Full text of DOI:10.1162/003355300554818

ESTIMATING THE LABOR MARKET SIGNALING
VALUE OF THE GED*
J OHN H. TYLER
RICHARD J . MURNANE
J OHN B. WILLETT
This paper tests the labor market signaling hypothesis for the General
Educational Development (GED) equivalency credential. Using a unique data set
containing GED test scores and Social Security Administration (SSA) earnings
data, we exploit variation in GED status generated by differential state GED
passing standards to identify the signaling value of the GED, net of human capital
effects. Our results indicate that the GED signal increases the earnings of young
white dropouts by 10 to 19 percent. We find no statistically significant effects for
minority dropouts.
The positive correlation between education and earnings is
one of the consistent findings of the human capital literature. In
the early 1970s, however, Arrow [1973] and Spence [1973] formu-
lated an alternative, information-based explanation for the educa-
tion-earnings relationship. Over the last 25 years many econo-
mists have conducted empirical work aimed at exploring the
signaling hypothesis. It has proved difficult, however, to distin-
guish between human capital and signaling explanations of the
observed relationship between education and earnings.1
* We are especially grateful to J effrey Kling, J oshua Angrist, Lawrence Katz,
and Caroline Minter Hoxby for helpful comments. We owe thanks to J anet
Baldwin of the GED Testing Service of the American Council on Education, Linda
Headley Walker of the State Education Department of New York, Leslie Averna of
the Connecticut Department of Education, and J ohn Sojat of the Florida Depart-
ment of Education for their assistance in providing data on GED testers, and to
personnel of the Office of Research and Statistics at the Social Security Administra-
tion’s Baltimore Headquarters, particularly Peter Wheeler and Russell Hudson,
for providing and assisting with Social Security earnings data. The support of
these individuals and organizations does not imply their endorsement of the
contents of this paper.
We gratefully acknowledge the support provided by the National Center for
the Study of Adult Learning and Literacy (NCSALL) and the Rockefeller, Russell
Sage, Spencer, and Smith Richardson Foundations. Work supported by NCSALL
was supported under the Educational Research and Development Centers Pro-
gram, Award Number R309B60002, as administered by the Office of Educational
Research and Improvement/National Institute on Postsecondary Education, Librar-
ies, and Lifelong Learning, United States Department of Education. The contents
do not necessarily represent the positions or policies of the National Institute on
Postsecondary Education, Libraries, and Lifelong Learning, the Office of Educa-
tional Research and Improvement, or the United States Department of Education,
and you should not assume endorsement by the Federal Government.
1. For recent attempts see Lang and Kropp [1986], J aeger and Page [1994],
and Park [1994].
2000 by the President and Fellows of Harvard College and the Massachusetts Institute of
Technology.
௠
The Quarterly J ournal of Economics, May 2000
431

432
QUARTERLY J OURNAL OF ECONOMICS
Ideal data for identifying the returns to a signal would
contain exogenous variation in signaling status among individu-
als with similar levels of human capital. In this paper we
approximate that ideal in estimating the returns to a particular
labor market signal, the General Educational Development (GED)
credential, and in so doing we provide a test of the signaling
hypothesis for this credential.
The GED was introduced in 1942 to provide a way for
veterans without a high school diploma to earn a secondary school
credential. The GED program has evolved markedly so that today
the credential is the primary ‘‘second chance’’ route to high school
certification for school dropouts in the United States. Each year
more than one million young people drop out of school, and
eventually about one-third of them will acquire a GED.
Using interstate variation in GED passing standards, we are
able to compare individuals who all chose to try to obtain a GED
and who have the same GED test scores, but who differ in GED
status because of differences in the stringency of the passing
standards in their state of residence. We will argue that this
‘‘natural experiment’’ research design allows us to net out the
effect of human capital on earnings, leaving clear estimates of the
signaling value of the GED credential in the labor market.
Our estimates indicate that for young white dropouts who are
on the margin of passing the GED exams, the signaling value of
the GED increases annual earnings by 10 and 19 percent. We do
not find statistically discernible returns to the GED signal for
young minority dropouts, a puzzle we discuss later in the paper.2
In the next two sections we discuss the GED credential and
our identification strategy. We then present our data, methods,
and results. The paper closes with an extensive discussion of the
potential threats to our identification strategy, an interpretation
of our results, and a summary.
I. THE GED IN THE LABOR MARKET AND IN THE LITERATURE
The five tests in the GED battery cover mathematics, writing,
social studies, science, and ‘‘interpreting literature and the arts.’’
2. Because of data limitations we are only able to identify individuals in our
data as being members of one of two racial/ethnic groups: (1) white, non-Hispanics
and (2) everyone else. The latter group includes African-Americans, Hispanics,
and other minority group members. For expository clarity we refer to the two
groups as ‘‘whites’’ and ‘‘minorities’’ in this paper.

LABOR MARKET SIGNALING VALUE OF THE GED
433
The writing component of the test includes an essay, and the rest
of the battery consists of multiple choice questions. In the years
relevant for this paper, the GED Testing Service (GEDTS) estab-
lished a minimum passing standard that required a minimum
score of at least 40 or a mean score of at least 45 over the five tests
in the battery. However, the GED program is jointly supervised by
the GEDTS and the state education agencies, and each state is
free to set standards higher than the GEDTS minimum. Most
states do so, creating the sources of variation critical to our study.
In 1996 a record 758,500 dropouts attempted the GED exams,
and half a million were awarded the certificate. Table I illustrates
the trends behind that record number. Based on October CPS
data, the first row of Table I shows that the high school completion
rate for 18–24 year-olds trended upward slightly from 1975
through 1992, and has been relatively steady through 1996.
Beginning in 1988 the October CPS allows determination of
TABLE I
HIGH SCHOOL COMPLETION RATES OF 18–24 YEAR-OLDS FROM 1975–1996, AND BY
METHOD OF COMPLETION FROM 1988–1996
Yeara,b
Completion
method
1975 1980 1985 1988 1989 1990 1991 1992 1993 1994 1995 1996
Total
(percent)
Completed
Diploma
Alternativec
White, non-Hispanic
Completed
Diploma
83.6 83.9 85.4 84.5 84.7 85.6 84.9 86.4 86.2 85.8 85.3 86.2
80.3 80.5 80.6 80.7 81.2 81.2 78.8 77.5 76.5
4.2 4.2 4.9 4.2 5.2 4.9 7.0 7.7 9.8
87.2 87.5 88.2 88.7 89.0 89.6 89.4 90.7 90.1 90.7 89.8 91.5
84.4 85.1 84.8 85.2 85.7 85.5 84.2 82.6 81.0
Alternative
Black, non-Hispanic
Completed
Diploma
4.2 3.9 4.8 4.2 5.0 4.7 6.4 7.2 10.5
70.2 75.2 81.0 80.9 81.9 83.2 82.5 82.0 81.9 83.3 84.5 83.0
76.1 76.9 77.9 77.3 75.9 76.1 75.2 75.4 73.0
Alternative
Hispanic
4.8 5.0 5.3 5.2 6.1 5.8 8.1 9.0 10.0
Completed
Diploma
62.2 57.1 66.6 58.2 59.4 59.1 56.5 62.1 64.4 61.8 62.8 61.9
54.4 54.8 54.8 53.4 56.6 58.2 54.2 54.0 55.2
Alternative
3.8 4.7 4.2 3.1 5.5 6.1 7.6 8.8 6.7
Source: Tables 13 and A25 in Dropout Rates in the United States: 1997, U. S. Department of Education,
Office of Educational Research and Improvement.
a. The numbers for the years 1992–1997 reflect new wording of the educational attainment items in the
CPS.
b. The numbers for the years 1994–1997 reflect changes in the CPS due to newly instituted computer-
assisted interviewing and the change in the population controls to the 1990 Census-based estimates, with
adjustment for the undercount in the 1990 Census.
c. Completed high school by means of an equivalency test such as a GED exam.

