661
[ Journal of Labor Economics, 2006, vol. 24, no. 3]
᭧ 2006 by The University of Chicago. All rights reserved.
0734-306X/2006/2403-0009$10.00
Bias-Corrected Estimates of
GED Returns
James J. Heckman, University of Chicago and American
Bar Foundation
Paul A. LaFontaine, Center for Social Program Evaluation,
American Bar Foundation
Using three sources of data, this article examines the direct economic
return to General Educational Development (GED) certification for
both native and immigrant high school dropouts. One data source—
the Current Population Survey (CPS)—is plagued by nonresponse and
allocation bias from the hot deck procedure that biases the estimated
return to the GED upward. Correcting for allocation bias and ability
bias, there is no direct economic return to GED certification. An
apparent return to GED certification with age found in the raw CPS
data is due to dropouts becoming more skilled over time. These results
apply to both native-born and immigrant populations.
I. Introduction
There has been rapid growth in the fraction of persons who achieve high
school certification by means of an equivalency exam rather than through
the traditional route of classroom attendance and high school graduation.
This project was supported by the Mellon Foundation, the Joyce Foundation,
the Pew Foundation, and NICHD R01-34598-03. We thank William Johnson and
James Ziliak for helpful comments on the first draft. We discussed our findings
with Barry Hirsch at Trinity University, San Antonio, March 2003. Additional
material for this article, referred to in the text, is available at http://jenni.uchicago
.edu/ged_imputation. Contact the corresponding author, James J. Heckman, at
jjh@uchicago.edu.
662 Heckman/LaFontaine
Fig. 1.—GED credentials issued as a percentage of public and private high school graduates,
United States, 1960–2001.
The primary vehicle for high school equivalency certification is the General
Educational Development (GED) program. In 1960, only 2% of all new
high school certificates were awarded through equivalency exams in the
United States. By 2001, over 20% of all new high school credentials were
produced through GED certification (see fig. 1). This rapid growth in exam
certification occurred despite apparently low direct economic returns to it.
Using data from the National Longitudinal Study of Youth (NLSY), Cameron
and Heckman (1993) find that, in terms of hourly wages, controlling
for differences in ability, male exam-certified high school equivalents are
statistically indistinguishable from high school dropouts who are uncertified.
Any differences in wages among exam-certified equivalents and uncertified
dropouts are completely accounted for by differences in ability.
There is no causal effect of GED certification on wages.1
Cameron and
Heckman conclude that whatever economic return there is to GED certification
must come through access to further postsecondary education and
training that certification provides. However, GEDs are much less likely
than ordinary high school graduates to complete 2- or 4-year colleges. A
large body of subsequent work, summarized in Boesel, Alsalam, and Smith
(1998), supports the position that GED recipients are more similar to dropouts
than to high school graduates along many dimensions.
Advocates of the GED testing program raised some potentially valid
criticisms of the Cameron and Heckman analysis following its publication
(Murnane, Willett, and Boudett 1999; Boudett, Murnane, and Willett 2000;
Jaeger and Clark 2006). First, Cameron and Heckman only considered
1
Later work by Cameron (1994) found similar results for NLSY females.
Bias-Corrected Estimate of GED Returns 663
labor market outcomes at ages 25 and 28. If GED certification opens up
access to occupations that are closed to high school dropouts, then the
effect of certification may not manifest itself until later in the life cycle.
A second concern is the small sample sizes available in the NLSY data.
Some argued that it would not be possible to assess the entire GED
program based on a few hundred NLSY participants. Finally, there may
be a disparate impact of the GED program across different race groups
or other subpopulations. For instance, a GED may send a different signal
for recent immigrants who acquire the credential than it does for nativeborn
dropouts. This article addresses these questions.
In 1998, the monthly portion of the Current Population Survey (CPS)
began distinguishing between the two types of high school completion.
The large sample sizes for various racial and ethnic groups, as well as the
wide range of available ages, appear to make the CPS ideal for addressing
some of the limitations of the Cameron and Heckman analysis. However,
four potentially serious problems and limitations plague the CPS data.
First, the CPS contains no measure of ability. Cameron and Heckman
found that the GED program is selective because it is the higher-ability
dropouts who attain GED certification. Once differences in ability between
GED recipients and uncertified dropouts are accounted for, wage
differentials disappear. Second, as found by Hirsch and Schumacher (2004)
in the context of estimating union-nonunion wage differentials, “match
bias” can result from the CPS’s method of imputing missing wages. We
find that the estimated returns to GED certification are substantially upward
biased because GED respondents who either refuse or fail to report
their wage information are frequently assigned (matched to) the wages of
traditional high school graduates. Third, CPS data show that a large fraction
of workers have no reported earnings because they are unemployed
or out of the labor force. Finally, bias may arise from low- and highincome
earners refusing to report earnings.
This article addresses the first three of these problems. We show that,
when estimation is performed carefully, the returns to GED certification
and other educational estimates using CPS data are similar to those obtained
from other, cleaner, data sources. We find that, after correcting for differences
in ability, GED recipients who do not continue on to college earn
the same wages as uncertified dropouts. This result applies to both males
and females across the age spectrum. We find no evidence of postcertification
life cycle wage growth attributable to the program. The apparent
return to GED certification for older age groups in the raw data is due to
a greater unobserved ability bias for older birth cohorts rather than from
a causal effect of GED certification. After correcting for problems with the
CPS data, the estimated GED-dropout difference in wages is the same in
comparable NLSY and CPS cohorts. The positive wage returns to GED
certification found in unadjusted CPS data arise from unobserved ability
664 Heckman/LaFontaine
bias and improper allocation of GED missing wages. We also show that
ability bias is greater when comparing foreign-born GED recipients and
foreign-born dropouts. After adjusting for ability, no statisticallysignificant
effect of the GED on wages is discernible for both native and foreign-born
males and females of all race and ethnic groups.
The plan of this article is as follows. In Section II, we discuss the CPS
and compare evidence from it with evidence from the NLSY. In Section
III, we present ability-bias-corrected returns to GED certification. In
Section IV, we discuss the issues of age and cohort effects, using a variety
of data sources. In Section V, we consider GED returns among immigrants.
Section VI concludes.
II. The Importance of Wage Imputation and Nonresponse
A. CPS Data
We use the monthly outgoing rotation groups from the CPS for the
period January 1998 to December 2003. Our sample consists of civilian
males and females age 20–64 who are either in their fourth or their eighth
month in the sample. We use a sample of dropouts, GED recipients, and
high school graduates who have completed no college, along with a sample
of 4-year degree holders for whom we cannot determine what type of
high school certificate they hold.2
For our wage analysis, we exclude those
people who are enrolled in school, those who are self-employed, those
who reported their ethnicity as Native American, Aleut, or Eskimo, and
those who had their education status or years of schooling responses
imputed. The self-employed are excluded because earnings are not available
for these individuals. All regressions also exclude those who earn
less than $.50 or more than $200 an hour (in 2000$). Data losses due to
these exclusions are listed in table 1. The main exclusions are those who
are not working or who are self-employed. For these groups, wage data
are unavailable. Other sample restrictions only account for a small fraction
of lost data.
B. CPS Problems and Limitations
Due to its large sample size, the long period over which it is collected,
and its perceived quality, the CPS has become the primary data source
for understanding a host of important economic issues, including the U.S.
earnings structure, racial wage gaps, and returns to education. The growing
nonresponse to income-related questions calls into question the quality
2
Due to the structure of the CPS monthly questionnaire it is not possible to
determine the GED status of those who continue on to college. For this reason,
our estimates of GED returns using the CPS are limited to the direct effect of
certification on outcomes. These estimates will be lower than an overall effect
inclusive of the indirect effects of postsecondary training.
Table 1
Exclusion Restrictions by Data Source
Native Males Native Females Foreign Males Foreign Females
CPS NLSY NALS CPS NLSY NALS CPS NALS CPS NALS
Potential observations 352,858 55,057 5,412 371,222 54,101 7,058 65,004 821 68,688 886
Not working 64,302 12,358 872 117,363 19,873 2,306 10,061 109 29,377 354
Working and enrolled 1,681 1,612 311 2,227 1,862 425 305 60 251 59
Self-employed* 40,311 3,334 0 21,064 2,107 0 5,772 0 2,921 0
Other race 3,065 0 30 2,761 0 38 124 . . . 128 . . .
Zero years of education 385 17 0 280 34 0 886 0 482 0
Imputed education 988 0 0 780 0 0 298 0 166 0
Earnings outliers 286 130 137 298 81 380 61 26 36 48
Total observations 239,400 37,961 4,106 225,517 30,621 3,952 47,295 629 35,174 429
% Not working .182 .224 .161 .316 .367 .327 .155 .133 .428 .400
% Working and excluded .170 .111 .096 .112 .105 .168 .139 .117 .105 .194
Note.—The total excluded observations is not the sum of the column since many individuals fall into multiple categories. Calculations are based on a sample of employed
dropouts and GED recipients and high school graduates with no college plus 4-year college graduates. The sample ages are 20–64 for the CPS, 20–39 for the NLSY, and 20–64
for the NALS.
* It is not possible to determine years of schooling or self-employment in the NALS data.
