
discrimination legislation on these outcomes must be interpreted with their implications
regarding ﬁrm behavior in mind. Primarily, in this section, I try to determine whether
age discrimination legislation has resulted in a change in the probability of being newly
hired among older individuals in states where age discrimination legislation has been
enacted.

The CPS does present an additional problem for this analysis, however. There is
no way to determine explicitly who is a newly hired worker. Moreover, tenure
information is not available on an annual basis for each year from 1964 to 1967. Thus, a
proxy variable for newly hired workers must be found.

A measure that is possible with the CPS and also is quite close to that used by
Hutchens (1986) identiﬁes as newly hired those workers that switched industries from the
prior year. Speciﬁcally, in the CPS, information on the industry of individuals in the
prior calendar year is given. The same information is provided about their current job.
Although in accord with the literature, this measure is inappropriate for my purposes
because there are many ways that an individual may be misclassiﬁed as a new hire by this

method. First, he may have switched companies within an industry. If he reports the

112

same industry for both years, however, he will not be classiﬁed as a new hire. Second, he
may have actually been with the company for both periods, but the job listed as his main
job currently may not have been his primary job from the prior year. This will result in
individuals being classiﬁed as a new hire that may not be. Third, an individual may have
left his employer for a period and then returned. These workers will be considered new,
even though it is unclear had their employment history been known they would have been
classiﬁed as new. Finally, the CPS did not begin using dependent coding until 1976.
When dependent coding is used, if an individual did not switch jobs, the same industry is
listed for the job last year that is listed for the current job. Prior to 1976, the variables
were independently coded, with no such check on consistency of the measures (Stewart
1999).

Given these problems, the paper uses two other measures of new hires. First,
individuals are considered a new hire if they were not employed at some point during the
prior year but are employed at the time of the interview. This measure deals with some
of the problems listed above, but it certainly does not represent a complete solution. For
example, individuals may have taken a leave of absence and returned to the same job.
For this reason, I use a ﬁnal deﬁnition of a newly hired worker that requires the person to
have actively searched for work while not employed in the prior year. While this ﬁnal
measure probably is better, it still may fail to classify some new hires. Speciﬁcally,
individuals that moved from one job to the next without a spell of non-employment in
between may be misclassiﬁed. With the above problems duly noted, I still proceed with

conﬁdence that the measures I use identify the bulk of newly hired workers.

113

B. Empirical Methods and Results
I ﬁrst employ a similar method to that used to look at the employment effects of

age discrimination legislation. Speciﬁcally, the following equation is estimated:

(9) N -.= xii1 + B;COV*AD + B3UNCOV*AD + e.

N is a dummy variable indicating the worker is a new hire. The remaining variables are
the same as those that were used in the prior analysis in the paper.

The estimated effects of age discrimination legislation on the probability of being
a new hire, when a new hire is measured as anyone moving from non-employment to
employment, are reported in Table 3. In column (1), a small decrease in the probability
of being a new hire is estimated for workers in the covered age ranges. A smaller
decrease in the probability of being a new hire is estimated for workers outside of the age
range. Neither effect is close to being signiﬁcantly different than zero, however. Similar
effects are estimated when individuals are required to have searched for work in the prior
year to be considered newly hired.

I next break up the estimated effects on covered workers into an effect on older
covered workers and younger covered workers. As in the prior section, effects are ﬁrst
estimated when workers are considered old if they are aged 60 years or more, and then
effects are reestimated when individuals aged 50 years or more are considered old. The
results in columns (1”) and (2”) suggest that the probability of older workers in the
covered age range being newly hired is negatively affected by age discrimination laws,

but the effects are not signiﬁcant.

114

A second and similar approach can be employed to look at the same question that
perhaps will provide a bit more information about the effects of legislation on ﬁrm hiring
practices. Speciﬁcally, effects are estimated throughout the age distribution. This may
be the more relevant question for the purposes of this paper, as the ultimate aim of the
paper is to assess the effect that age discrimination legislation has on aged-based
preferences in ﬁrm hiring, which may stem either directly from the legislation or
indirectly through the increased or decreased use of delayed payment contracts. Thus, I
look at the effect of age discrimination legislation on the hiring of older workers of
different age groups, using younger workers as a comparison group.