434
QUARTERLY J OURNAL OF ECONOMICS
whether high school ‘‘completion’’ culminated in a regular diploma
or an alternative certification, and the GED is by far the most
common alternative form of certification [Cameron and Heckman
1993]. The second and third rows of Table I show that the steady
school completion rate over the 1990s masks opposite moving
trends: completion via a regular high school diploma trended
downward, while completion via GED certification trended up-
ward. The last three panels of the table show that these trends
were true for all racial/ethnic groups.
In general, GED holders are a relatively advantaged group
compared with uncredentialed dropouts. On average they com-
plete more years of schooling before dropping out, they have
higher levels of measured cognitive skills, and their parents have
more education. Even after controlling for these relative advan-
tages, however, there are reasons to believe that GED holders
might fare better in the labor market than dropouts without the
credential.
First, for dropouts with low levels of basic cognitive skills or
with English language deficiencies, raising skills to levels re-
quired to pass the GED battery may require considerable work
resulting in increased human capital. While Cameron and Heck-
man [1993] report a median study time for the GED of only about
twenty hours, the wording in the survey question on which their
estimates are based could lead to an underestimation of the total
GED preparation time. For example, it is not clear whether
dropouts who required participation in English as a Second
Language (ESL) or Adult Basic Education (ABE) courses prior to
GED preparation would count the time spent in those classes as
GED preparation on the survey question. Also, there is a long
right-hand tail to this distribution so that the least skilled
GED-holders may spend considerable time preparing to pass the
tests. Thus, the amount of human capital associated with acquir-
ing a GED, especially for low skilled dropouts, is still unclear.3
Returns to a GED can also be explained by a signaling
hypothesis. Assuming a distribution of underlying productivity in
the pool of dropouts, the more productive dropouts would like to
signal their higher productivity to employers and receive a higher
3. In addition to human capital returns related to test preparation, there may
be an additional human capital payoff from increased access to postsecondary
education and training associated with GED acquisition. In a later section we will
show that any human capital effects resulting from this route are likely to be very
small for the dropouts who drive our results.

LABOR MARKET SIGNALING VALUE OF THE GED
435
wage offer. The GED credential may serve as this signal in the
labor market. If the costs of acquiring the GED and the wage
offers based on GED status are such that the expected net benefits
of GED acquisition are negative for the less productive dropouts
and positive for the more productive dropouts, then we could
observe a separating equilibrium where GED acquisition is
positively correlated with productivity and wages.4
Since the returns to a GED can be explained by either human
capital or signaling theory, the empirical task is to estimate the
signaling value of the credential net of any human capital effects
associated with GED acquisition. A first step is establishing the
correlations between GED acquisition and labor market outcomes.
Table II presents estimates from log earnings regressions
using High School and Beyond data on individuals who were
sophomores in 1980. The dependent variable is the log of average
1990–1991 earnings, years when most of the sample would have
been 25–26 years of age.
Across all of the models, regular high school graduates
consistently show large advantages over dropouts, regardless of
race/ethnicity. In the models in columns 1 and 2 that do not control
for sophomore test scores, GED-holders have earnings that are
about 16 percent higher than uncredentialed dropouts—well
estimated for white dropouts and less well estimated for minority
dropouts. Adding test score controls in the third and fourth
columns, however, reduces the estimated GED effect for both
whites and minorities so that neither estimate is statistically
significant.
Since the positive but statistically insignificant estimates of
the ‘‘GED effect’’ pick up both human capital and labor market
signaling effect on earnings, one interpretation of these estimates
is that any signaling effect of the GED is very small. However, a
second interpretation we favor is that these baseline equations
are misspecified because they do not allow the effect of the GED on
earnings to vary according to the skill level of the dropout. This is
important in interpreting our main results, since our estimates
are based only on very low skilled GED-holders. We return to this
subject later.
4. The actual pecuniary costs associated with GED acquisition are generally
low, ranging from zero dollars in some states to around $50 in other states. Thus,
the binding costs for low productivity dropouts are more likely associated with
opportunity costs tied to long preparation time in upgrading skills or in the psychic
costs associated with the school-like preparation for and taking of the seven and
three-quarter hour battery of tests.

436
QUARTERLY J OURNAL OF ECONOMICS
TABLE II
EARNINGS REGRESSIONS FOR INDIVIDUALS IN THE HIGH SCHOOL AND BEYOND
SURVEY (DEPENDENT VARIABLE IS THE LOG OF THE AVERAGE OF 1990–1991
ANNUAL EARNINGS AND STANDARD ERRORS ARE IN PARENTHESES.)
Whitesa
Minoritiesb
Whites
Minorities
Intercept
9.518**
(0.057)
Ϫ0.397**
(0.019)
0.162*
9.320**
(0.079)
Ϫ0.306**
(0.030)
0.164
9.455**
(0.059)
Ϫ0.40**
(0.022)
0.094
9.199**
(0.081)
Ϫ0.298**
(0.032)
0.083
Female
GED
(0.072)
0.536**
(0.057)
(0.109)
0.581**
(0.073)
(0.072)
0.380**
(0.059)
0.012**
(0.002)
0.0002
(0.003)
0.002
(0.003)
Ϫ0.008*
(0.004)
0.007*
(0.003)
Yes
(0.109)
0.400**
(0.075)
0.008**
(0.002)
0.002
(0.005)
0.008ϳ
(0.005)
0.003
(0.005)
0.009*
(0.004)
Yes
0.131
2,810
High school graduate
Math test score
Reading test score
Writing test score
Science test score
Vocabulary test score
Region dummiesc
R2
Nd
Yes
0.11
5,403
Yes
0.087
2,810
0.133
5,403
a. White, non-Hispanic.
b. Everyone not white, non-Hispanic.
c. The regions are South, West, North Central, and Northeast.
d. Included in the sample were all dropouts (those with and without a GED as of the 1992 survey) and
individuals with at least a high school degree who were interviewed in the 1992 HSB survey, who were not in
the military in 1992 and not in college in both 1990 and 1991, who did not have zero earnings in both 1990 and
1991, and who had nonmissing values for their 1980 sophomore test scores.
In a widely cited study, Cameron and Heckman [1993] used
data from the National Longitudinal Survey of Youth (NLSY) to
show that male conventional high school graduates consistently
outperform male GED-holders in the labor market. Their work,
however, is less definitive on the question of how GED-holders
fare relative to uncredentialed dropouts. They find that male
GED-holders have wages that are between 3 and 6 percent higher
than uncredentialed male dropouts, but the results are not
statistically significant. They conclude that ‘‘. . .GED recipients lie
between [uncredentialed] dropouts and [regular high school]
graduates in their economic standing, but are much closer to
dropouts’’ [Cameron and Heckman 1993].

LABOR MARKET SIGNALING VALUE OF THE GED
437
While informative, their comparisons between uncreden-
tialed dropouts and GED-holders are potentially problematic for
two reasons. First, data limitations required that all uncreden-
tialed dropouts in the data—both those who chose not to acquire a
GED and those who attempted to acquire a GED but failed the
exams—form the comparison group in their study. It is question-
able whether either of these dropout groups provides a satisfac-
tory counterfactual.
Constraints imposed by model specification, however, are
potentially the most important and interesting reason to reexam-
ine the returns to the GED. Similar to the specifications used in
Table II, all of the Cameron and Heckman models constrain any
GED effect on wages or employment to be the same for all
dropouts, regardless of the skills they possess. Models that do not
allow for a differential effect of the GED by skill may miss
important elements of the way the GED works in the labor
market. This may be an especially important consideration regard-
ing a signaling hypothesis. For example, it may be that the GED is
a more important signal for dropouts who leave school with low
skills than for those who leave school with higher skill levels.
As a result of our identification strategy, our estimates
identify the signaling effect of the GED only for those dropouts on
the margin of passing the GED exams—that is, for the very least
skilled GED holders. If the returns to a GED are larger for this low
skilled population than for more highly skilled dropouts, then our
results are potentially consistent with those of Cameron and
Heckman.
II. ESTIMATING THE SIGNALING VALUE OF THE GED, USING
DIFFERENTIAL STATE GED PASSING STANDARDS
In this paper we take advantage of interstate variation in
GED passing standards to generate variation in GED status
among individuals with the same observed levels of human
capital. Our identification strategy is simple. Given two sets of
states, one with a lower GED passing standard than the other, we
compare individuals with marginally passing GED test scores in
the lower passing standard states with individuals who have the
same GED test scores in the higher standard states. However,
because of the more stringent GED passing standard, our compari-
son individuals in the higher passing standard states will not
have the credential. Once we have accounted for differences in