666 Heckman/LaFontaine
Fig. 2.—CPS monthly outgoing rotation groups, percentage of allocated earners,
1979–2003. Calculations are based on CPS Monthly Outgoing Rotation Groups from 1979
to 2003. The sample is restricted to individuals between the ages of 16 and 65 who are
members of the civilian labor force and are earnings eligible. Allocation flags are unavailable
from 1994 to August of 1995. Allocated earners from 1989 to 1993 are those people who
have missing values for unedited weekly earnings since Census-provided allocation flags are
unreliable.
of the data and its comparability across time. Figure 2 shows that prior
to 1994 the percentage of those who chose not to report earnings was
relatively stable at around 15%. After 1994, earnings nonresponse rose
from a low of 24% in 1995 to nearly 34% in 2003.3
Increasing rates of
nonresponse, greater numbers of workers selectively withdrawing from
the labor force, and the CPS practice of not collecting wage information
from the self-employed have resulted in substantial fractions of respondents
with missing wage data among certain race, sex, and age groups.
Table 2 reveals that only about 50% of white and Hispanic males in each
outgoing rotation group report earnings due to the combination of these
factors. Wage data for black males are only available for around 38% of
the sample due to higher rates of income nonresponse among the employed
and higher incidence of unemployment among this population.
The situation is worse for women due to their lower labor force participation
rates. Unlike the NLSY, which surveys each person individually,
3
The dramatic increase in allocation after 1994 is primarily due to the implementation
of the newly redesigned CPS questionnaire. The new questionnaire
asks a longer, more complex series of questions in order to determine weekly
wages, and the new data processing procedures set weekly wages to missing if
even one of these questions is met with either a refusal or a “don’t know” response.
Table 2
Sources and Extent of CPS Missing Wage Data by Race for the Full Sample
White Males Black Males Hispanic Males Foreign Males
20–29 30–39 40–49 50–59 20–29 30–39 40–49 50–59 20–29 30–39 40–49 50–59 20–59
Potential wage observations 95,928 122,760 140,283 109,744 12,956 14,819 15,032 10,654 21,697 20,714 14,283 8,049 87,019
Unemployed 4,580 3,529 3,748 2,648 1,311 846 763 392 1,286 884 561 323 3,523
Out of the labor force 11,003 6,734 10,341 16,999 2,851 1,891 2,663 2,947 2,109 1,366 1,375 1,509 8,696
Self-employed 4,132 14,577 22,713 19,102 265 749 834 719 611 1,331 1,288 714 8,493
Military 1,595 2,006 854 221 239 330 114 17 238 154 65 6 493
Nonresponse 21,753 27,912 33,282 24,692 3,384 4,433 4,790 3,087 5,034 4,781 3,367 1,875 21,714
Wage observations 52,865 68,002 69,345 46,082 4,906 6,570 5,868 3,492 12,419 12,198 7,627 3,622 44,100
% Reporting wages .551 .554 .494 .420 .379 .443 .390 .328 .572 .589 .534 .450 .507
Proxy responses 3,0731 37,080 37,766 24,655 1,780 1,794 1,918 1,547 8,046 7,133 4,445 2,110 25,158
% Self-reporting wages .231 .252 .225 .195 .241 .322 .263 .183 .202 .245 .223 .188 .218
White Females Black Females Hispanic Females Foreign Females
20–29 30–39 40–49 50–59 20–29 30–39 40–49 50–59 20–29 30–39 40–49 50–59 20–59
Potential wage observations 99,672 128,051 145,846 113,599 17,926 20,086 19,987 13,893 20,726 20,741 15,244 8,826 89,359
Unemployed 3,658 3,210 3,139 2,006 1,650 1,173 855 341 1,115 966 600 240 3,236
Out of the labor force 21,677 28,971 28,539 32,172 4,620 3,932 4,389 4,687 7,513 6,786 4,422 3,684 30,779
Self-employed 2,727 8,900 11,819 9,710 216 515 564 407 276 634 686 441 4,540
Military 188 116 86 11 59 53 31 0 23 10 9 0 49
Nonresponse 18,930 23,214 30,358 22,277 4,023 5,682 6,047 3,741 3,111 3,295 2,887 1,411 16,784
Wage observations 52,492 63,640 71,905 47,423 7,358 8,731 8,101 4,717 8,688 9,050 6,640 3,050 33,971
% Reporting wages .527 .497 .493 .417 .410 .435 .405 .340 .419 .436 .436 .346 .380
Proxy responses 23,274 19,702 21,934 14,447 2,337 2,114 2,101 1,145 3,981 3,281 2,553 1,259 14,983
% Self-reporting wages .293 .343 .343 .290 .280 .329 .300 .257 .227 .278 .268 .203 .212
Note.—Based on CPS 1998–2003 monthly outgoing rotation groups. Potential wage observations are those people in their fourth or eighth month in samples who are in
the civilian labor force. These are the individuals for whom wage and job information questions are asked.
668 Heckman/LaFontaine
the CPS survey is administered by telephone to one person who responds
for his or her entire household. Potentially exacerbating the nonresponse
problem is that the accuracy of the available wage information may also
be questionable. For males, over 60% of the wage and labor force information
is given by a proxy respondent, and these respondents may
not be privy to all income-related information. The percent of available
self-reported wages is extremely low—around 25% for males and 30%
for females. The propensity to report earnings also varies across race
groups. In particular, black males and females are 10% more likely not
to report earnings than either their white or Hispanic counterparts.
Unfortunately, the CPS does not provide enough information to determine
the nature of this response bias. We present some evidence on
the severity of this potential bias using NLSY data. Nonresponse bias
may not be large, since our estimates obtained from CPS data closely
track those estimates from cleaner data sources where we can control for
this potential bias.
C. CPS Imputation Strategy
To avoid computing national statistics based on a sample with a large
proportion of missing data and in an attempt to correct for possible nonresponse
bias caused by missing wage data, the CPS allocates missing earnings
using a “hot deck” imputation method. A hot deck assigns the wages
of respondents to nonrespondents based on a limited set of demographic,
education, and occupational characteristics.4
A common practice among
researchers is to treat allocated values as observed when using CPS survey
data. In a widely cited paper, Angrist and Krueger (1999) claim that CPS
wage allocation is empirically unimportant. This article shows that CPS
allocation methods and the resulting match bias are of first-order economic
importance in estimating returns to GED certification.
“Match bias,” a phrase due to Hirsch and Schumacher (2004), arises from
the limited number of categories used to impute nonrespondent wages. Of
particular interest to this article, the matching of wage nonrespondents to
wage respondents is based on only three levels of educational attainment:
high school dropout, high school graduate with up to but not including a
bachelor’s degree, and bachelor’s degree and above. Given these education
categories, it is clear that estimated returns for those who graduate high
school and do not attend college will exhibit an upward bias since nonrespondents
will frequently be matched to those who complete some college.
On the other hand, estimated returns for those who complete above
a bachelor’s degree will be biased downward as a result of nonrespondents
4
Currently, the CPS matches nonrespondents to respondents in the monthly
data based on the following categories: gender (2), race (2), age (6), occupation
(13), hours worked (8), education (3), and tips and overtime receipt (2).
Bias-Corrected Estimate of GED Returns 669
being assigned the wages of those with only a bachelor’s degree. Clearly,
all CPS educational estimates will be affected by this type of educational
mismatching within allocation cells. Hirsch and Schumacher (2004) and
Bollinger and Hirsch (2006, in this issue) present a more detailed discussion
of the CPS hot deck procedure and the resulting bias in estimates for various
educational categories.
GED-allocated wages exhibit a particularly severe form of this type of
misallocation bias since nonrespondents who hold GED credentials are
frequently assigned the wages of high school graduates who may have
postsecondary education up to but not including a bachelor’s degree. If
a wage differential exists between GED recipients and high school graduates,
then this differential will approach zero as the proportion of GED
nonrespondents increases. As nonresponse has grown from less than 15%
to over 30% in recent years, the upward bias in estimated returns to the
GED has increased proportionally.
Table 3 shows that, for native males, the estimated return to GED
certification is overstated by over 35% when CPS-allocated wages are
included in the sample. After dropping the allocated wages, the estimated
return to GED certification drops from .14 log points to .09. For females,
as shown in table 4, the bias tends to be generally smaller in magnitude
but is still over 25%. The estimated return decreases from .15 log points
to just under .11 for the full sample of females. As predicted, excluding
allocated earners also decreases estimated returns to high school graduation
and college completion. However, this decrease is not of the same
magnitude as is found for GED recipients. The resulting reduction for
the full sample of males is just over 5% for college graduates and just
under 12% for high school graduates who did not attend college. The
observed effects of CPS allocation for the female sample are similar. Overall,
imputation tends to increase the estimated college dropout and high
school dropout wage differentials and leaves the college–high school differential
largely unaffected. The most serious bias is observed in the GED
category.
Tables 3 and 4 show that the returns are different across racial, sex, and
ethnic groups, although not dramatically so. Returns to certification are
always higher for females compared to males, and minorities have higher
returns than whites. Both Hispanic males and females show the highest
returns to GED acquisition among all racial groups. However, the differences
across groups are not dramatic. The largest estimated difference
between pooled and separated race estimates is only .04 log points.