Given that the hiring of younger workers may be affected by age discrimination
legislation as well, this group is not meant to serve as a control group for the purposes of
drawing inferences regarding the causal effect of age discrimination laws. It merely
serves to highlight where in the age distribution the hiring effects of the legislation are
likely to bite. This will present an accurate picture of whether ﬁrm hiring practices are

changing. Speciﬁcally, the following equation is estimated:

(10) N = X61 + 820AG + B3AD + BaAD*OAG + 8.

Individual subscripts are again omitted. OAG is whether an individual falls in the
particular older age group in question. The older age groups are those aged 40-44, 45-59,
50-54, 55-59, 60-64, and 65 or more years.23 AD is whether an individual lives in a state

with an age discrimination law. Separate regressions are run for each older age group.

 

23 These groups were chosen because when a lower limit on the age of protected workers is set,
the limit is most often the age of 40.

115

For example, to estimate the effects on those aged 50-54 years, the sample consists only
of individuals either below the age of 40 or between the ages of 50 and 54. State dummy
variables are always included in the X vector in addition to the standard demographics
characteristics and year dummy variables.

The results are reported in Table 4, which is broken down into panels using the
two measures of new hires. In the ﬁrst column of each panel (columns (1) and (2)), the
differential hiring effects on older workers are reported. Remember that the reference
group consists of those below the age of 40 and each row presents results from the
estimation of separate regressions. The estimates in both panels reveal that there tends to
be a negative differential effect on the hiring of older workers due to age discrimination
legislation. The effects are not statistically signiﬁcant for the most part, however.

In the second column of each panel (columns (1’) and (2’)), the total effect of the
legislation on workers in each age group is presented. This simply is the sum of the
estimates of 133 and B4 in equation (10). The estimated effects are consistent with age
discrimination legislation reducing the probability of being a newly hired older worker.
In fact, when new hires are required to have looked for work in the prior year, there are
substantial reductions in the probability of being newly hired for the 45-49, 55-59, 60-64,
and 65-70 age groups.24

It is important to note, however, that the differential hiring effects on older
workers are not signiﬁcant for any age group other than the 65-70 year olds. The
differential effects are not even close to being signiﬁcant for any other age group. Given
that age discrimination legislation covered those over 65 in just four states during this

time period, there is little evidence of an increase in the propensity to hire younger

116

workers in states enacting age discrimination legislation. For the most part, there seems
to be a reduction in the probability of being newly hired for workers of all ages.
Moreover, these results do not tell the complete story given the fact that the probability of
employment is higher in states with age discrimination laws. It is quite possible that the
hiring results are not reﬂecting the fact that more ﬁrms are using delayed payment
contracts and preferring to hire younger workers. It may simply be the case that older

workers are remaining with their ﬁrms for longer tenures.

C. Age Discrimination Laws and Worker Retention

In an effort to explain why the probability of being newly hired is negatively
affected by age discrimination legislation for some groups of older workers, I test for the
existence of retention effects. Speciﬁcally, I attempt to determine whether older workers
are more or less likely to move from the status of employed to the status of non-employed
(or unemployed) where age discrimination legislation has been enacted.

To conduct the test, I ﬁrst replace the dependent variable in equation (9) with a
dummy variable indicating the worker has moved from being employed in the prior year
to being non-employed in the crurent year.25 The estimated effect will essentially be the
opposite of a retention effect, but it should provide greater insight into whether age
discrimination legislation causes more older workers to be retained at their job. The
results are reported in the left panel of Table 5 and suggest that workers in covered age
ranges are more likely to move from employment to non-employment. In fact, when the

effects on the covered age range are broken down into an effect on older workers and an

 

2’ These are signiﬁcant at the .10 level, at the very least.

117

effect on younger workers, the increased probability of job loss seems to fall largely on
older workers. I also arrive at similar estimates when the dependent variable is measured
as movements from employment in the prior year to current unemployment. The results
are reported in the right panel. Note that the increased probability of job loss is
signiﬁcant for older workers.