438
QUARTERLY J OURNAL OF ECONOMICS
state labor markets, we can estimate the impact of the GED on
earnings by comparing the mean earnings of the two groups. Since
these individuals have the same GED test scores, any human
capital effect on earnings captured by the GED test scores is
removed from the estimates. There may be other forms of
unobservable human capital differences between the two groups
that would bias our estimates. We address the most obvious and
problematic of these in a later section in the paper and conclude
that they do not pose serious problems for the main estimates in
this paper. Thus, we interpret the estimates from this methodol-
ogy as the signaling effects of the GED on the earnings of dropouts
who would choose to obtain the GED and are at the margin of
passing.5
Our data contain information on 1990 GED candidates from
45 states. All of the states in our sample define their GED passing
standards by a combination of the minimum test score and the
mean score over the five tests. In our research we are able to use
three of the possible seven different passing standards that
existed across the nation in 1990. They are, in ascending order of
difficulty:
● a minimum score of at least 40 or a mean score of at least
45,
● a minimum score of at least 35 and a mean score of at least
45, and
● a minimum score of at least 40 and a mean score of at least
45.
The nature of our data, in combination with the standards above,
dictate that we use minimum-mean score combinations to con-
struct different GED score groups. The score groups were con-
structed so that, based on score group and state in which the GED
was attempted, we can be assured of GED status. Thus, we
constructed ten GED ‘‘score groups’’ based on the intersection of
the minimum and mean score ranges represented in the rows and
columns of Table III.
In Table III individuals in score groups 1 and 2 would have
GED scores below the passing standard in any state in 1990.
Individuals in score group 5 or higher would have scores above the
passing standard of any state in our data in 1990. It is only in
score groups 3 and 4—the ‘‘affected score groups’’—that we find
5. Thus, our results are not estimates of the effect of the GED on the earnings
of a random dropout.

LABOR MARKET SIGNALING VALUE OF THE GED
439
TABLE III
GED SCORE GROUPS FORMED BY COMBINING MINIMUM AND MEAN SCORES
(OUTLINED CELLS ϭ VARIATION IN GED-STATUS BY STATE, DARK SHADING ϭ ALL
POSSESS GED, NO SHADING ϭ NONE HAVE GED.)
Mean score
Minimum score
Ͻ45
Ͼ ϭ 45
20–34
35–39
40–44
45–46
47–48
49–50
51–52
53ϩ
Score group 1
Score group 2
Score group 3
Score group 4
Score group 5
Score group 6
Score group 7
Score group 8
Score group 9
Score group 10
variation in GED status by the state where the GED exams were
taken in 1990. In each of these two scoring ranges individuals in
states with the lower passing standard—the ‘‘treatment states’’—
have a GED, while individuals in ‘‘comparison states’’ do not, due
to higher passing standards.
Thus, we have these three natural experiments to test the
signaling hypothesis as it pertains to the GED. We denote these
experiments by the affected score group in each experiment.
—Experiment 4, where variation in GED status by state is in
score group 4, the treatment states are those states that award a
GED in score groups 4 and higher, and the comparison states are
those that award a GED in score groups 5 and higher.
—Experiment 3, where variation in GED status by state is in
score group 3, the treatment states are those states that award a
GED in score groups 3 and higher, and the comparison states are
those that award a GED in score groups 5 and higher.
—Experiment 3*, where variation in GED status by state is
in score group 3, the treatment states are those states that award
a GED in score groups 3 and higher, and the comparison states are
those that award a GED in score groups 4 and higher.
Since the GED passing standards are set by the individual
state education departments, and since these agencies are embed-
ded in the state political system, the interstate variation in pass-
ing standards on which we rely is not transparently exogogenous.
The fact that GED passing standards in general, and the ones we
exploit in particular, have been relatively stable over time,

440
QUARTERLY J OURNAL OF ECONOMICS
however, suggests that states do not set the GED passing stan-
dard based on current state economic conditions.6 We assume that
the variation in passing standards we use is exogenous condi-
tional upon state fixed effects.
III. RESEARCH DESIGN
A. Data
To exploit the sources of GED signaling status described
above, we needed information on demographic characteristics,
GED test-scores, state in which tested, and earnings for a sample
of dropouts. The first three types of information came from
administrative records of the GED Testing Service and the state
education agencies of New York, Florida, and Connecticut.7 In the
data from New York, Florida, and Connecticut, we have informa-
tion on the universe of dropouts aged 16–21 who last attempted
the GED tests in 1990. For data obtained from the GED Testing
Service, we have, from each of 42 states, a sample of dropouts who
were aged 16–21 when they attempted the GED battery in 1990.
Using Social Security numbers supplied by GED candidates
at the time of testing, Social Security Administration (SSA)
programmers attached Social Security-taxable earnings data to
our GED test-file data, providing valid matches for about nine out
of ten dropouts.8 The resulting micro-level data set contained
demographic and GED test score information collected when the
dropouts in the sample attempted the GED in 1990, together with
their Social Security earnings records for each year from 1988 to
1995. The SSA does not release individual earnings to research-
ers. As a result, the SSA used the micro-level data to construct an
aggregated data set satisfying SSA confidentiality requirements
and our research design.9 The final aggregated data set consists of
990 cells, defined by the state where the GED was attempted,
6. For example, out of the 42 states used in our study, only Louisiana in 1982,
California in 1984, New York in 1985, Arkansas in 1987, and Washington in 1988
changed their GED passing standard between 1980 and 1990.
7. We augment the GED Testing Service data with data from New York,
Florida, and Connecticut because these states were missing from the GED Testing
Service data.
8. Providing SSNs on the GED test form is voluntary, but inspection of the
Florida and Connecticut samples indicates that about 98 percent of examinees 21
and younger (our target group) provide this information.
9. SSA confidentiality regulations require that no information be released in
cells where there are fewer than three observations. Other SSA limitations are
discussed in the Data Appendix.

LABOR MARKET SIGNALING VALUE OF THE GED
441
GED score group, gender, and whether the GED candidate was
white or minority.10 Each cell contained average FICA earnings
for a particular year for the individuals in the cell, cell frequency,
and the standard deviation of earnings.
Table IV presents summary statistics for the aggregated
data. In the table we possess the universe of young dropouts who
attempted the GED in 1990 in New York, Florida, and Connecti-
cut, and so, relatively speaking, the number of individuals contrib-
uted by those states is large (column 1). The analytic sample is
about 70 percent white (column 6) and 55 percent male (column
7). Columns 8 and 9 present mean earnings five years after
dropouts attempted the GED battery, and these earnings are
generally low.11 While not reported here, the variances of 1995
earnings for any particular subset of our sample are generally
large relative to the mean, as is typical with annual earnings.
Column 2 shows the minimum score group in which individuals
would have a GED in a given state, reemphasizing the three
different passing standards present in our data. Columns 3, 4, and
5 of Table IV designate whether a state falls into the treatment-
group or comparison-group in each experiment that we use,
demonstrating how some states contribute to the treatment group
in one experiment and to the comparison group in another.
We use New York and Florida to form the comparison group in
two of our three experiments. We single out these states because
New York and Florida data contain the accumulated best scores of
dropouts who have either attained a GED or who have ‘‘stopped
out’’ as of 1990.12 The 1990 GEDTS data, however, contain only a
one-year ‘‘snapshot’’ of GED-attempters, and some unknown
portion of individuals with failing scores in the 1990 GEDTS data
will have subsequently retaken and passed the battery in years
1991–1995.13 Therefore, any comparison group constructed from
10. To avoid small cells in the affected score groups that would be censored by
the SSA, we could not stratify more finely on race/ethnicity and on age.
11. Earnings in New York are low compared with those in other states. In
additional data we find that the positive earnings of dropouts in New York are the
highest of any state, indicating that, in our data, many young dropouts in New
York had zero Social Security taxable earnings in 1995.
12. By this we mean that we are certain that nonpassers in the New York and
Florida data did not return after 1990 to retake the exams and potentially receive a
GED in any of the years prior to our measurement of earnings. In these states,
anyone who took any portion of the tests after 1990 would not be included in the
data, regardless of their passing status.
13. For example, using the Connecticut and Florida individual-level data,
we find that about 70 percent passed the GED on the first try in 1988. Out of those
who did not pass, about 40 to 50 percent retested and acquired a GED within two