In order to assess how sensitive these estimates are to nonresponse and
match bias, we implement a hot deck imputation procedure that differs
from the CPS hot deck only in that it matches using more preciselydefined
educational groups. This is done both to show that it is the exclusion of
670
Table 3
CPS OLS Log Hourly Wage Regressions for Males by Race
Model 1 Model 2 Model 3
Including Allocated Earners Excluding Allocated Earners Reallocating Missing Wages
All Whites Blacks Hispanics All Whites Blacks Hispanics All Whites Blacks Hispanics
GED, no college .137 .135 .146 .163 .088 .083 .105 .117 .086 .080 .092 .109
(.005) (.006) (.016) (.016) (.006) (.007) (.020) (.018) (.007) (.008) (.021) (.019)
High school, no college .209 .209 .207 .209 .184 .180 .195 .197 .183 .181 .191 .203
(.003) (.004) (.009) (.010) (.004) (.005) (.012) (.012) (.005) (.005) (.013) (.012)
College graduate .571 .570 .584 .591 .540 .534 .590 .573 .546 .540 .585 .584
(.004) (.004) (.012) (.015) (.004) (.005) (.015) (.017) (.005) (.005) (.017) (.016)
High school–Dropout .209 .209 .207 .209 .184 .180 .195 .197 .183 .181 .191 .203
College-Dropout .571 .570 .584 .591 .540 .534 .590 .573 .546 .540 .585 .584
College–High school .362 .360 .377 .382 .357 .354 .395 .375 .362 .359 .394 .381
Adjusted R2
.287 .272 .221 .282 .321 .306 .278 .313 .314 .312 .248 .299
Observations 236,666 203,012 21,182 11,824 158,314 137,892 11,868 8,100 236,666 203,012 21,182 11,824
F-test ( ):Pr 1 F
GED p Dropout .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000
GED p High school .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000
Note.—All dummy variables are defined exclusively. Dropouts are the excluded category. Persons enrolled in school at each age are deleted as are those people who have
wages less than $.50 or more than $200 an hour (in 2000$), those who are self-employed, those who were not born in the United States, those who are younger than 20 years
of age or older than 64, those who did not complete at least 1 year of schooling, those who are Aleut and Eskimo or Native American, or those had their completed schooling
or GED status imputed by the CPS. Controls for central city status, married with spouse present, year of survey, region of residence, as well as a quadratic in age and race
dummies (where appropriate), are included in each regression but not shown. Reported standard errors (in parentheses) are corrected for heteroscedasticity and clustering with
the Huber-White sandwich estimator except when reimputing wages. Standard errors after reimputation are calculated using the method outlined in Shao and Sitter (1996).
671
Table 4
CPS OLS Log Hourly Wage Regressions for Females by Race
Model 1 Model 2 Model 3
Including Allocated Earners Excluding Allocated Earners Reallocating Missing Wages
All Whites Blacks Hispanics All Whites Blacks Hispanics All Whites Blacks Hispanics
GED, no college .150 .140 .174 .196 .110 .102 .121 .157 .108 .099 .117 .144
(.005) (.006) (.013) (.016) (.006) (.007) (.016) (.018) (.007) (.007) (.015) (.017)
High school, no college .237 .236 .217 .257 .215 .216 .191 .234 .210 .205 .199 .226
(.003) (.004) (.007) (.010) (.004) (.005) (.009) (.012) (.005) (.005) (.009) (.013)
College graduate .673 .666 .712 .708 .647 .639 .698 .700 .639 .629 .689 .683
(.004) (.004) (.009) (.014) (.004) (.005) (.011) (.016) (.004) (.005) (.010) (.017)
High school–Dropout .237 .236 .217 .257 .215 .216 .191 .234 .210 .205 .199 .226
College-Dropout .673 .666 .712 .708 .647 .639 .698 .700 .639 .629 .689 .683
College–High school .437 .430 .494 .450 .432 .423 .508 .466 .429 .424 .490 .457
Adjusted R2
.277 .263 .308 .313 .307 .291 .355 .351 .305 .287 .342 .347
Observations 223,046 185,465 26,160 10,866 154,742 130,817 15,716 7,815 223,046 185,465 26,160 10,866
F-test ( ):Pr 1 F
GED p Dropout .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000
GED p High school .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000 .000
Note.—See the note to table 3 for sample definitions and regression controls. Standard errors are in parentheses.
672 Heckman/LaFontaine
GED status as a match criterion in the CPS hot deck that causes the match
bias and to correct for possible nonresponse bias in our final estimates.
We impute wages using the CPS hot deck with an added GED educational
category. In order to account for the uncertainty associated with
the imputed wage estimates of nonrespondents, we use the bootstrapping
algorithm of Shao and Sitter (1996). This procedure produces unbiased
estimates of standard errors by reimputing missing wages for the bootstrap
replicates.
The last columns of tables 3 and 4 show that the estimates obtained
from either reallocating wages or dropping those who do not report earnings
are nearly identical. This is entirely consistent with the findings of
Hirsch and Schumacher (2004), who show similar results comparing the
wages of union and nonunion workers. This evidence does not prove the
absence of nonresponse bias, but it is suggestive that for our present
analysis the nonresponse bias is minimal.
Bollinger and Hirsch (2006, in this issue) present a detailed description
of the CPS allocation procedure and an analysis of its shortcomings. They
also present an analysis of the implications of census allocations on other
outcomes besides wages. The primary focus of our article is on estimating
the direct effects of GED certification on the wages of dropouts. CPS
imputation bias is only a part of our story, but it is the main thrust of
the Bollinger-Hirsch analysis.
We focus on estimating the return to the GED using a variety of data
sets and methods to adjust for selection effects and ability bias to show
that estimated direct returns to GED certification are very low. For the
remainder of this article, we use the most expedient method of dealing
with allocated values—and the one advocated by Bollinger and Hirsch
(2006, in this issue)—by dropping employed workers who do not report
earnings rather than imputing missing wages. Due to the richer set of
conditioning variables available in the NLSY compared to the CPS, we
are able to correct NLSY-based estimates for sample selection bias due
to employment status using both parametric and semiparametric selection
correction models; this is described in more detail in the next section.
III. Ability Bias
Even though the exclusion of allocated earners dramatically reduces the
size of the estimated return to a GED credential, the resulting wage relative
to the wage of dropouts is both positive and statistically significant for
both males and females across all race groups. Cameron and Heckman
(1993) found that positive returns to GED certification could be attributed
entirely to ability bias. Those who choose to take the GED examination
are a select group from the dropout pool. The distributions of measured
ability of the people who choose to take the GED and those who do not
Bias-Corrected Estimate of GED Returns 673
are very different. The CPS data do not include any measures of ability.
Unobserved ability may be driving the observed wage differences between
education categories. Accordingly, we turn to other strategies to control
for ability bias and to richer data sets.
A. NLSY Data
This section uses the National Longitudinal Survey of Youth (NLSY79)
to control for ability bias. The NLSY is a representative sample of young
Americans who were between the ages of 14 and 21 at the time of the
first interview in 1979. The NLSY is made up of three subsamples: (1) a
random sample of 6,111 noninstitutionalized civilian youths, (2) a supplemental
sample of 5,295 youths designed to oversample civilian Hispanics,
blacks, and economically disadvantaged whites, and (3) a sample
of 1,280 youths who were ages 17–21 as of January 1, 1979, and who
were enlisted in the military as of September 30, 1978. The NLSY collects
information on parental background, schooling decisions, labor market
experiences, and cognitive test scores. Our sample includes only the random
sample and the black and Hispanic oversamples of the 1979–2000
waves. Our wage analysis is carried out separately for males and females
and excludes those who are enrolled in school, those who have wages less
than $.50 or greater than $200 per hour, and those who are self-employed.
In 1980, the Armed Services Vocational Aptitude Battery (ASVAB) was
administered to all NLSY respondents, with a completion rate of about
94% for the sample. We use the AFQT test score as our measure of
ability.5
Figure 3 presents the distributions of AFQT scores by education
and race for the NLSY. The differences in ability between GED recipients
and dropouts for both males and females of all races are large and statistically
significant.6
In fact, GED recipients have nearly the same measured
ability as high school graduates who do not continue on to college
across all races.
B. Estimation
In order to determine the importance of ability bias in generating the
estimated returns to GED certification using CPS data, we compare CPS
estimates to those obtained in the NLSY, both including and excluding the
5
The ASVAB consists of a battery of 10 tests: general science, arithmetic reasoning,
word knowledge, paragraph comprehension, numerical operations, coding
speed, auto and shop information, mathematics knowledge, mechanical comprehension,
and electronics information. The Armed Forces Qualification Test
(AFQT) is the sum of word knowledge, arithmetic reasoning, paragraph comprehension,
and numeric operations components of the ASVAB and is a general
measure of trainability used by the military for enlistment screening and job
assignment.
6
Wilcoxon rank sum tests of stochastic dominance show strong differences.
674 Heckman/LaFontaine
Fig. 3.—Density of NLSY AFQT scores by race and gender
AFQT score. Tables 5 and 6 show that the estimated returns to certification
across race groups using NLSY data for respondents who are between 20
and 39 years of age are similar to those obtained from the CPS. The exception
is for black males. For this group, the CPS estimate is higher.
Returns to GED certification are also positive and, in all but one case, they
are statistically significant across all race groups using standard significance
levels. However, when the AFQT score is controlled for, the estimated
GED effect is essentially zero for males. The estimated effect for females
is still slightly positive across all race groups, but it is always statistically
insignificant. All wage differentials between GED recipients and dropouts
can be eliminated by accounting for ability. The positive GED effects obtained
in the CPS arise from an unobserved ability bias that results from
high-ability dropouts self-selecting into the GED program.
To test the robustness of the NLSY estimates to sample selection bias
problems that may arise from excluding workers on the basis of their
labor force status, we estimate a parametric selection correction model
due to Heckman (1979).7
As shown in the last columns of tables 5 and
6, accounting for selective participation in the work force does not overturn
the conclusion that GEDs are paid the wages of high school dropouts
at the same ability level.