These results are the opposite of what was expected. Older workers do not appear
to be retained at their jobs with greater probability in states enacting age discrimination
legislation. If anything, it appears as if there are more older workers searching for work.
There are several possible explanations for the results. First, it may the case that age
discrimination legislation is not improving conditions at all for older workers, and the
increased use of delayed payment contracts may be making it even harder for them to
ﬁnd employment. Second, there may be problems with my data. I only know whether a
worker received a wage in the prior period. From this, I determine his employment
status.

To investigate further what is happening, I return to the positive employment
effects experienced by older workers that are estimated in Table 2. If it is true that the
hiring of older workers falls (or remains the same) with legislation and the probability of
becoming unemployed increases, why are there are positive employment effects? Instead
of looking at transitions into and out of employment, here I estimate the effects of age
discrimination legislation on the probability of being unemployed and the probability of
being retired. The results are reported in Table 6. As shown, there is an increase in the

probability of unemployment for all workers in states with enacted legislation. At the

 

25 Unfortunately, the best measure of employment status in the prior year is whether an individual
earned a wage at any point in the prior year.

118

same time, however, there are huge reductions in the probability of retirement among

older workers in states where age discrimination laws have been adopted.

D. Discussion of Results

What do the estimates of this section mean in terms of labor market outcomes for
older individuals? While the results do suggest that there is a reduction in the retirement
of older workers, it also appears as if more older individuals are becoming unemployed
and fewer are ﬁnding work. Thus, to illustrate the magnitude of these alternative effects,
I calculate what the implications of the estimates are for a typical state that adopts age
discrimination legislation.

1 present the results for individuals aged 50 years or more in Table 7. I ﬁrst look
at the situation prior to the adoption of age discrimination legislation. In the ﬁrst column,
the number of people in the sample that ﬁt each category is listed. For example, 564
individuals aged 50 years or more looked for work for at least one week in the prior year
and found work by the time of the March CPS sample.26 In the adjacent column, these
numbers are magniﬁed to reﬂect what would be expected in a state with a population of
one million white male individuals aged 18 to 70 years. Thus, 11,517 individuals aged
50 years or more in the state normally move from looking for work to becoming non-
employed. There are 15,499 workers that normally are employed in the prior year but are

not employed in the current year (not listed). A total of 4,431 of these people are looking

 

26 Since the aim is to determine the effect of legislation on the affected age groups, I exclude from
the calaculations in Table 7 those individuals aged 65 years or more because legislation
frequently does not cover these individuals. For example, the original federal legislation only
covered those aged 40-65 years.

119

for work in the current year and 11,068 are not.27 Also listed are the number of retired
workers overall in the state (26,097) and the number of unemployed (5,534).

With the advent of age discrimination legislation, the results of this paper suggest
some changes in these numbers. In the third column, the relevant estimates obtained
earlier in the paper from linear probability models are listed. Robust standard errors
appear below each estimate. The estimates are used to contruct conﬁdence intervals of
the changes in the numbers of individuals in the state that fall in the various categories in
the left-hand column. Thus, in the hypothetical state, between 40 and 231 fewer
individuals over the age of 50 will move from being unemployed to employed following
the adoption of the legislation. Also signiﬁcant are the increases in the number of
individuals moving from employed to not employed but looking for work and the number
of total unemployed. These changes pale in comparison, however, to the huge reductions
in the number of retired workers (between 389 and 749 fewer retirees among affected
individuals aged 50 years or more).28

The Table 7 exercise is highly suggestive of the dominant effect of age

discrimination legislation being the reduction of retirements. Firms appear to respond to

 

27 A sizable number of those in the latter group are retirees, but some are individuals that are
unable to work or on unpaid absenses from their job. This variable was not used in prior analysis,
but it is important to the overall story of this section so it is now included.