442
QUARTERLY J OURNAL OF ECONOMICS
TABLE IV
SUMMARY STATISTICS AND TREATMENT/COMPARISON-GROUP DESIGNATION BY STATE
FOR YOUNG DROPOUTS WHO ATTEMPTED THE GED IN 1990
Col. 2
Minimum
score
Cols. 3, 4, & 5
Treatment or
comparison in . . . who are . . .
Cols. 6 & 7
% of sample
Cols. 8 & 9
Mean 1995
earnings
State
groupa
Col. 1
N
group
for GEDb Exp3 Exp3* Exp4 White Male No GED GED
AZ
CT
FL
KY
NY
TN
TX
VA
3457
1692
17905
1949
19134
1668
5027
1657
2852
4
4
5
4
5
4
3
4
4
4
4
4
—
—
C
—
C
C
—
—
C
—
C
T
C
C
C
T
—
C
T
C
T
—
T
T
65% 57% 7176
71% 51% 8779 10413
9336
74% 56% 7767
83% 58% 6797
53% 56% 6777
94% 50% 8479
60% 51% 7755
78% 59% 7905
79% 54% 7559
75% 55% 8184 10327
91% 52% 8273 9727
85% 55% 9031 10364
9694
8149
8764
9079
9286
9662
8813
—
T
—
—
—
—
—
Miscc
Atlantic states 3237
New England 1801
North Central 2313
Low pass
T
T
T
C
C
statesd
Miscellaneous
statese
4138
3
T
T
—
76% 55% 7899
8732
4202
4713
4
5
—
—
C
C
T
—
80% 55% 7750
64% 57% 8017
9032
9351
WA and CA
High pass
statesf
4255
80000
5
—
C
—
—
—
67% 55% 6258
8669
—
Totalg
69% 55%
—
a. States AZ through VA had enough observations to serve as their own ‘‘state group.’’ Other states had to
be grouped into either geographic or other logical groupings due to small numbers of observations. See the
Data Appendix for a listing of the states in each group not otherwise detailed here.
b. There were seven different GED passing standards in 1990 that ranged from the low standard
represented by score group 3 to that of Wisconsin, which required a mean of 50 and a minimum of 40.
c. Alaska, Hawaii, Idaho, Montana, Alabama, Iowa, Kansas, and Nevada.
d. This group is composed of Mississippi, Louisiana, and Nebraska, and together with Texas they
comprise the group of states with the lowest passing standard, a mean of 45 or a minimum of 40.
e. This group is composed of Colorado, Illinois, Pennsylvania, and Minnesota, the only states where there
were statistically significant differences between individuals with and without imputed writing scores. (See
above and the Data Appendix for a discussion of imputed writing scores in the GEDTS data.)
f. This group is composed of Arkansas, Delaware, Maryland, Oklahoma, South Dakota, and Utah, and
along with Washington and California, these states have the same higher standards as do New York and
Florida.
g. Note that with about 36,000 observations combined, New York and Florida heavily influence the
statistics in this row.
nonpassing GED candidates in the GEDTS data will be contami-
nated with individuals who subsequently obtained a GED. This
occurs in Experiment 3*, and the results from this experiment
years. The result is that the overall pass rate went from 70 percent on the first try
in 1988 to an eventual pass rate of about 85 percent by the end of 1990.

LABOR MARKET SIGNALING VALUE OF THE GED
443
provide a conservative, lower bound estimate of the signaling
effect of the GED.
B. Statistical Analyses
The key treatment-control comparison is captured in the
following model describing the annual earnings of dropouts in an
unobserved individual-level world:
(1) Yi ϭ ␤0 ϩ ␤ЈSTi ϩ ␤ЈSGi ϩ ␣(Ti ء ASGi) ϩ ␤3Femalei ϩ ui,
1
2
where
Y ϭ individual i’s annual earnings in some year after the
GED was attempted;
ST ϭ vector of state dummies;
SG ϭ vector of dummies representing the GED score group
containing individual i;
T ϭ 1 if individual i is in a lower standard treatment state, 0
otherwise;
ASG ϭ 1 if individual i is in the affected score group and 0 if
individual i is in a higher score group, where the ‘‘af-
fected score group’’ is either score group 3 or score group
4 depending on the experiment;14
Female ϭ 1 if individual i is a female, 0 if a male;
u ϭ error term, with zero expectation and orthogonal to
predictors.
For reasons we explain later, we stratify our data into white
and minority dropouts, and we would like to fit equation (1)
separately for each of these two race/ethnicity groups. However,
we cannot fit the hypothesized model in equation (1) because we
do not have individual-level data on earnings. Fortunately, based
on equation (1), the parameter of interest, ␣, representing the
effect of the GED on the earnings of dropouts whose GED-status is
affected by state-of-residence is
(2) ␣ ϭ E [Yi 0Ti ϭ 1,ASGi ϭ 1] Ϫ E [Yi 0Ti ϭ 0,ASGi ϭ 1]
Ϫ (E [Yi 0Ti ϭ 1,ASGi ϭ 0] Ϫ E [Yi 0Ti ϭ 0,ASGi ϭ 0]),
14. We omit those who score below the affected score group range because, in
the GEDTS data, these score groups are contaminated with individuals who have
a GED, and whom we cannot identify. This contamination downwardly biases
estimates of a. Estimates using these lower score groups are available from the
authors upon request. These ‘‘contaminated’’ estimates are consistently lower than
our preferred estimates, regardless of the experiment and the racial group.

444
QUARTERLY J OURNAL OF ECONOMICS
suggesting that a method-of-moments estimator of ␣ can be
constructed from sample means. This difference-in-differences
estimator is
␣ˆ ϭ (YT Ϫ YC) Ϫ (YTHi Ϫ YCHi),
where
Y ϭ mean earnings of any group;
T indexes the treatment group—individuals in the affected score
group who are in a low-passing-standard state in a given
experiment, and thus have a GED;
C indexes the comparison group—individuals in the affected score
group who are in a high-passing-standard state, and thus do
not have a GED;
THi indexes individuals in treatment states that are in a scoring
group higher than the affected score group, and thus have a
GED;
CHi indexes individuals in comparison states that are in the same
high scoring group as the THi group, and thus also have a
GED.
We use all individuals in score groups 5 through 10 to form
the high comparison groups, THI and CHi, in the difference-in-
differences estimation. This avoids arbitrarily deciding which
score group to use in this capacity and maximizes sample size in
the high comparison groups.15 Finally, notice that the second
difference in the difference-in-differences estimator, YTHi-YCHi
,
removes state fixed effects from the estimate.
We can estimate ␣, and conduct hypothesis tests, using
SSA-provided aggregate within-group statistics on earnings.16 We
identify ␣ using two assumptions. The first is that the differential
state GED passing standards generate exogenous variation in
GED status. This is embodied in the assumed independence of the
interaction and error terms. The second assumption is that
equation (1) contains no higher order interactions. Later, we
examine empirical evidence regarding violations of these assump-
15. Additional analyses verify that our findings are not qualitatively different
when we use any one of the score groups as the high comparison group, as opposed
to using everyone simultaneously.
16. The standard errors for the estimator are obtained in the standard
manner, with no assumptions of equality of variances across groups.