One method for controlling for unobserved ability is to use fixed effects
7
At our Web site (http://www.jenni.uchicago.edu/ged_imputation), we report
estimates based on a semiparametric factor model structure (based on Carneiro,
Hansen, and Heckman [2003]). The parametric and semiparametric estimatesagree
(see tables A1–A4, which are at the Web site or available from the authors).
Table 5
NLSY OLS and Parametric Selection Corrected Hourly Wage Regressions for Males by Race
Model 1 Model 2 Model 3
No Selection on AFQT Including AFQT Controlling for AFQT and Selection*
All Whites Blacks Hispanics All Whites Blacks Hispanics All Whites Blacks Hispanics
GED, no college .065 .068 .049 .092 Ϫ.004 Ϫ.004 .003 .006 Ϫ.008 .004 .000 Ϫ.020
(.020) (.035) (.030) (.040) (.021) (.034) (.031) (.042) (.021) (.034) (.031) (.044)
High school, no college .131 .165 .080 .140 .044 .071 .026 .032 .035 .065 .029 .001
(.014) (.021) (.024) (.031) (.015) (.023) (.025) (.032) (.015) (.022) (.024) (.034)
College graduate .477 .472 .500 .523 .274 .276 .312 .253 .257 .261 .307 .207
(.018) (.024) (.037) (.055) (.022) (.031) (.043) (.057) (.022) (.031) (.043) (.057)
AFQT score . . . . . . . . . . . . .113 .109 .113 .125 .110 .104 .111 .123
. . . . . . . . . . . . (.008) (.012) (.015) (.016) (.008) (.012) (.015) (.018)
High school–Dropout .131 .165 .080 .140 .044 .071 .026 .032 .035 .065 .029 .001
College-Dropout .477 .472 .500 .523 .274 .276 .312 .253 .257 .261 .307 .207
College–High school .346 .307 .420 .383 .230 .205 .286 .221 .221 .197 .279 .206
Adjusted R2
.303 .299 .261 .212 .331 .324 .296 .250 . . . . . . . . . . . .
Observations 33,573 18,199 9,009 6,365 32,054 17,351 8,735 5,968 36,706 19,126 11,168 6,412
F-test ( ):Pr 1 F
GED p Dropout .001 .055 .107 .022 .842 .899 .925 .882 .701 .909 .993 .650
GED p High school .000 .004 .257 .219 .010 .018 .390 .516 .021 .053 .284 .605
Note.—All dummy variables are defined exclusively. Dropouts are the excluded category. Persons enrolled in school at each age are deleted as are those people who have
wages less than $.50 or more than $200 an hour (in 2000$), are younger than 20 years of age or older than 39, or are self-employed. Controls for central city status, married
with spouse present, year of survey, and region of residence, as well as a quadratic in age and race dummies (where appropriate), are included in each regression but not shown.
Reported standard errors (in parentheses) are corrected for heteroscedasticity and clustering with the Huber-White sandwich estimator.
* We use a parametric model selection correction due to Heckman (1979). For both males and females, the participation equation includes race dummies, family income in
1979, mother’s and father’s education, broken home status at 14, urban status at 14, south at 14, number of siblings, local unemployment rate, age, and age squared. For the
female model, spouse’s income, number of children in the household, and dummies for the presence of a baby or toddler in household are also included.
Table 6
NLSY OLS and Parametric Selection Corrected Hourly Wage Regressions for Females by Race
Model 1 Model 2 Model 3
No Selection on AFQT Including AFQT Controlling for AFQT and Selection
All Whites Blacks Hispanics All Whites Blacks Hispanics All Whites Blacks Hispanics
GED, no college .113 .093 .122 .111 .027 .012 .033 .027 .017 .000 .015 .032
(.021) (.029) (.039) (.041) (.021) (.030) (.035) (.043) (.021) (.031) (.034) (.045)
High school, no college .225 .199 .248 .247 .130 .123 .141 .123 .101 .096 .107 .116
(.016) (.023) (.032) (.030) (.016) (.024) (.029) (.034) (.016) (.027) (.028) (.036)
College graduate .651 .607 .667 .769 .429 .413 .415 .507 .376 .372 .345 .475
(.019) (.026) (.037) (.041) (.023) (.032) (.041) (.052) (.023) (.038) (.039) (.054)
AFQT score . . . . . . . . . . . . .131 .118 .151 .146 .126 .123 .135 .131
. . . . . . . . . . . . (.009) (.118) (.016) (.019) (.009) (.012) (.016) (.021)
High school–Dropout .225 .199 .248 .247 .130 .123 .141 .123 .101 .096 .107 .116
College-Dropout .651 .607 .667 .769 .429 .413 .415 .507 .376 .372 .345 .475
College–High school .426 .408 .419 .522 .299 .290 .274 .384 .276 .276 .238 .359
Adjusted R2
.309 .298 .312 .307 .339 .323 .349 .349 . . . . . . . . . . . .
Observations 28,489 16,225 7,341 4,923 27,567 15,645 7,195 4,727 42,707 22,186 12,923 7,598
F-test ( ):Pr 1 F
GED p Dropout .000 .002 .002 .007 .187 .689 .268 .534 .428 .999 .673 .394
GED p High school .000 .000 .000 .001 .000 .000 .001 .013 .000 .001 .002 .049
Note.—See the notes to table 5 for discussion of the selection model and exclusions. Standard errors are in parentheses.
Bias-Corrected Estimate of GED Returns 677
models. Although the CPS was not originally intended as a longitudinal
data set, many researchers construct 2-year panels from the fourth and
eighth survey months. We now exploit this longitudinal structure in an
attempt to correct for ability bias using the CPS sample.
A number of important caveats need to be given before presenting the
estimates based on a fixed effect analysis. First, the CPS survey follows
households and not individuals from one survey to the next. A person
who moves out of a household will not appear in the next survey. This
is of particular importance for our estimation because the subpopulation
we are interested in—those who attain a GED between survey rounds—
tends to be younger in age and significantly more likely to move between
survey rounds compared to older individuals. This biases longitudinal
samples toward those who are more stable, that is, toward those who do
not move between surveys. Second, changes in GED status could be due
to the mismatching of individuals or errors in reporting education from
proxy responses. While every effort is made to eliminate error due to the
first consideration by matching individuals on a number of demographic
characteristics, the second source of error is less easily dealt with. Because
the CPS surveys one member of a household, and he or she responds for
the entire household, changes in the educational status of an individual,
particularly their GED status, occur quite frequently when different members
of a household respond. This type of misreporting may be particularly
severe for GED recipients given that a proxy respondent may be
unaware that someone has a GED and because a GED is often assumed
to be the same degree as a regular high school diploma and therefore is
frequently reported as such. Finally, if a person does not report wages or
is not working in either the fourth or eighth survey months, then we
cannot use them in the estimation. Using only households with wages
reported in both interviews leads to a small sample of individuals for
whom we can estimate a fixed effect model, and the bias inherent in the
sample from this exclusion is unknown.
We present estimates from two longitudinal models using CPS data
that attempt to control for the ability bias that plagues OLS estimates.
The first model is a standard fixed effects regression that differences out
individual specific effects. The second model identifies those who obtain
a GED any time during the sample period and then enters a dummy
variable into the wage equation indicating whether an individual is in a
pre–GED attainment state or a post-GED state. Comparing the pre-GED
and post-GED coefficients helps to determine the causal effect of GED
certification on wages. No difference in pre- and post-earnings indicates
that the GED effect is zero and that cross-section estimates seriously
overstate the value of a GED. A positive difference in pre- and post-GED
earnings is evidence that supports the claim that the GED has a direct
effect on earnings. In addition, if pre-GED individuals are already earning
678 Heckman/LaFontaine
significantly more than dropouts before certification, then this is evidence
that preexisting productive factors, such as unmeasured ability, are driving
the higher wage returns of GED recipients, not any true direct effect of
the GED.
Excluding allocated earners is of particular importance when estimating
longitudinal models in the CPS. Tables 7 and 8 show that dramatically
different conclusions are reached depending on the treatment of allocated
earners. Including allocated earners results in large differences in earnings
pre- and postcertification for both males and females. After dropping
allocated earners, we find no evidence of a positive treatment effect of
the GED on earnings. GED recipients earn the same in both the preand
post-GED states, and they earn more before certification than other
dropouts. Fixed effects models strengthen the conclusion that positive
GED returns from cross-section estimates are not causal. The inclusion
of allocated earners once again generates an apparently large and statistically
significant positive effect of certification for both males and females.
Dropping allocated observations results in a zero estimated direct effect
of certification after controlling for unobserved individual effects. Estimates
from the NLSY sample confirm the conclusions drawn from the
CPS.
IV. Cohort versus Age Effects and Further Evidence on Ability Bias
Proponents of the GED program argue that a GED title may confer
little initial benefit but that, after time, GED holders will experiencehigher
wage growth than dropouts who do not certify. This claim is based on
an analogy with the returns to college. In the early years after completing
schooling, college graduate earnings do not exceed those of high school
graduates of a comparable age. In later years, their earnings far exceed
the returns to high school graduates as returns to investment are harvested.
If the GED is an investment with long-term yields, we would expect to
see higher wage differentials between GED recipients and high school
dropouts at older ages. Tables 9 and 10 shed light on this question by
estimating the return to GED certification in the CPS by age groups for
white males and females, respectively. We focus on whites because the
minority samples in NALS and the NLSY are too small.8
We consider
only GED recipients who get no further education in order to estimate
direct effects of certification on wages. For white males, we find evidence
that apparently supports the notion that the GED is an investment. GED
recipients in each successive age category have higher estimated returns
8
The estimates for minorities are consistent with those for whites, but the cells
are small and the standard errors are large. See tables A5 and A6 in the table
appendix (on the Web site at http://www.jenni.uchicago.edu/ged_imputation or
available from the authors) for these results.