28 A rather peculiar result is the rather small reduction in the number of workers moving from
employed to not working and not looking for work (in fact, a positive change cannot be ruled
out). If the total number of retired workers has dropped signiﬁcantly, it would seem as if this
reduction should be greater. A possible explanation is rnismeasurement that stems from the fact
that employment status in the prior year is solely determined by whether one received positive
income from wages and salary. It is possible that employment status in the prior year is also
positively and signiﬁcantly affected by age discrimination legislation, as firms react to impending
legislation by increasing the employment of older workers. I tested for this possibility by
estimating the effect of legislation on employment status in the prior year and found a signiﬁcant
positive effect on the employment of older workers that was similar in magnitude to the Table 2
estimates. Thus, the transition into retirement will be partially masked by the increase in the

120

legislation by retaining more of their older workers. There are modest declines in the
ability of older workers to ﬁnd jobs and increases in unemployment rates for older
workers, but these effects may largely be due to the fewer job openings that the lower

rates of retirement bring.

V. Conclusion

Lazear (1979) suggested that age discrimination legislation might beneﬁt current
older workers at the expense of future labor market efﬁciency. Later empirical work has
shown not only that the importance of ﬁrm mandatory retirement provisions in enforcing
these implicit contracts is not as great as initially thought, but the predicted losses of
efﬁciency have not occurred. In fact, the only evidence that tests the effect of age
discrimination legislation on the existence of delayed payment contracts ﬁnds evidence
consistent with an increase in these ﬁrm-employee relationships (Neumark and Stock
1999).

The fact that delayed payment contracts may increase with age discrimination
legislation raises an interesting empirical question that my paper addresses. Delaying
payments to the end of an individual’s worklife introduces a ﬁxed hiring cost according to
Hutchens (1986). Thus, ﬁrms will tend to hire younger workers to maximize expected
tenure. This propensity toward hiring younger individuals presents a major problem for
older workers in labor markets in terms of job mobility. While age discrimination
legislation aims to improve labor market conditions for older individuals, the increased

use of delayed payment contracts stemming from the laws may exacerbate age-based

_‘

measure of prior-year employment. This also may mean that the employment effects reported in
Table 2 and the hiring effects reported in Table 3 are understated.

121

hiring preferences.

I use state variation in age discrimination legislation from the 19603 to look at the
effects of age discrimination legislation on a variety of individual-level outcomes. The
aim is to draw inferences on what the estimates mean in terms of the effect of legislation
on ﬁrm hiring practices and the employment of older workers.

Consistent with evidence presented by Neumark and Stock (1999), the evidence
of this paper suggests that employment for older workers increases as a result of age
discrimination laws. The effect of the legislation on hiring outcomes is less clear. For
older workers, the estimates show that legislation does result in a decrease in the
probability of being hired and an increase in the probability of moving from employment
to unemployment. These dual results indicate that hiring outcomes may be worse for
older workers, suggesting that ﬁrms are increasingly preferring to hire younger workers
due to the increased use of delayed payment contracts.

Alternatively, however, the huge decrease in the probability of retirement that is
estimated may suggest that age discrimination legislation leads to an increase in the
length of time that older workers are retained at work and thus lowers the frequency in
which ﬁrms hire workers. This, coupled with the fact that results in this paper also show
that the probability of being unemployed increased for both young and old in states
adopting age discrimination legislation, suggests that that there were no shifts in
preferences in ﬁrm hiring due to the increased use of delayed payment contracts. Instead,
the legislation increased retention. This may have been due either to the reduction of
involuntary retirements that were previously part of delayed payment contracts or the

result of employer fears of being sued for wrongfully terminating the employment of

122

older workers.

123

‘ ﬂ“-

Table 1: Summary of the Existence and Coverage of State Age Discrimination Laws

 

 

State Age discrimination Year law was ﬁrst Age group that
law enacted in 1967 or enacted legislation covers
earlier (with during 1964 -1967
appropriate
enforcement)
Alabama No
Alaska Yes 1960 46-70
Arkansas No
Arizona No
California Yes 1961 40-64
Colorado No
Connecticut Yes 1959 40-65
Delaware Yes 1960 45-65
District of Columbia No
Florida No
Georgia No
Hawaii Yes 1963 18-70
Idaho Yes 1965 18-59
Illinois No
Indiana No
Iowa No
Kansas No
Kentucky No
Louisiana Yes 1934 18-49
Maine Yes 1965 18-70
Massachusetts Yes 1937 (amended 1966) 45-65
40-60 (as of 1966)
Maryland No
Michigan Yes 1965 35-60
Minnesota No
Missouri No
Mississippi No
Montana No

124

Table 1 (cont’d).