LABOR MARKET SIGNALING VALUE OF THE GED
445
tions, confirming that any resulting biases have little impact on
our estimates.
In computing our difference-in-differences estimates, we
weight the calculation of within-group earnings to account for the
different gender distributions within treatment and comparison
groups, and across states within these groups. In other analyses
not reported here, we confirm that our findings are not sensitive to
the choice of weights.
IV. RESULTS
A. Difference-in-Differences Estimates of the Impact of the GED
on the Earnings of Young Dropouts
Table V presents separate estimates of the effect of the GED
on 1995 earnings for white and minority dropouts from three
different experiments involving young dropouts who attempted
the GED in 1990.17 We focus initially on estimates generated by
experiments 3 and 4. In these experiments, acquisition of a GED
is associated with about a $1500 increase in annual earnings for
young white dropouts, an increase of approximately 19 percent.
The point estimates for whites from the two different experiments
($1473 and $1531) are remarkably similar given that the states
making up the treatment group in each experiment are different.
Both experiments 3 and 4 use New York and Florida as
comparison-group states. Experiment 3*, like experiment 3, uses
Texas, Louisiana, Mississippi, and Nebraska as treatment states.
However, experiment 3* excludes New York and Florida and
instead uses states from the GEDTS data in the comparison
groups, including states where the GED is awarded in score group
4. As explained above, when GEDTS data are used to construct
the comparison group, the group will be contaminated by an
unknown number of individuals who obtained a GED after 1990.
Because of this, the 1995 earnings of the comparison group will be
inflated (assuming a positive treatment effect of the GED on
earnings). Thus, the $907 difference-in-differences estimate in
experiment 3* represents a downwardly biased, conservative
estimate of the impact of the GED on earnings.
17. Preliminary analyses indicated that results for white males and females
were similar and the results for minority males and females were similar. Thus, to
increase statistical power, we aggregated across the genders, retaining the
white-minority distinction.

446
QUARTERLY J OURNAL OF ECONOMICS
TABLE V
DIFFERENCE-IN-DIFFERENCES ESTIMATES OF THE IMPACT OF THE GED ON 1995
EARNINGS OF DROPOUTS WHO TESTED IN 1990 (STANDARD ERRORS ARE
IN PARENTHESES.)
Experiment 4
State passing
Experiment 3
State passing
Experiment 3*
State passing
Low-High
standard
High contrast
Low-High
standard
High contrast
Low-High
standard
Low High contrast
standard is
Low
Panel A: Whites
standard is
standard is
Low
Test score is
Low
9628 7849
(361) (565)
9981 9676
1779
(670)
305
(103)
1473*
(678)
9362 7843
(400) (312)
9143 9165
1509
(507)
Ϫ23
(149)
1531**
(529)
9362 8616
(400) (219)
9143 9304
(135) (135)
746
(456)
Ϫ162
(150)
907ϳ
(481)
High
(80)
(65)
(135)
(63)
Difference-in-differences
for whites
Panel B: Minorities
Test score is
Low
6436 8687
(549) (690)
7560 8454
Ϫ2252
(882)
Ϫ894
(207)
Ϫ1357
(906)
7005 7367
(347) (347)
7782 8375
Ϫ363
(495)
Ϫ593
(233)
231
7005 6858
(347) (290)
7782 7568
(214) (133)
147
(452)
214
(252)
Ϫ67
(518)
High
(184)
(96)
(214)
(93)
Difference-in-differences
for minorities
(548)
** ϭ significant at the 0.01 level, * ϭ significant at the 0.05 level, ϳ ϭ significant at the 0.10 level.
Experiment 4: Test Score Low: score group ϭ 4; Test Score High score groups ϭ 5–10.
Passing Standard Low: 35 minimum score and 45 mean score; Passing Standard High: 40 minimum score
and 45 mean score.
Low Passing Standard states: All states except for TX, LA, MS, NE, FL, NY, CA, WA, and CT; High
Passing Standard states: NY and FL.
Experiment 3: Test Score Low: score group ϭ 3; Test Score High score groups ϭ 5–10.
Passing Standard Low: 40 minimum score or 45 mean score; Passing Standard High: 40 minimum score
and 45 mean score.
Low Passing Standard states: TX, LA, MS, and NE; High Passing Standard states: NY and FL.
Experiment 3*: Test Score Low: score group ϭ 3; Test Score High: score groups ϭ 5–10.
Passing Standard Low: 40 minimum score or 45 mean score; Passing Standard High: 35 minimum score
and 45 mean score.
Low Passing Standard states: TX, LA, MS, and NE; High Passing Standard states: all states except TX,
LA, MS, NE, NY, FL, and CT.
The results for nonwhite dropouts differ sharply from the
results for white dropouts. The three experiments yield no
statistically significant evidence that acquisition of a GED results
in higher earnings for minority dropouts. We return to the
minority results later. Based on the results from experiments 4, 3,
and 3*, our estimates are robust to the use of different treatment
and comparison groups.
B. Timing of GED Treatment Effects
To avoid underestimating the impact of the GED by measur-
ing earnings too close to receipt of the credential, we have

LABOR MARKET SIGNALING VALUE OF THE GED
447
concentrated on using the 1995 earnings of young dropouts who
attempted the GED in 1990 as our outcome measure. To explore
both pretreatment earnings differences as well as the possibility
that benefits accruing to GED-acquisition do indeed take time to
develop, we construct equivalent difference-in-differences esti-
mates using earnings in the years 1988–1989 and 1991–1995.
These comparisons continue to use 1990 test-takers as the
analytic sample, and we concentrate on white dropouts.
Figures I–III show year-by-year difference-in-differences esti-
mates for experiments 4, 3, and 3*. Posttreatment earnings in
Figures I and II (years 1–5 after the GED attempt) suggest that it
takes time for the GED to pay off for young white dropouts. Figure
I shows that, in the first two years after GED-acquisition,
GED-holders actually earn less than uncredentialed dropouts
with the same GED scores. Over time GED-holders in the
treatment group gain on their uncredentialed counterparts in the
comparison group, so that by the fifth year after GED-acquisition,
they are earning $1473 more per year. And, while it appears from
Figure II that treatment individuals in experiment 3 can make
the GED pay off sooner than those in experiment 4, the trends are
the same. Estimates from experiment 3* are not well suited for
studying the timing of GED effects because each year after 1990
the comparison group becomes increasingly contaminated by
retesters who return to obtain a GED. We consider pretreatment
earnings in a later section.
V. THREATS TO IDENTIFICATION IN THE NATURAL
EXPERIMENT ESTIMATES
There are four general sources of potential bias in our
primary results in Table V. First, to the extent that individuals
use the GED to access postsecondary education or training, our
failure to control for these human capital effects would lead to
overestimates of the signaling value of the GED. Also, there are
three different types of endogeneity bias that could affect our
estimates. We explore these potential biases in this section.
A. Postsecondary Education and Training Effects
Most academic postsecondary programs and many training
programs, as well as the loan and grant programs that can help
pay for them, require a school leaving certificate. For most
dropouts who desire admittance to these programs, the GED
serves this purpose. It follows that acquisition of a GED could lead

448
QUARTERLY J OURNAL OF ECONOMICS
FIGURES I–III
Pretreatment and first through fifth year Difference-in-Differences Estimates for
Young White Dropouts. (** ϭ Significant at the 0.01 Level; * ϭ Significant at the
0.05 Level; ϳ ϭ Significant at the 0.10 Level.)