Table 7
OLS Pre- versus Post-GED and Fixed Effects Estimates for CPS and NLSY White Males
CPS Pre-GED vs. Post-GED
OLS Wage Regressions*
CPS Fixed Effects Wage
Regressions† NLSY Pre-GED
vs. Post-GED
OLS Wage
Regressions‡
NLSY Fixed
Effects Wage
Regressions§
With
Allocations
Without
Allocations
With
Allocations
Without
Allocations
Pre-GED .049 .073 . . . . . . .098 . . .
(.037) (.043) . . . . . . (.037) . . .
Post-GED .124 .078 .116 .033 .082 Ϫ.036
(.009) (.011) (.057) (.056) (.036) (.050)
High school, no college .196 .167 . . . . . . .188 . . .
(.006) (.007) . . . . . . (.021) . . .
College graduate .556 .523 . . . . . . .524 . . .
(.006) (.008) . . . . . . (.025) . . .
Adjusted 2
R .318 .355 .012 .009 .320 .086
Observations 96,711 67,232 14,451 9,981 18,956 4,041
F-test: (Pr 1 F)
Pre-GED p Dropout .185 .094 . . . . . . .009 . . .
Pre-GED p Post-GED .042 .091 . . . . . . .709 . . .
Post-GED p Dropout .000 .000 .041 .550 .026 .475
* See the notes to table 3 for sample definitions and regression controls. The exception is this sample is matched white males between the ages of 18 and 49.
† High school and college graduates are omitted in fixed effects regressions as well as any time invariant controls listed under table 3.
‡ See the notes to table 5 for sample definitions and regression controls. The only exception is that this sample is between the ages of 18 and 46.
§
High school and college graduates are omitted in fixed effects regressions as well as any time invariant controls listed under table 5.
680
Table 8
OLS Pre- versus Post-GED and Fixed Effects Estimates for CPS and NLSY White Females
CPS Pre-GED vs. Post-GED
OLS Wage Regressions*
CPS Fixed Effects Wage
Regressions† NLSY Pre-GED
vs. Post-GED
OLS Wage
Regressions‡
NLSY Fixed
Effects Wage
Regressions§
With
Allocations
Without
Allocations
With
Allocations
Without
Allocations
Pre-GED .029 .065 . . . . . . .056 . . .
(.035) (.037) . . . . . . (.049) . . .
Post-GED .127 .096 .011 Ϫ.074 .086 Ϫ.050
(.009) (.011) (.058) (.056) (.032) (.054)
High school, no college .213 .203 . . . . . . .207 . . .
(.007) (.008) . . . . . . (.025) . . .
College graduate .662 .644 . . . . . . .616 . . .
(.007) (.008) . . . . . . (.028) . . .
Adjusted R2 .297 .327 .004 .013 .298 .055
Observations 87,139 62,945 9,603 6,933 16,989 2,671
F-test: (Pr 1 F)
Pre-GED p Dropout .409 .081 . . . . . . .254
Pre-GED p Post GED .005 .390 . . . . . . .520
Post-GED p Dropout .000 .000 .857 .186 .009 .352
Note.—Standard errors are in parentheses.
* See the note to table 4 for sample definitions and regression controls. The exception is this sample is matched white females between the ages of 18 and 49.
† High school and college graduates are omitted in fixed effects regressions as well as any time invariant controls listed under table 4.
‡ See the notes to table 6 for sample definitions and regression controls. The only exception is that this sample is between the ages of 18 and 46.
§
High school and college graduates are omitted in fixed effects regressions as well as any time invariant controls listed under table 6.
681
Table 9
CPS-NLSY Comparison, OLS and Selection Corrected Hourly Wage Regressions for White Males
CPS
CPS NLSY
Cohort*
34–46
NLSY
Excluding AFQT Including AFQT
NLSY AFQT and
Selection
20–29 30–39 40–49 50–59 20–29 30–39 20–29 30–39 20–29 30–39
GED, no college .031 .082 .104 .130 .076 .052 .067 -.031 Ϫ.040 Ϫ.024 Ϫ.031
(.011) (.013) (.014) (.017) (.014) (.035) (.043) (.035) (.042) (.035) (.042)
High school, no college .112 .173 .234 .220 .195 .152 .206 .057 .062 .047 .057
(.008) (.008) (.010) (.012) (.009) (.020) (.027) (.022) (.029) (.022) (.029)
College graduate .363 .544 .615 .589 .598 .387 .584 .198 .318 .175 .305
(.009) (.009) (.010) (.012) (.010) (.024) (.030) (.031) (.038) (.031) (.037)
AFQT score . . . . . . . . . . . . . . . . . . . . . .111 .153 .104 .149
. . . . . . . . . . . . . . . . . . . . . (.012) (.015) (.012) (.015)
High school–Dropout .112 .173 .234 .220 .195 .152 .206 .057 .062 .047 .057
College-Dropout .363 .544 .615 .589 .598 .387 .584 .198 .318 .175 .305
College–High school .250 .371 .381 .369 .403 .235 .377 .142 .256 .128 .248
Adjusted R2
.246 .283 .267 .228 .294 .214 .278 .244 .317 . . . . . .
Observations 29,120 40,190 38,916 24,418 34,184 10,625 8,284 10,180 7,930 11,795 8,501
F-test ( ):Pr 1 F
GED p Dropout .001 .000 .000 .000 .000 .134 .113 .367 .344 .487 .459
GED p High school .000 .000 .000 .000 .000 .003 .001 .006 .009 .026 .025
Note.—See the notes to tables 3 and 5 for sample definitions and controls. Parametric selection model estimates are shown. See the note to table 5 for details of the estimation
procedure. Standard errors are in parentheses.
* This is a cohort of persons from the CPS in the years 1998–2003 who were born in the years 1957–64, the birth years of the NLSY cohort.
Table 10
CPS-NLSY Comparison, OLS and Selection Corrected Hourly Wage Regressions for White Females
CPS
CPS NLSY
Cohort*
34–46
NLSY
Excluding AFQT Including AFQT
NLSY AFQT and
Selection
20–29 30–39 40–49 50–59 20–29 30–39 20–29 30–39 20–29 30–39
GED, no college .095 .102 .119 .105 .108 .084 .119 .011 .011 Ϫ.004 Ϫ.001
(.013) (.014) (.014) (.018) (.015) (.033) (.044) (.034) (.046) (.034) (.046)
High school, no college .164 .229 .251 .229 .243 .172 .222 .092 .118 .029 .108
(.009) (.010) (.011) (.011) (.011) (.026) (.033) (.028) (.036) (.027) (.035)
College graduate .527 .703 .683 .619 .704 .483 .732 .298 .510 .194 .501
(.010) (.010) (.011) (.012) (.012) (.029) (.035) (.036) (.044) (.036) (.044)
AFQT score . . . . . . . . . . . . . . . . . . . . . .126 .142 .120 .151
. . . . . . . . . . . . . . . . . . . . . (.013) (.019) (.013) (.018)
High school–Dropout .164 .229 .251 .229 .243 .172 .222 .092 .118 .029 .108
College-Dropout .527 .703 .683 .619 .704 .483 .732 .298 .510 .194 .501
College–High school .363 .474 .432 .391 .461 .310 .510 .207 .393 .165 .393
Adjusted R2
.323 .321 .261 .230 .290 .217 .300 .244 .327 . . . . . .
Observations 26,307 35,136 38,342 25,211 31,642 9,442 6,914 9,110 6,671 13,182 9,307
F-test ( ):Pr 1 F
GED p Dropout .000 .000 .000 .000 .000 .010 .007 .738 .811 .899 .991
GED p High school .000 .000 .000 .000 .000 .001 .006 .002 .005 .213 .004
Note.—See the notes to tables 3 and 5 for sample definitions and controls. Parametric selection model estimates are shown. See the note to table 5 for details of the
estimation procedure. Standard errors are in parentheses.
* This is a cohort of persons from the CPS in the years 1998–2003 who were born in the years 1957–64, the birth years of the NLSY cohort.
Bias-Corrected Estimate of GED Returns 683
to certification. For white females, the pattern of returns is quite different,
being nearly constant across age groups.
It is not clear whether the higher returns to GED certification for males
at older ages are due to age or cohort effects. It is not possible to answer
the age versus cohort question using cross-sectional data such as the CPS
(see Heckman and Robb 1985). It may be that the acquisition of the GED
title causes the wage differential to increase between male GED recipients
and dropouts at older ages, or it may be that older birth cohorts exhibit
higher returns due to unobservable differences in quality between GED
recipients and dropouts that are not present in more recent birth cohorts.
Comparing CPS to NLSY data and data from the National Adult Literacy
Survey (NALS) that will be discussed further in Section IV.A, we find
that higher estimated returns for older groups are due to cohort differences
and not increased wage growth resulting from GED acquisition.
By comparing GED estimates for a cohort comparable to the NLSY
cohort in the CPS to estimates reported by Cameron and Heckman (1993)
at younger ages, Jaeger and Clark (2006) claim to find evidence of strong
GED life cycle wage growth. They report that estimated returns to GED
certification in the monthly CPS data for the NLSY cohort—those born
between 1957 and 1964—far exceed the estimates reported at age 25 and
28 in Cameron and Heckman’s analysis. They conclude that by the time
GED recipients are in their late thirties to early forties, the GED title has
helped them “catch up” to high school graduates and to far exceed the
wage growth exhibited by high school dropouts who do not exam certify.