Nebraska Yes 1963 40-65
North Carolina No
North Dakota No
New Hampshire No
New Jersey Yes 1962 22-70
New Mexico No
Nevada No
New York Yes 1958 40-65
Ohio Yes 1961 40—65
Oklahoma No
Oregon Yes 1959 25-65
Pennsylvania Yes 1956 40-62
Rhode Island No
South Carolina No
South Dakota No
Tennessee No
Texas No
Utah No
Virginia No
Vermont No
Washington Yes 1961 40-65
Wisconsin Yes 1959 40-65
West Virginia No
Wyoming No

 

Note: Colorado, North Dakota, Texas, and Illinois each had age discrimination statutes on the
books by 1967, but there were no agencies entrusted with enforcing the legislation of the
respective states (Northrup 1980). Colorado and North Dakota residents are removed from the
sample, as their inclusion may contaminate the control group due the existence of the age
discrimination laws. No such action is taken for Texas or Illinois residents. The law in Texas
only applies to public sector employees, who are not included in my sample in the ﬁrst place.
Illinois residents are not excluded as their legislation took effect after the end of the sample
period (March 1967). Residents aged 40-65 years in all states were covered by age
discrimination legislation starting in 1968, but there was limited enforcement of the law until
1979. Rhode Island and Indiana had enforcement bodies but no penalties were set for breaking
the laws in these states. All remaining information in the table comes from either the Monthly
Labor Review (1964-1968), Neumark and Stock (1999), or Northrup (1980).

125

dado mm uni 295m

25. deﬁnes—go was .383 <m2m 63.3 3.58 .0? .30» .8 $9280 02:05 309850on =< 693—an mﬁon .«o tom—BBEQ 05 :0 acute
3 35.925 on :38 83838 05 35 8 OS .3 8:922: Pa 88533 =< .momoﬁcoaa E 83% 89:0 Basso 503680.5638388303
.2258 b:BBo.& 58:: Set 2.2% EcEenoEEo co sous—£3. conga—€86 owe me acute 2: .8 83:58 05 Ba cat83— ”8oz

 

 

mo> e Z 303825 Boziom<
8% on; 03.3.5, >883 88m
”eons—2: 30.980
80: A35 amd Ammé 8Q: Ammd
2.? mod- 8.7 NS. mad. 8;. women owe e228 do 8&8 e383
3m. 3 Sand
3.0 cad 20 «58» cm Bee: €8.83 x
30: $03
vNN mud .520 do So who» ow £8.83 on
€33 and
mad 5.— 20 £8» ow Bus: €8.83 on
God and
86 amﬁ .520 8 Bo wan» ow Boo—.83 x
AR: and 8:228.
. mt: .. .. El 3 concave mow—8o own E mace—53
9% Q S no a: E

 

.eoEeEam he haeeeem 2: no sesame: ceaseﬁre: one, do seem 2:. a 2%...

126

 

.2000 2 020 0.088 2:. 00000822 530w.“

000 .0080. 80000000 .0880. <m—2m .0880. 02.58 .090 .50» 00.0 28800 00208 8000050000 :< .002: >050: memos «0 5:38.08 05 :0 $0005
00 0058.52 00 >08 55800.0 05 .05 0v. 2: .3 0000208 0.5 00.08000 =< 080500.50 8 .5250 88.5 0.5080» 882800.5508008880:
20008 52030.8 508. 80.0 22 30: a 2.50 :0 0002080. 800088.520 0mm 00 300000 05 «0 0058000 05 05 0080002 ”0002

 

 