LABOR MARKET SIGNALING VALUE OF THE GED
449
to higher average levels of human capital through increased
access to postsecondary education and training programs. Since
we do not account for this potential human capital effect, interpre-
tation of our estimates as returns to a signal could overstate the
case. Analysis of other data indicates, however, that these types of
human capital effects do not seriously threaten the signaling
interpretation of our estimates.
In HSB data we find that the least skilled GED-holders (those
in the lowest quartile of the sophomore math test) accumulated on
average less than five postsecondary credits by age 28. Based on
an academic year being equal to 30 credits, this is one-sixth of an
academic year. Other authors [Kane and Rouse 1995; Murnane,
Willett, and Boudett 1999] estimate that a year of postsecondary
academic work has a return of about 5 percent relative to high
school graduation. This suggests that less than one percentage
point (1/6 ϫ 0.05) of any estimate based on low skilled GED-
holders might be due to postsecondary education human capital
accumulation rather than labor market signaling.
Regarding postsecondary training, Murnane, Willett, and
Boudett [1997] use the NLSY to show that both male and female
GED-holders are more likely to obtain off-the-job training (i.e.,
training in a government-sponsored program or a proprietary
school) than are comparable uncredentialed dropouts.18 However,
in later work [Murnane, Willett, and Boudett 1999] they also find
no effect of this off-the-job training on the wages of male GED-
holders, and Boudett [1998] finds no effect of off-the-job training
on the wages of female GED-holders. Taken together, these
analyses suggest that neither postsecondary education nor off-
the-job training carries substantial earnings effects for low skilled
dropouts who are on the margin of passing the GED exams.
B. Endogeneity Bias Associated with Individual Behavior
Interpreting the results in Table V as the causal impact of the
GED signal on earnings rests on an assumption that our treat-
ment and comparison groups are balanced on unobservable
characteristics that might influence their earnings. If the differ-
ent passing standards influence individual behavior in systematic
ways, then this assumption may be violated. There are three
decisions that might be influenced by differential GED passing
18. Since the GED signal may help dropouts obtain jobs that offer training,
any returns to post-GED on-the-job training are returns to the signaling value of
the GED.

450
QUARTERLY J OURNAL OF ECONOMICS
standards across states: the decision to attempt the test, the
decision to migrate to another state, and the decision about how
much effort to exert on the test. Each of these potential decisions
has different implications for our identification strategy.
First, when faced with higher standards, some individuals on
the margin who would otherwise test may elect not to attempt the
GED battery. Assuming that this behavior is negatively correlated
with productivity-enhancing traits such as persistence, self-
confidence, and motivation, then this type of selection would
result in an overestimate of the mean earnings of potential
GED-holders in comparison group states. The net effect would be
a downward bias in the estimated effect of the GED on earnings.
Endogenous migration decisions are a problem if more moti-
vated dropouts tend to migrate to lower standard states in order
to increase their chances of passing the exams or because of other
factors that are also related to earnings. We examine this possibil-
ity by examining the migration patterns of dropouts in the NLSY.
Rows 1, 2, and 3 in Panel A of Table VI represent the mobility
patterns of GED-holders who were in low, medium, and high
TABLE VI
STATE MIGRATION PATTERNS BETWEEN LOW, MEDIUM, AND HIGH GED PASSING
STANDARD STATES FOR GED-HOLDERS AND UNCREDENTIALED DROPOUTS IN THE
NLSY (WEIGHTED TABULATIONS, ROW PERCENTAGES SUM TO 100)
Panel A:
GED-holdersa
GED passing standard in state of
residence in year of GED acquisition
Passing standard in state
of residence at age 14
Low
Medium
High
Low
Medium
High
89.0%
3.4%
2.1%
8.5%
90.9%
8.1%
2.6%
5.6%
89.9%
Panel B:
Uncredentialed dropouts
Passing standard in state
of residence at age 14
Low
GED passing standard in
state of residence at age 21
Low
86.7%
2.5%
1.5%
Medium
6.6%
91.9%
9.0%
High
6.8%
5.6%
89.5%
Medium
High
a. To match the age group used in this paper, the percentages in Panel A are based on the 512
GED-holders in the NLSY who were aged 16–21 when they obtained their GED and who obtained the
credential in 1979 or later. The migration patterns of all GED-holders in the NLSY are very similar to this
younger age group. We excluded those dropouts who obtained their GED prior to 1979 because we cannot
determine the state of residence at GED acquisition for earlier years. Finally, there are five states (SC, NM,
ND, WI, and NJ ) with standards that do not fit our low, medium, and high categories.

LABOR MARKET SIGNALING VALUE OF THE GED
451
standard states (respectively) at age fourteen. The percentages
across the rows sum to 100, so that 89 percent of the individuals
who started out in low standard states were still in a low standard
state when they obtained their GED, 8.5 percent had moved to a
medium standard state, and 2.6 had moved to a high standard
state by the time of GED acquisition.
Systematic migration from high to low standard states would
be indicated by relatively little migration out of the upper
left-hand cell and relatively more migration to the left out of both
the middle cell of the table and the lower right-hand cell of the
table. Instead, we see that about 10 percent of GED-holders,
regardless of the passing standards in their state of residence at
age fourteen, obtain their GED in a state with a higher or lower
standard. Furthermore, among those GED-holders who do move
between age fourteen and GED acquisition, there is no evidence of
systematic movement to lower standard states.
As a further check on the mobility patterns of GED-holders,
Panel B of Table VI presents the migration probabilities (as of age
21) of the uncredentialed dropouts in the NLSY. The results are
similar to those of Panel A, again suggesting that there are no
systematic links between GED state passing standards and the
migration decisions of young dropouts.19
The third source of bias that could result from individual
responses to state passing standards relates to test-taking effort.
The concern is that individuals in our treatment group may have
higher unobserved skills than individuals in the comparison
group. This situation could occur if some of the more highly skilled
dropouts in the lower standard states exerted just enough effort to
narrowly pass the exams. In this case our treatment group would
be contaminated with individuals whose skills were higher than
their GED scores indicated. We can explore this potential source
of bias with our data.
To do this, we assume that some individuals in lower stan-
dard states whose true skills are higher than what is required to
‘‘just pass’’ exert just enough effort to pass the GED battery. As a
result, some of these higher ability individuals are in score groups
3 or 4 in our data. Figure IV displays the sorting that would result
19. A separate problem would arise if GED candidates did not migrate, but
simply crossed state lines to attempt the test in a lower standard state. We note
that all states have residency requirements to attempt the GED. Furthermore, the
chief state GED officer in Texas informed us that proof of residency was often
requested in testing sites situated close to the borders with other states [Erwin
1999].

452
QUARTERLY J OURNAL OF ECONOMICS
FIGURE IV
Score-Group Sorting around Low and Medium GED Passing Standards (Shaded
Cells ϭ Score Groups Where the GED is Awarded)
from this assumption. The top row of Figure IV depicts the ten
observed GED score groups. The next two rows depict hypotheti-
cal distributions of true, underlying skills across the ten GED
score groups used in this study when some higher ability individu-
als exert just enough effort to pass. The middle row illustrates a
hypothetical distribution of true skills for low standard states
(where the GED cutoff is at score group 3) and the bottom row
illustrates this situation for medium standard states (where GED
cutoff is at score group 4). As an example, according to the middle
row of Figure IV, observed score group 3 in low standard states not
only has individuals whose true underlying skills place them in
this score group (‘‘true 3s’’), but also individuals with higher skills
(‘‘true 5s and 6s’’) who exerted just enough effort to pass. This type
of ‘‘ability contamination’’ of score group 3 in the low standard
states would cause us to overestimate the effect of the GED on
earnings.
We can adjust our estimates to account for this type of
differential effort by assuming the following:
(1) the observed proportion of individuals in the affected
score group in the higher standard states represents what
the true skill distribution in the affected score group in
the lower standard states would be in the absence of
differential effort across the states20 (e.g., for experiment
20. As an anonymous reviewer pointed out, this assumption would likely not
hold if there were differential participation rates across lower and higher standard
states. However, we assume that fewer individuals attempt in the higher standard
states. If more were to attempt, then presumably the percentage who ‘‘just fail’’ in
these higher standard states would go up, and based on the algorithm we describe,