Tables 9 and 10 show that this conclusion arises as an artifact of inclusion
of allocated earners in the Jaeger and Clark samples.9
We construct
an NLSY birth cohort in the CPS. It is the sample in the CPS survey
years 1998–2003 that was born in the period 1957–1964, the same years
in which the NLSY cohort is born. In 1998, these people are ages 34–41.
In 2003 they are ages 40–46. After excluding those who do not report
their earnings, the estimated GED returns for the NLSY-comparable cohort
constructed from the CPS data are nearly identical to the estimates
obtained from the NLSY when the sample is in their twenties and again
in their thirties.
Both data sources show that GED recipient wage growth is not greater
than that exhibited by high school dropouts. Furthermore, the positive
wage differences between GED recipients and uncertified dropouts is
completely accounted for by the inclusion of the ability measure for males
and females of all ages. However, the returns to college remain. This is
9
The log hourly wage regressions in the NLSY and CPS comparisons include
similar covariates and are based on the same sample restrictions to make the
estimates comparable.
684 Heckman/LaFontaine
clear evidence of investment occurring in college. However, there is no
investment occurring in GED certification.
Tables 11 and 12 strengthen this conclusion by comparing male and
female estimates of the CPS-NLSY cohort with cross-sectional estimates
obtained from the NLSY sample at ages 25, 28, 30, 35, and 38. We again
see that the estimated returns to GED certification and high school graduation
for this cohort are remarkably similar between the two data sources
and across ages. The estimated GED-dropout difference at ages 35 and
38 is no different than those previously found by Cameron and Heckman
(1993) at ages 25 and 28. According to official published statistics from
the GED testing service, over 75% of GED recipients acquire the degree
before the age of 25. Therefore, the majority of the wage sample at ages
35 and 38 have had their diplomas for over 10 years, which is ample time
for any positive net benefits to accrue. If GED recipients have not shown
positive wage growth within 10 years of obtaining the title, it is highly
unlikely that they will do so later. Both the NLSY and CPS data strongly
reject the hypothesis of postcertification life cycle wage growth posited
by Jaeger and Clark (2006), as well as Murnane et al. (1999) and Boudett
et al. (2000), once match bias is accounted for and estimation is performed
on comparable cohorts.
Controlling for ability differences in the NLSY data produces no statistically
significant differences in wages between GED recipients and
dropouts who do not certify for both males and females at all ages. It is
possible that the differences in wages between GED recipients, high school
graduates, and dropouts observed in the CPS can be completely accounted
for by unobserved ability differences as well. Given that the NLSY cohort
shows little life cycle wage growth, it is also plausible that the higher
returns to GED certification seen for older birth cohorts in CPS data are
due to a growing difference in this ability bias between GED recipients
and dropouts. Two—not necessarily mutually exclusive—possibilitiesmay
explain the data. The first is that, as the GED program has expanded
rapidly over the last 30 years, the quality of GED recipients may have
declined. Second, the quality of dropouts may have improved. Figure 4
shows that the quality of dropouts, as measured by their years of completed
schooling, has improved across cohorts, while GED quality has
remained roughly constant. Male and female dropouts of all races have
obtained greater levels of schooling, while the completed secondary
schooling levels of GEDs are nearly constant across all birth cohorts. The
greater schooling attainment of dropouts may indicate that the skill gap
between GED recipients and dropouts is closing across cohorts, or it may
be the consequence of social promotion. Both factors may be at work.
We now turn to the National Adult Literacy Survey (NALS) data to
explore this issue further. It provides data on literacy skills of successive
cohorts.
Table 11
CPS-NLSY Comparison, OLS Log Hourly Wage Regressions for Males by Age
NLSY Excluding AFQT Score
CPS NLSY
Cohort* NLSY Including AFQT Score
25 28 30 35 38 34–46 25 28 30 35 38
GED, no college .059 .043 .015 .050 .079 .085 Ϫ.034 Ϫ.037 Ϫ.065 Ϫ.043 Ϫ.084
(.038) (.035) (.034) (.041) (.061) (.015) (.039) (.038) (.036) (.041) (.062)
High school, no college .170 .141 .161 .157 .194 .208 .059 .044 .056 .029 Ϫ.015
(.024) (.023) (.024) (.028) (.041) (.010) (.026) (.024) (.025) (.030) (.041)
College graduate .373 .443 .496 .650 .714 .611 .134 .235 .269 .382 .333
(.033) (.029) (.029) (.034) (.048) (.011) (.040) (.035) (.037) (.040) (.053)
AFQT score . . . . . . . . . . . . . . . . . . .134 .124 .133 .149 .192
. . . . . . . . . . . . . . . . . . (.013) (.013) (.013) (.014) (.019)
High school–Dropout .170 .141 .161 .157 .194 .208 .059 .044 .056 .029 Ϫ.015
College- Dropout .373 .443 .496 .650 .714 .611 .134 .235 .269 .382 .333
College–High school .203 .302 .335 .493 .520 .403 .074 .191 .214 .353 .348
Adjusted R2
.168 .229 .258 .314 .358 .308 .207 .269 .294 .352 .415
Observations 2,247 2,367 2,400 2,287 1,088 30,549 2,165 2,254 2,298 2,196 1,039
F-test ( ):Pr 1 F
GED p Dropout .119 .217 .650 .219 .193 .000 .390 .330 .067 .287 .173
GED p High school .001 .003 .000 .003 .032 .000 .006 .018 .000 .038 .189
Note.—See the notes to tables 3 and 5 for sample definitions and controls. Standard errors are in parentheses.
* This is a cohort of persons from the CPS in the years 1998–2003 who were born in the years 1957–64, the birth years of the NLSY cohort.
686
Table 12
CPS-NLSY Comparison, OLS Log Hourly Wage Regressions for Females by Age
NLSY Excluding AFQT Score
CPS NLSY
Cohort* NLSY Including AFQT Score
25 28 30 35 38 34–46 25 28 30 35 38
GED, no college .096 .117 .109 .114 .149 .107 .014 .007 Ϫ.014 .028 .022
(.047) (.048) (.048) (.046) (.062) (.015) (.048) (.049) (.049) (.047) (.065)
High school, no college .210 .234 .275 .272 .315 .237 .113 .123 .125 .161 .160
(.033) (.035) (.034) (.035) (.051) (.011) (.034) (.037) (.036) (.038) (.055)
College graduate .489 .640 .728 .799 .858 .700 .277 .417 .432 .573 .570
(.036) (.039) (.037) (.038) (.062) (.011) (.042) (.048) (.045) (.048) (.073)
AFQT score . . . . . . . . . . . . . . . . . . .137 .142 .180 .142 .146
. . . . . . . . . . . . . . . . . . (.014) (.016) (.017) (.018) (.023)
High school–Dropout .210 .234 .275 .272 .315 .237 .113 .123 .125 .161 .160
College-Dropout .489 .640 .728 .799 .858 .700 .277 .417 .432 .573 .570
College–High school .279 .405 .452 .527 .543 .463 .164 .294 .307 .412 .411
Adjusted R2
.176 .261 .318 .319 .311 .297 .213 .295 .361 .350 .342
Observations 1,855 1,832 1,873 1,857 913 29,452 1,803 1,782 1,812 1,800 885
F-test ( ):Pr 1 F
GED p Dropout .041 .014 .025 .013 .017 .000 .765 .879 .783 .554 .733
GED p High school .003 .002 .000 .000 .001 .000 .011 .003 .001 .001 .004
Note.—See the notes to tables 3 and 5 for sample definitions and controls. Standard errors are in parentheses.
* This is a cohort of persons from the CPS in the years 1998–2003 who were born in the years 1957–64, the birth years of the NLSY cohort.
687
Fig. 4.—Average years of secondary schooling for dropouts and GEDs by year of birth
688 Heckman/LaFontaine
A. NALS Data
The National Assessment of Literacy (NALS) is a decennial survey
administered by the NCES to a random sample of the U.S. adult population
to determine their literacy skills. The 1992 sample used in this
section consists of a random sample of 13,600 adults ages 16 and over
and a state supplement of 11,344 adults. The NALS testing battery consists
of three separate tests designed to measure three types of skills: prose,
document, and quantitative skills. Unlike the CPS, the NALS sample does
not ask respondents to report their hours of work. Therefore, all comparisons
between CPS and NALS data are based on weekly wage regressions.
These regressions exclude those individuals who have weekly
wages less than $100 or more than $4,000 (in 2000$), those who are
younger than 20 years of age or older than 64, and those who are Aleut,
Eskimo, or Native American. Controls for central city status, married
with spouse present, year of survey, and region of residence, as well as a
quadratic in age and race dummies (where appropriate), are included in
each regression.10
B. NALS Test Scores
As measured by the NALS test scores, people who choose to take the
GED test are as capable in their basic cognitive skills as high school
graduates and are more capable than high school dropouts who choose
not to certify. Figure 5 shows the distributions of total NALS test scores
derived from the average over all three components of the NALS battery
by race and education status for the native born. The distributions of
NALS scores for high school graduates and GED recipients are nearly
identical across all races, while dropouts have lower scores. In terms of
basic literacy skills, the GED exam effectively sorts between those who
pass the exam and those who do not.
Since the gap in years of schooling completed between dropouts and
GED recipients is narrowing across birth cohorts, we might expect to
find the cognitive skill gaps between the groups to be narrowing as well.