$30 30.9 Ewe § .0 0: .0 § Q
3.0- 3.0- 3.0- $0. $0. moo- ewes one 0228 0o 858 e383
Gwdv 92.8
mud—Al MW.~I ... U—O §0~A OW hO—UG—a whOMHOB X
80.9 a; .0
3.7 8.7 .520 00 20 0.50» on 802003 x
Amwdv C 2.9
0:- amA- 20 800» 00 508. 802.03 x
$0.8 Gad
Ed- Rd- 520 00 20 800» cc 802.03 x
Amwdv AS .NV 00020.80—
mﬁT mm._- ,3 005>00 0098 0mm 8 802003
08 g 5 p: n: 3
as

 

 

22 30: a 00. 52800 30:40. 00.52800 50 02.. 05 2.80 08
808 0_ 80 2 :00. .0350 55 00.05 20 080. m 0.0 l2..03....0.0|= 2050. all? 80.2....I03.....8.0h=|.0.- =<tl

 

 

3000 B0: 0.8 .80» 02.3 05 u0 ta .8.. 0.003 00: 20 8:00:05
5 552.: a. 28.5 388800 038 :52 use: Co 00228.. 20 .8 503003 8028088 54. do .800 2:. an e35.

127

000002008 500-080 0:0 .880 80000000 .0880 <mzm .0830 88.88 .080 .500 .80 800800 00808 8000005000

=< .008: 330: 0800 00 8030020 0.0 :0 800000 00 0220508 00 008 0058000 05 85 00 000 00 0000088 20 0208000 =< .80

00 000 08 .508. 805 850 002w 0002002 05 505 80803 .50—0 :0 50000 88200000 05 2002 0005 000020 5.0 88800 0.0.0. 8000—800.
000088.580 0mm 00 50000 .500 0.0 2002 .. . .. .83 02008 88800 0:0. 000050200 8 50000 22.5 050080 808800-8003008205:
80008 8000020 .508. 820 28 30: 0 0800 :0 0008080— 000088580 08 00 050000 05 00 0088000 08 20 0020030 ”802

 

 

600 $08 $00 308
N0.N- 0.0.0- 0m.m- 00. 7 00-3
30.: $0.00 30.3 80.:
SN- 00.0- 0w.- 00.0- 00-00
and $000 80.3 30.00
0m.- 00.0- 0nd- 00.0- 0m-mm
02.: G08 $00 3.8
000- :...0- 0N8- 00.0- 3.0m
80“.: $000 80.8 000. S
m _ .m- m 0 .0 00.0- 00.0- 00.3.
80.: 80.00 and 000.00
0:- vmd- 3.0- 3.0- 3.00
”00082 080 8:30:00
08 8 0.50003 :0 5000m—
08 S 0: 5
500 .800 .500 50.550 0080 05

05 80808080 8800 80850 003 .50—003 82.80 08.80 0800 0800 5 0000—080 00: 200.83 80.80 28 30: 00 000805
389$ .020 028 832 a use: 0.. 0828.. 20 no 502000.. 8030.838 030.. .88. 2:. 3. can...

128

.mobﬁo— mm uni 29:8 25. $5382.: Hoxéwa 28
.88..” £3335 £33m <m2m £38m .558 .0? .30» a8 «35:8 ova—2: 30:850on =< .32 com a wﬁoaotoaxo .8 mam—BBS; 05 co 308%
mu c8298.: on may: 83850 05 35 8 8— .3 332:8 Pa 83850 =< .momoﬁcoaa 5 325a E95 2353 .c3m_m:oo-bmoumacoow880:

208.: binge; .325 Bob neg Mono .8— 986: :o condime— =oum=uut8€ own we «quote 05 mo 88:58 05 0.8 39.83— ”202

 

 

 

A38 awe 82¢ 22.: §.: Ed
So N3 56 3o- 25. 8.? $5.. ow“ 8:58 Ho 0238 e383
2&8 2 EV
cad de- Bo who» on 825 38:95 x
8de A5:
mm; cad .520 8 20 who» on ﬂow—83 x
868 § . S
3: :d- 20 when co “one: €383 x
Amie Ana. 3
no; mm; .520 B 20 E3» co €8.83 x
93.9 $9: sown—ammo—
SA SA 3 @838 women“ own 5 £8.83
5V 5 S p: a: E
tllmo a; 3'0 8 386550 So 8: 05 :56 E—