LABOR MARKET SIGNALING VALUE OF THE GED
453
3 the correct proportion of true 3s in the low standard
states is given by the proportion of observed 3s in the
higher standard states), and
(2) the mean earnings of higher skilled individuals who score
less than their true abilities are the same as the observed
mean earnings of similarly skilled individuals who sort
themselves into their true score groups.
To illustrate how we use these assumptions to generate
adjusted estimates, consider experiment 3 using just Texas, a low
standard state, and Arizona, a medium standard state. Individu-
als in Texas in score group 3 have a GED, while those in Arizona in
score group 3 just missed getting a GED. Say, that we see 4
percent of the distribution of GED attempters in Texas in score
group 3, but only 1 percent of the distribution of attempters in
Arizona in that score group. We are worried that the larger
percentage in Texas is the result of higher ability individuals
exerting just enough effort to pass in that state, placing them in
score group 3. As a result, the observed earnings in score group 3
in Texas are a weighted average of score group 4 types and higher
ability types, and are biased upward from what they would be if
everyone scored according to their underlying ability.
To ‘‘purge’’ the estimated earnings in score group 3 of the
earnings of the higher ability types, we will assume, based on
what we observe in Arizona, that out of the 4 percent in score
group 3 in Texas, only one-fourth are true score group 3 types,
while the remaining three-fourths are actually, say, true ‘‘5s.’’ If
we assign to the three-fourths who we assumed are 5s, the
observed mean earnings for score group 5 in Texas, we can then
estimate the earnings of the one-fourth whom we believe to be
true 3s. These are our adjusted earnings estimates.
Adjusted estimates based on these assumptions and algo-
rithm are in Table VII. We assume that the treatment group in
each of the experiments is primarily contaminated by true 5s or
true 6s, and so we employ three different assumptions in Table VII
regarding the mix of 5s and 6s. First, we assume no contamina-
tion, which simply redisplays our original results from Table V for
ease of comparison. Next, we assume a 50–50 mix of 5s and 6s, and
then to get a very conservative, lower-bound estimate, we assume
only 6s contaminate the treatment group.
our adjusted estimates would be closer to our unadjusted estimates. Thus, if
anything, we believe that differential participation gives us conservative adjusted
estimates.

454
QUARTERLY J OURNAL OF ECONOMICS
TABLE VII
ADJ USTED ESTIMATES IN THE PRESENCE OF DIFFERENTIAL TESTING EFFORT ACROSS
LOW AND HIGH STANDARD STATES FOR YOUNG WHITE DROPOUTS (STANDARD
ERRORS ARE IN PARENTHESES.)
Adjusted estimates under
Adjusted numbers
of observations
different skill composition
assumptions
Treatment group
contamination
[NT
Ϫ
assumption
of . . .
Adjusted
estimate
a
b
c
d
pc
Ntreat
NT
pc(Ntreat )
pc(Ntreat )] e
Experiment 0.011 18332 653
4
202
451
no contaminationf
1473*
(678)
the contamination
is 50% true 5s
1166
(1416)
and 50% true 6s.
all contamination
is from true 6s.
612
(1526)
Experiment 0.031
3
6137 471
190
281
1531**
(529)
no contamination
contamination is
50% true 5s and
50% true 6s.
1715
(1138)
1374
(1219)
all contamination
is from true 6s.
Experiment 0.062
3*g
6137 471
380
91
907ϳ
(481)
no contamination
contamination is
50% true 5s and
50% true 6s.
936
(568)
881
(573)
all contamination
is from true 6s.
** ϭ significant at the 0.01 level; * ϭ significant at the 0.05 level.
a. c ϭ the observed proportion of individuals from high standard states in the affected score group.
b. Ntreat ϭ the total number of observations in the treatment states, across all score groups.
c. T ϭ the number of observations in the affected score group in the lower standard states.
p
N
d. The assumed number of treatment group individuals with the correct true score.
e. The assumed number of treatment group individuals with true scores that would place them in score
group 5 or 6; these are the ‘‘contaminating’’ individuals.
f. We include our original estimates in this table (the estimates we obtain under an assumption of no
contamination) for ease of comparison.
g. The mean earnings of score group 5 for the treatment states in this experiment are actually slightly
lower than the mean earnings of the treatment group here, and so a calculation using the earnings of true 5s
would result in a higher adjusted estimate.

LABOR MARKET SIGNALING VALUE OF THE GED
455
The simple story from Table VII is that while all of the
adjusted estimates are noisy, the point estimates in Table VII are
not far from our original difference-in-differences estimates dis-
played in Table V. It is only when we assume that all of the 451
‘‘contaminants’’ in the affected score group in experiment 4 are
true 6s that our estimates are driven substantially down (to $612).
Under a more reasonable assumption that the contamination is
equally spread between 5s and 6s, the original estimate falls 20
percent to $1166. Meanwhile, in experiments 3 and 3*, the
stronger contaminating assumption yields adjusted estimates
that are only 10 and 3 percent below the original estimates.
In summary, differential effort across low and high standard
states may lend an upward bias to our estimates. The evidence,
however, is that the magnitude of this bias, especially in experi-
ments 3 and 3* is not so severe as to discredit our original
estimates.
C. Endogeneity Bias Associated with State Behavior
A second area of potential bias consists of differential state
policies between low and high standard states. We examine two
different types of policies that would bias our results: state
minimum wages and state course graduation requirements.
If states with lower GED standards tended to have higher
minimum wages, then our estimates could be biased upward. The
reasoning is as follows. It could be that higher scoring GED-
holders are not generally affected by minimum wage laws (as a
result of their relatively higher skills), while lower scoring
GED-holders are. If high minimum wage states are states that
systematically have relatively low GED passing standards, then
our treatment group individuals may have higher earnings rela-
tive to the comparison group for reasons not associated with the
effect of the GED on earnings.
An examination of the data suggests that this is not a serious
concern, however, simply because there is little variation in 1995
state minimum wages. Most states in the United States in 1995
had a minimum wage of $4.25. Only nine states had higher
minimum wages, ranging up to $5.30 in Alaska. However, none of
the four states in our lowest passing standard category had a
minimum wage higher than $4.25, and states with higher mini-
mum wages are relatively equally divided between low and high
passing standard states.
One could argue that it is the ‘‘effective minimum’’ wage,

456
QUARTERLY J OURNAL OF ECONOMICS
rather than the absolute minimum, that is critical for our
analyses. Lee [1999] shows that the same absolute minimum
wage can differentially compress wages at the bottom end of the
wage distribution in low versus high wage states. If lower GED
passing standard states are also systematically lower wage states,
then this type of wage compression might cause the wages of low
scoring GED-holders (those at the margin of passing) and higher
scoring GED-holders to be closer in low standard states, biasing
our estimates upward. This type of wage effect is essentially a
state-skill interaction among GED-holders that is not included in
our model. In the subsection on ‘‘Employer Behavior ’’ which
follows, we present evidence suggesting that missing state-skill
interactions are not important factors in explaining our results.
Another potential policy difference that could affect our
estimates concerns high school graduation requirements. Higher
graduation requirements could be associated with higher dropout
rates. As a result, the underlying ability of the pool of dropouts
could be different in states with different high school graduation
requirements. In particular, if states with low GED passing
standards have systematically higher graduation requirements
for high school, then unobserved ability and motivation in the pool
of dropouts in these states may be greater than in other states. In
this case, our estimates would have an upward bias.
Our examination of course graduation requirements in 1990
shows, however, that lower GED passing standard states had
lower course graduation requirements rather than higher ones.
The mean number of Carnegie units21 required for graduation by
GED passing standard was 18.4 in the lower standard states, 19.3
in the medium standard states, and 20 in the higher standard
states. Thus, to the extent that there is a systematic relationship
between course graduation requirements and GED passing stan-
dards, it does not work in a direction that would lead to an upward
bias in our results.
D. Endogeneity Bias Associated with Employer Behavior
Finally, it could be that employers behave differently in low
versus high GED passing standard states in ways that would bias
21. A Carnegie unit is a unit for measuring the amount of high school
academic work by the number of classroom hours spent in one subject. Developed
in 1899 by a committee of the National Education Association, one Carnegie unit in
a subject equals 120 classroom hours lasting 40 to 60 minutes each and meeting
four to five times a week, 36 to 40 weeks during the school year. Most high schools
require sixteen Carnegie units per subject for graduation, including required
courses such as English, mathematics, American history, and science.