Figure 6, which presents NALS score distributions across different birth
cohorts, shows that this is indeed the case. The distributions of scores
for GEDs have remained nearly identical to those of high school graduates
across all birth cohorts for males and females. As dropouts have obtained
more years of schooling, their test score distributions are becoming more
similar to GEDs across birth cohorts, but they are still statistically significantly
different, even in the most recent cohort. This pattern of test
scores could produce the cross-section finding of greater return to the
10
The amount of data lost due to these exclusion restrictions for the NALS
sample is comparable to data loss generated from similar restrictions on the CPS
sample.
Bias-Corrected Estimate of GED Returns 689
Fig. 5.—Density of NALS test scores by race for the native born
GED by age solely as a consequence of diminished ability bias for more
recent cohorts. In addition, the rise in GED certification may be due in
part to diminishing participation costs of preparing for the exam by uncertified
dropouts. Whereas in 1950, when passing the GED would have
required a substantial investment and skill acquisition for a sixth-grade
droput, the average dropout from today’s public school system with 10
years of education requires only minor preparation to pass the exam.11
C. Estimation
The returns to GED certification found in the NALS92 sample for
males and females ages 20–64 closely match those found in the CPS
1998–2003. Tables 13 and 14 show that male GED recipients have 6.6%
higher weekly wages than dropouts before controlling for ability. Female
GED recipients earn 9.4% more than dropouts. However, these positive
returns to certification are completely eliminated once we control for the
NALS test score. As with the male NLSY sample, GED recipients earn
less than dropouts at the same level of ability. Once again, this effect is
not statistically significant. Female GED recipients show a small, but
statistically insignificant, positive return to certification, adjusting for ability,
much as we saw in the NLSY data. It is evident that not controlling
for ability in CPS data leads to an overestimate of the wage returns to
GED certification. All positive returns to certification can be completely
accounted for by selection into the GED program based on ability.
11
A 1980 study found that the median study time of GED examinees was only
20 hours. By 1989, they were preparing 30 hours for the test (Boesel et al. 1998).
690 Heckman/LaFontaine
Fig. 6.—Density of NALS test scores by birth cohort for the native born
The NALS distribution of test scores across birth cohorts shows that
the ability differential between GED recipients and dropouts is diminishing.
In a cross section, this results in the pattern of wage returns to
the GED across ages that is observed for males in the CPS. Older age
groups show a higher return to certification. This is a spurious age pattern
due solely to a greater ability gap between GEDs and dropouts in earlier
cohorts.
Table 15 makes this point clearly by comparing estimated weekly wage
returns in both the CPS and NALS for two birth cohorts. The first is
the pre-NLSY cohort (those born before 1957 in CPS and NALS), and
the second includes the NLSY cohort (those born 1957–64 in the CPS
and NALS) and those born afterward. Once again, we see the pattern of
higher returns for the older cohort in both the NALS and CPS data.
However, controlling for the NALS test score, across all birth cohorts,
there is no statistically distinguishable wage benefit for both male and
female GED recipients. The available evidence suggests that the GED
program has always selected the most able from the dropout pool and
that the direct wage benefits across all certification cohorts range from
small to nonexistent once we account for this selection on ability.
V. GED Returns among Immigrants
Jaeger and Clark (2006) argue that the GED has an even greater signaling
effect for immigrants than for the native born. However, their study does
not control for the ability differences between education groups. It is possible
that the GED program is even more selective in the immigrant population
than it is for natives, so that only the most able immigrants with
Table 13
NALS-CPS Comparison, OLS Log Weekly Wage Regressions for Males by Race
CPS
NALS
Excluding Test Score Including Test Score
All Whites Blacks Hispanics All Whites Blacks Hispanics All Whites Blacks Hispanics
GED, no college .085 .079 .107 .115 .066 .079 .057 .008 Ϫ.022 Ϫ.003 -.028 Ϫ.079
(.007) (.007) (.023) (.021) (.043) (.050) (.108) (.138) (.043) (.050) (.110) (.140)
High school, no college .193 .190 .200 .212 .221 .241 .163 .191 .126 .147 .092 .092
(.004) (.005) (.013) (.013) (.024) (.030) (.046) (.079) (.025) (.031) (.049) (.080)
College graduate .577 .571 .616 .619 .658 .664 .688 .639 .441 .443 .514 .407
(.005) (.005) (.016) (.019) (.026) (.031) (.063) (.091) (.032) (.038) (.076) (.111)
NALS score . . . . . . . . . . . . . . . . . . . . . . . . .148 .156 .115 .129
. . . . . . . . . . . . . . . . . . . . . . . . (.013) (.016) (.031) (.047)
High school–Dropout .193 .190 .200 .212 .221 .241 .163 .191 .126 .147 .092 .092
College–Dropout .577 .571 .616 .619 .658 .664 .688 .639 .441 .443 .514 .407
College–High school .384 .381 .415 .407 .437 .423 .525 .447 .315 .297 .422 .316
Adjusted R2
.316 .301 .261 .303 .389 .352 .354 .316 .407 .371 .371 .337
Observations 158,603 136,796 11,704 8,026 4,077 3,236 589 245 4,077 3,236 589 245
F-test ( ):Pr 1 F
GED p Dropout .000 .000 .000 .000 .122 .115 .596 .955 .603 .948 .799 .574
GED p High school .000 .000 .000 .000 .000 .000 .314 .182 .000 .001 .264 .215
Note.—All dummy variables are defined exclusively. Dropouts are the excluded category. Persons enrolled in school at each age are deleted, as are those people who have
weekly wages less than $100 or more than $4,000 (in 2000$); are not born in the United States; are younger than 20 years of age or older than 64; or are Aleut, Eskimo, or
Native American. Controls for central city status, married with spouse present, year of survey, region of residence, and a quadratic in age and race dummies, where appropriate,
are included in each regression but are not shown. Robust standard errors are shown in parentheses.
692
Table 14
NALS-CPS Comparison, OLS Log Weekly Wage Regressions for Females by Race
CPS
NALS
Excluding Test Score Including Test Score
All Whites Blacks Hispanics All Whites Blacks Hispanics All Whites Blacks Hispanics
GED, no college .127 .110 .142 .158 .094 .088 .083 .087 .023 .019 .008 .054
(.007) (.009) (.020) (.021) (.037) (.047) (.085) (.083) (.037) (.047) (.084) (.086)
High school, no college .241 .234 .235 .266 .229 .215 .252 .233 .158 .149 .179 .192
(.005) (.006) (.011) (.014) (.023) (.031) (.046) (.067) (.024) (.032) (.046) (.070)
College graduate .704 .686 .783 .766 .737 .706 .860 .731 .561 .530 .678 .637
(.005) (.006) (.013) (.019) (.026) (.033) (.056) (.099) (.032) (.039) (.065) (.119)
NALS score . . . . . . . . . . . . . . . . . . . . . . . . .145 .154 .135 .064
. . . . . . . . . . . . . . . . . . . . . . . . (.015) (.018) (.029) (.051)
High school–Dropout .241 .234 .235 .266 .229 .215 .252 .233 .158 .149 .179 .192
College-Dropout .704 .686 .783 .766 .737 .706 .860 .731 .561 .530 .678 .637
College–High school .463 .453 .548 .499 .508 .492 .607 .497 .403 .382 .499 .445
Adjusted R2
.252 .235 .336 .309 .304 .279 .371 .379 .320 .295 .387 .384
Observations 150,841 126,097 15,272 7,577 3,952 2,950 750 238 3,952 2,950 750 238
F-test ( ):Pr 1 F
GED p Dropout .000 .000 .000 .000 .011 .059 .330 .295 .528 .671 .927 .533
GED p High school .000 .000 .000 .000 .000 .002 .041 .094 .000 .002 .034 .109
Note.—See the notes to tables 3 and 13 for sample definitions and regression controls. Standard errors are in parentheses.
693
Table 15
NALS-CPS Comparison, OLS Log Weekly Wage Regressions by Cohort of Birth
NALS Males NALS Females
Males, 1940–56 Males, 1957–69 Females, 1940–56 Females, 1957–69
CPS NALS NALS CPS NALS NALS CPS NALS NALS CPS NALS NALS
GED, no college .126 .108 .003 .084 .067 .013 .136 .106 .048 .128 .073 Ϫ.009
(.013) (.069) (.071) (.011) (.062) (.062) (.013) (.057) (.057) (.012) (.061) (.062)
High school, no college .226 .254 .139 .209 .196 .136 .253 .244 .180 .259 .243 .171
(.008) (.040) (.043) (.007) (.038) (.038) (.009) (.039) (.041) (.009) (.040) (.040)
College graduate .616 .757 .515 .636 .523 .353 .691 .833 .675 .751 .662 .476
(.009) (.042) (.052) (.008) (.041) (.049) (.010) (.041) (.051) (.009) (.044) (.050)
NALS score . . . . . . .154 . . . . . . .132 . . . . . . .121 . . . . . . .165
. . . . . . (.020) . . . . . . (.023) . . . . . . (.024) . . . . . . (.025)
High school–Dropout .226 .254 .139 .209 .196 .136 .253 .244 .180 .259 .243 .171
College-Dropout .616 .757 .515 .636 .523 .353 .691 .833 .675 .751 .662 .476
College-High school .390 .503 .376 .427 .327 .217 .438 .589 .495 .491 .419 .305
Adjusted R2
.248 .360 .380 .310 .319 .336 .206 .304 .314 .248 .281 .302
Observations 51,798 1,730 1,730 61,594 1,530 1,530 53,104 1,754 1,754 55,810 1,432 1,432
F-test ( ):Pr 1 F
GED p Dropout .000 .119 .964 .000 .283 .838 .000 .063 .405 .000 .234 .882
GED p High school .000 .019 .030 .000 .022 .027 .000 .007 .009 .000 .002 .001
Note.—See the notes to tables 3 and 13 for sample definitions and regression controls. Standard errors are in parentheses.