 

 

 

808 2 So :308 338a 8a 35 805 E0

088 a 33.33 on; ﬁoxuoaéodlwaobso =<

355.8 2: no :3 05 3 toys 05 85 £8 32. ES .30» Sr:— 2: a: ﬁsh 68— aa 58 68.33 .2632:
an 352—3 mm 035.5, «ace—.095 :2. «.25 mama?— uc 3:332.— 2: .8 552m? catamarans ow< «a 895 2E. am «Sun.

129

.855: m: 00?. 03:80 05. 0:038:35 30:0? 0:: .88: £03830
.0880 <m—2m .0380 35:: .0m: :00» :0: .0.—0:200 00205 8000050090 =< 00:00.: 0:: 60.3383: w:_0: :0 b52008: 05 :0 0:00:00

00 00:05:03: 0: :08 00:05:00 05 :05 00 oo— 3 03:53:: 0:: 00:08:00 =< 00:05:05: :_ 300:0 0:0::0 0:06:80 ::0:m_m:00.b_0:0000000:0:0:
2000:: D2338: 30:: 80:: E05050: 0:: 50:32:88: :0 5:033: 5:23:86 0%: :0 0:00:00 05 :0 00:08:00 05 0:0 00:33: ”0:02

 

 

cumdv Ammdv ﬁnd: 309 ASN.8 8mg
3.: 2..— mﬁm £6 £6 56 0&5: 0w: 3:030 :0 028:0 0:00:03
308 Eng
wad mwd 20 0:00: on :0::: 0:00:03 0:
80.8 Awmdv
wﬁm- 0a.: :020 :0 20 0:00» om 0:00:03 x
80.0: 308
mmd- 3.0 20 0:00» cc :0::: 0:00:03 0:
3N: and:
mmw- mm: :020 :0 20 0:00» ow 0:00:03 0:
Che €08 5:023—
:06. 5.: .3 :0:0>00 wows: 0w: :_ 0:00:03
pm: 8 S p: S: 3
I::0::0::0m :JIQIIEE 0. 80::

 

3080.50”— .E: «ago—95:9 :0 55:000.:— 05 :0 gun—0E0:— :050£E::005 0w< :0 :00hﬂ 0.5. 6 030,—.

130

.53. 8 “Vaughn—BE on

2:05 83:58 2: .58 BE? 9 3858.. 29:8 a no 683 0.83 .592. 05 mo 338:8 05 35 .830 .58 83>» .858. 28 mo noun—anon
a ~23 38m 28853»: a no 63.3 0.3 E382— 05 mo 52 BE. .3 B>o $26365 0622: 8: moot €353 3.850 £32.28 .«o 95%

05 .3 mo own 05 :8: a: 28.83 bee E260 bﬁoﬁb noun—£3. $5 530 $73239: 03.3 woos—oi “Ea Bang 05 .8 “mo.— 05 5 33:58
05 Ease 9 com: 35 mo 29533 a m_ £5. datum 2953 $34.09 05 wctsc coax—mmwo— conga—tom? own 8923 833 E R3: 35
mE=E>€E .8 2953 a “8 Brae 8a 38 2: mo waEEmB 05 “a bowuﬁo 05 85 E 35 9.3.6365 no 898:: 05 £828 “ﬂu 05 5 ”802

 

 

8898
39m 2%va 386 #36 2h coho—9:25
808.9
wanna ammuovﬁv £86- 306m mum; @838
A3 5.8 3:302 Ho: “:6 3.83 .«o So €5.55
89: 33:3 $8.? Sq: 3 Ea 5» BE 5 E8 2:8 a 83%.»
8898 v.83 8m waioo. €5.55 98
5.6 $.va god 5? SN 3% 5.8 5 i8 2:8 a 83380
Amwoodv coho—mac mucous 98 :83
an: $3.8m; M: :3. Sn: v8 a £2 a .2 3% 8E 5 “.83 é 8x8.
”om B>o mﬁ=E>€E @8050 $39.89*
28m
05 5 cou££w2
cows—ammo. “can 3 @858 :ouﬂﬂwﬂ 282.
085 E “38:: owswno mo 3:35 gonna 89¢ 02?. 33058»: 2953 mmo
c0383...“ eunucmcoo $3 0383mm a 5 ~38: Z Sat 39:52