LABOR MARKET SIGNALING VALUE OF THE GED
457
our estimates. These types of biases show up as omitted interac-
tions in our model. In particular, omitted state–score or state–
GED interactions are of concern. State–score interactions that
favor high-scoring dropouts in high-standard states would bias
our estimates upward, while interactions that favor high-scoring
dropouts in low-standard states would bias our estimates down-
ward. State–GED interactions that favor GED-holders in high
standard states would bias our estimates upward, while such
interactions that favor GED-holders in low standard states would
bias our estimates downward.
It is possible that employers in states where the GED-passing
standard is relatively high give more credence to and place more
emphasis on the credential than do employers in low standard
states. In this case, ceteris paribus, GED-holders would earn more
on average in high standard states than similar GED-holders in
lower standard states, and this type of mechanism would cause an
upward bias in our estimates.
Ex ante, we cannot sign the bias associated with possible
state–score interactions. We can, however, bring ex post evidence
to bear on the question, fitting models where we let the effect of
GED test scores, as captured by a variable constructed to approxi-
mate a continuous GED score, have different slopes in low,
medium, and high standard states.22 Table VIII shows estimates
from a model fitted over cells representing score groups above the
GED passing standard in each state. This model, fit separately for
whites and minorities, allows for fixed state effects, a linear effect
of GED test score, and then different slopes for GED test score in
each of the three different groups of states defined by low,
medium, and high GED passing standards.
The nonsignificant coefficient estimates on the passing stan-
dard by score interactions suggest that state–skill interactions
are not a serious problem for our estimates. Furthermore, the
point estimates, imprecise as they are, indicate that any state–
skill interactions would tend generally to favor more highly
skilled individuals in higher standard states.
These results, combined with the most plausible state–GED
interaction effects, suggest that any bias from either or both of the
two effects would bias our estimates upward. To assess how
serious any total bias from these effects might be, we use the
22. We discuss the construction of this variable and the weighted least
squares fitting of the model in the Statistical Appendix.

458
QUARTERLY J OURNAL OF ECONOMICS
TABLE VIII
a
WEIGHTED LEAST SQUARES REGRESSIONS TESTING FOR STATE-SKILL INTERACTIONS
(DEPENDENT VARIABLE IS THE CELL MEAN OF 1995 ANNUAL EARNINGS, AND
STANDARD ERRORS ARE IN PARENTHESES.)b
Whites
Minorities
Intercept
10385**
(1765)
Ϫ4316**
(138)
27.2
5178
(3443)
Ϫ1429**
(269)
72.0
Female dummy
GED scorec
(34.0)
54.8
(40.2)
21.5
(69.0)
Ϫ14.8
(87.0)
99.1
GED score* medium states dummy
GED score* high states dummy
(45.0)
Yes
(86.0)
Yes
State dummies
F-test of H0:
GED score* medium states and
GED score* high states interactions are jointly zero
Number of cells
0.347
266
0.273
192
Number of individual level observations
28,193
7,778
** ϭ significant at the 0.01 level; * ϭ significant at the 0.05 level; ϳ ϭ significant at the 0.10 level.
a. Low, medium, and high refer to states with low, medium, and high GED passing standards. Low
standard states are the omitted group.
b. Construction of the weights and our use of the cell mean data is described in the Statistical Appendix.
c. The ‘‘GED score’’ variable is the within-state average of the GED mean score (the mean of the five tests
in the GED battery) in each GED score group. The models are fit using only GED score groups 5–10 and using
all available states.
results from experiments 3 and 4 in Table V. We expect any
upward bias due to omitted interactions to be greater in experi-
ment 3 than in experiment 4. The reason is that experiment 3 uses
treatment states with the lowest standards in the United States
in comparisons with the high standard states of New York and
Florida, while experiment 4 uses treatment states with medium
standards in comparisons with New York and Florida. Thus, the
spread between the experiment 3 and experiment 4 estimates
gives us some measure of how our results are being affected by the
total bias attributable to omitted interactions.23 That spread,
23. This result rests not only on our stated assumption regarding the
direction of the bias of the state–test-score interactions, but also on the assump-
tions that (1) any state–test-score interaction is monotonically increasing in test
scores and (2) that at least within the narrow scoring range (score groups 3 and 4)
represented by the two experiments, the treatment effect is the same. We think
both of these assumptions are reasonable, given that we are looking at a relatively
well-defined and narrowly defined sample: namely, high school dropouts who score
above a certain range on the GED battery of exams.

LABOR MARKET SIGNALING VALUE OF THE GED
459
however, is less than 4 percent ($1473 versus $1531). Therefore,
we conclude that omitted interaction bias cannot go far in
explaining our estimates.
In summary, as with almost any quasi-experimental setting,
there are scenarios under which our identifying assumptions
would not hold. We have examined the most plausible violations of
our identifying assumptions and have found our estimates for
white dropouts to be robust to various specification tests.
E. Pretreatment Contrasts
As a final specification check we examine pretreatment
difference-in-differences contrasts. In general, a necessary, but
not sufficient, condition for the validity of inferences drawn from
an experiment is that there be no pretreatment differences
between treatment and comparison outcomes. Figures I and III
show no statistically discernible pretreatment differences in the
mean earnings of the treatment and comparison groups in
experiments 4 and 3*. In particular, point estimates of pretreat-
ment earnings comparisons in experiment 3* ($110 and $58) are
almost zero. In contrast, Figure II shows that the pretreatment
comparisons in experiment 3 are negative and statistically signifi-
cant. We might therefore expect the treatment group to fare worse
than the comparison group in the absence of treatment, suggest-
ing that estimates of the treatment effect may be downwardly
biased with these groups.
VI. RESULTS FOR MINORITY DROPOUTS
Unlike our findings for whites, we find no statistically signifi-
cant GED-effect for minority group members. We discuss two
potential explanations for these differential minority results.
Using a separate data set from Florida that allows us to
determine the testing center at which individuals took the GED
exams, we find that about 17 percent of the 16–21 year-old
minority males in Florida who obtained their GED in 1995 (the
earliest year available in these data) did so while incarcerated.
The comparable figure for young white males is 4 percent, and for
white females and minority females the figures are 0.3 and 1
percent, respectively. Thus, one explanation for our minority
results may lie in the relatively high proportion of young minority
males who obtain their GED while incarcerated.

460
QUARTERLY J OURNAL OF ECONOMICS
There are two complementary reasons why individuals in our
data who obtained their GED while incarcerated might have
depressed earnings relative to individuals who obtained it in a
more conventional setting. First, the stigma of incarceration may
depress the postprison earnings of dropouts, and eliminate any
positive signaling value of the GED credential. Second, many
dropouts who obtained a GED while incarcerated may have zero
earnings five years later because they are still in prison. With the
data used in this project, we have no way of sorting out the
importance between these two complementary explanations for
the difference between the signaling value of the GED for white
and minority dropouts.
A second explanation for the minority results hinges on the
following: (a) different behavioral assumptions behind why indi-
viduals acquire the credential, (b) different white-minority distri-
butions associated with those behavioral assumptions, and (c)
employers engaging in statistical discrimination when faced with
lack of information.24 Under this explanation, one set of dropouts
acquires a GED because they either place a value on the creden-
tial or they perceive that employers place a value on the creden-
tial. Another set of dropouts acquires the credential incidentally
because it is ‘‘quasi-compulsory’’ in a program from which they are
seeking benefits. Examples of such programs are Aid to Families
with Dependent Children (AFDC), J ob Training Partnership Act
(J TPA) programs, the J ob Corps, and many community level
youth employment programs directed toward young dropouts.
The GED may be a signal of productive attributes for the
group that actively pursues the signal and a signal of ‘‘government
program participation’’ for those who acquired it in an incidental
fashion. Employers may discount the value of the ‘‘incidental’’
GED relative to the ‘‘actively pursued’’ GED. If the distributions of
‘‘actively pursued’’ and incidental GED-holders differ by race/
ethnicity, then we would expect the signaling returns to the
credential to differ across the white and minority groups.
Exploring this ‘‘government-program-GED’’ hypothesis is dif-
ficult, because the only estimates we can generate for adult
dropouts come from information on J TPA classroom programs,
while the only estimates for dropouts younger than 21 come from
the relatively small yearly cohorts of the J ob Corps. There is no
24. We are indebted to Caroline Hoxby for the ideas in this part of the
discussion.