 

35m Eugen: a a agazus .5. 8.35% a. Ecusié K 2.3

131

W,V

 

Var

W0

 

 

WT

VT

 

Figure l

132

References

Burkhauser, R. and J. Quinn. 1983. “Is Mandatory Retirement Overrated? Evidence
from the 19708” Journal of Human Resources 18(3): 337-358.

EEOC Ofﬁce of Research, Information, and Planning. 1999. Age Discrimination in
Employment Act Charges (Washington: EEOC).

Gustman, A. and T. Steinmeier. 1986. “A Structural Retirement Model” Econometrica
54(3): 555-584.

Hellerstein, J., D. Neumark, and K. Troske. 1999. "Wages, Productivity, and Worker
Characteristics: Evidence from Plant-Level Production Functions and Wage Equations,"
Journal of Labor Economics 17(3): 409-446.

Hirsch, B., D. Macpherson, and M. Hardy. 2000. "Occupational Age Structure and
Access for Older Workers," Industrial and Labor Relations Review 53(3): 401-418.

Hutchens, R. 1986. “Delayed Payment Contracts and a Firm’s Propensity to Hire Older
Workers” Journal of Labor Economics 4(4): 439-457.

. 1988. “Do Job Opportunities Decline with Age?” Industrial and Labor Relations
Review 42(1): 89-99.

Hurd, R. 1990. "Research on the Elderly: Economic Status, Retirement, and
Consumption and Saving." Journal of Economic Literature 28(2): 565-637.

Idson, T. and R. Valletta. 1996. “Seniority, Sectoral Decline, and Employee Retention:
An Analysis of Layoff Unemployment Spells.” Journal of Labor Economics 14(4): 654-
676.

Johnson, R. and D. Neumark. 1997. “Age Discrimination, Job Separations, and
Employment Status of Older Workers: Evidence form Self-Reports” Journal of Human
Resources 32(4): 779-81 1.

Kotlikoff, L. and D. Wise. 1989. The Wage Carrot and the Pension Stick (Kalamazoo:
Upjohn Institute).

Lazear, E. 1979. “Why is there Mandatory Retirement?” Journal of Political Economy
87(6): 1261-1284.

. 1981. “Agency, Earnings Proﬁles, Productivity, and Hours Restrictions”
American Economic Review 71(4): 606-620.

Medoff, J. and K. Abraham. 1980. “Experience, Performance, and Earnings” Quarterly
Journal of Economics 95(4): 703-736.

133

Neumark, D. and W. Stock. 1999. “Age Discrimination Laws and Labor Market
Efﬁciency” Journal of Political Economy 107(5): 1081-1125.

Northrup, J. 1980. Old Age, Handicapped, and Vietnam-Era Antidiscrimination
Legislation (Philadelphia: University of Pennsylvania), pp. 191-222.

Oi, W. 1962. "Labor as a Quasi-ﬁxed Factor." Journal of Political Economy 70: 538-
555.

Orr, D. 1998. "Strategic Bankruptcy and Private Pension Default" Journal of Economic
Issues 32(3): 669-87.

Pergamit, M. and J. Veum. 1999. “What is a Promotion?” Industrial and Labor
Relations Review 52(4): 581-601.

Polsky, D. 1999. “Changing Consequences of Job Separation in the United States ?”
Industrial and Labor Relations Review 52(4): 565-580.

Rosen, B. and T. Jerdee. 1977. “Too Old or Not Too Old?” Harvard Business Review
55(6): 97-106.

Stewart, J. 1999. "Did Job Security Decline in the 19908?" Bureau of Labor Statistics,
mimeo

134

. RRRIES

"‘lilillllllllllllllllll '

 

5365