694 Heckman/LaFontaine
Fig. 7.—Density of NALS test scores for the foreign and native born by education
higher skills GED certify. Failure to control for these factors would cause
an even wider disparity between the GED and dropout literacy and cognitive
distributions than is found in native born populations, which would
result in a higher perceived return for this subpopulation if these differences
were not accounted for in estimation. Figure 7 reveals that the distributions
of literacy levels for foreign-born dropouts, GED recipients, and high
school graduates are dramatically different. While GED recipients and high
school graduates are nearly identical in terms of literacy,immigrantdropouts
have extremely low literacy and quantitative skills. These vast differences
in basic skills among foreign-born educational groups call into question
the comparability of wage returns between them, since the types of jobs
available to them will be very different as well. This evidence suggests that
it is even more important to adjust for literacy and cognitive skill differences
among the foreign born than it is for native-born populations in order to
accurately determine the value of a GED credential for immigrants.
Immigrants who take the GED also come into the country with higher
levels of completed schooling in their home countries than immigrants who
do not take the GED. Table 16 shows that foreign-born GED recipients
and high school graduates are far more likely than foreign-born dropouts
to have attended secondary schooling in their native country. The majority
of immigrant dropouts only complete elementary school or less. Both high
school graduates and GED recipients are also more likely to have been
schooled solely in the United States, as evidenced by the percentage who
did not attend school before arriving in the United States. GED recipients
also have the highest probability of entering the country having completed
a postsecondary vocational training program. All of these factors point
Bias-Corrected Estimate of GED Returns 695
Table 16
NALS Foreign Years of Schooling Completed before Entering
the United States
Males Females
Dropouts
GED
High
School
Dropouts
GED
High
School
Did not attend school .104 .154 .145 .084 .158 .126
Primary (grades K–3) .151 .039 .039 .158 .000 .049
Elementary (grades 4–8) .494 .115 .089 .524 .263 .113
Secondary (grades 9–12) .223 .577 .648 .197 .474 .635
Vocational training .002 .077 .011 .009 .053 .014
College .007 .000 .017 .006 .000 .005
Other .000 .000 .006 .004 .000 .005
NA .019 .039 .045 .018 .053 .054
Observations 431 26 179 513 38 162
toward the possibility that the GED program is even more selective for
immigrants than it is for natives and that large wage differencesexistbetween
foreign GED recipients and foreign dropouts before they certify.
We now present CPS and NALS estimates of the returns to GED
certification among the foreign born. We estimate the same regression
model as was used to analyze the native-born population, except that we
also add controls for country of birth, citizenship status, and cohort of
entry into the United States. Table 17 shows that the CPS match bias that
results from matching foreign-born nonrespondent GED recipients and
high school graduates to native wage donors by the hot deck overstates
the value of both degrees by about .05 log points for males and .06 log
points for females. In contrast to the results for the native born, if we
drop the unallocated workers, we cannot reject the hypothesis that GED
certification is equivalent to high school graduation for both males and
females, using a 10% level of statistical significance as the criterion. The
data reject the null hypothesis that there are no direct wage benefits of
obtaining a GED compared to staying in the dropout state, so that there
appears to be a positive effect of GED certification over the dropout state.
The positive estimated returns to GED certification among the foreign
born in the CPS appear to be driven by unobserved ability bias. Figure
7 shows that, in the NALS data, GED recipients and those dropouts who
choose not to certify have very different skill distributions. Table 17 shows
that unobserved skill differences account for all differences between GED
recipients and uncertified dropouts and that the positive wage returns to
certification estimated in CPS data are spurious due to selection on ability.
Another interesting comparison that can be made in the NALS and
CPS data is one between native and foreign-born educational groups
shown in table 18. Not adjusting for ability, the ordering in the returns
to education between the groups is as expected, except for the ordering
Table 17
NALS-CPS Comparison, OLS Log Weekly Wage Regressions for the Foreign Born
CPS NALS
Including Allocated
Values
Excluding Allocated
Values Excluding Test Score Including Test Score
Males Females Males Females Males Females Males Females
GED, no college .186 .157 .134 .090 .109 .086 .012 -.045
(.016) (.018) (.019) (.020) (.113) (.112) (.110) (.111)
High school, no college .159 .189 .100 .138 .093 .095 Ϫ.024 Ϫ.049
(.006) (.007) (.007) (.008) (.057) (.067) (.058) (.070)
College graduate .603 .641 .574 .591 .614 .659 .319 .397
(.009) (.009) (.011) (.012) (.064) (.071) (.076) (.084)
NALS score . . . . . . . . . . . . . . . . . . .155 .153
. . . . . . . . . . . . . . . . . . (.024) (.029)
High school–Dropout .159 .189 .100 .138 .093 .095 Ϫ.024 Ϫ.049
College-Dropout .603 .641 .574 .591 .614 .659 .319 .397
College–High school .445 .452 .474 .453 .521 .564 .343 .446
Adjusted R2
.337 .309 .376 .325 .508 .350 .540 .391
Observations 46,912 33,996 31,498 22,747 629 429 629 429
F-test ( ):Pr 1 F
GED p Dropout .000 .000 .000 .000 .333 .446 .914 .685
GED p High school .084 .086 .063 .027 .885 .935 .751 .973
Note.—In addition to the regression controls listed in the note to table 13, all regressions include additional controls for cohort of entry, world region of birth, and whether
or not the person is a citizen of the United States. Standard errors are in parentheses.
Bias-Corrected Estimate of GED Returns 697
Table 18
NALS-CPS Comparison, Native versus Foreign Born
CPS
NALS
Excluding Test
Score
Including Test
Score
Males Females Males Females Males Females
Native, dropout .151 .103 .108 .070 Ϫ.055 Ϫ.104
(.007) (.008) (.043) (.050) (.043) (.052)
Foreign, GED .215 .192 .231 .148 .069 Ϫ.010
(.020) (.020) (.118) (.122) (.116) (.121)
Native, GED .235 .229 .171 .170 Ϫ.089 Ϫ.080
(.008) (.009) (.051) (.057) (.054) (.061)
Foreign, high school .152 .217 .209 .187 .061 .004
(.007) (.008) (.057) (.067) (.056) (.068)
Native, high school .344 .341 .323 .301 .059 .051
(.006) (.007) (.041) (.048) (.045) (.052)
Foreign, college gradute .609 .679 .723 .729 .393 .425
(.012) (.013) (.056) (.067) (.060) (.071)
Native, college graduate .730 .803 .766 .812 .367 .451
(.007) (.007) (.042) (.049) (.050) (.058)
NALS score . . . . . . . . . . . . .161 .154
. . . . . . . . . . . . (.011) (.013)
Adjusted R2
.351 .262 .406 .306 .430 .326
Observations 183,759 167,142 4,735 4,412 4,735 4,412
F-test ( ):Pr 1 F
Native dropout p
Foreign dropout .000 .000 .011 .165 .209 .045
Native GED p
Foreign GED .316 .062 .612 .859 .154 .566
Native high school p
Foreign high school .000 .000 .026 .049 .967 .410
Foreign GED p
Native dropout .000 .000 .290 .514 .205 .428
Foreign GED p
Native high school .000 .000 .428 .199 .672 .599
Native GED p
Foreign dropout .000 .000 .001 .003 .098 .187
Native GED p
Native dropout .000 .000 .115 .015 .382 .560
Note.—Regression controls are as listed in the note to table 13. Foreign dropouts are the excluded
category. Allocated earners are excluded. Standard errors are in parentheses.
for GED recipients. Despite the lower cognitive ability of foreign-born
GED recipients, as shown in figure 7, they earn the same on average as
native GED recipients for both males and females. After adjusting for
ability in the NALS data, an interesting result emerges. Both male and
female native dropouts and GED recipients earn less than their foreign
counterparts, although this difference is not always statistically significant.
This finding would not be predicted by a one-ability model of earnings.
We conjecture that the foreign born have compensating favorable noncognitive
traits such as motivation and industriousness that offset their
lower cognitive ability levels. A recent paper by Heckman, Stixrud, and
Urzua (2006, in this issue) finds that both native GED recipients and
698 Heckman/LaFontaine
dropouts have low noncognitive skills that account for their relatively
poor economic and social outcomes. Our evidence suggests that foreignborn
GED recipients may differ from native-born recipients in these
important traits. These issues are explored more fully in a forthcoming
book (see Heckman and LaFontaine 2007).
Given the small immigrant sample available in the NALS data, we must
be cautious in drawing any firm conclusions about the value of GED
certification among the foreign born. However, the evidence suggests that
those immigrants who choose to GED certify are very different from
those who do not and that any study of the value of GED certification
among this population needs to be able to account for this selection.12
VI. Conclusion
This article shows the importance of accounting for the CPS hot deck
procedure in order to obtain unbiased estimates of the return to education
using CPS data. Misallocation of nonrespondent GED recipients to high
school status results in a sizable overestimate of the value of GED certification.
This bias does not arise from nonresponse and is more sizable
among certain subpopulations such as the foreign born. Correcting for
match bias is important in order to have conceptually comparable estimates
of the returns to the GED across different data sources. Researchers
should pay closer attention to how missing wages are allocated. Alternative
allocation procedures may dramatically affect their conclusions.
The importance of this warning is highlighted by our finding of a low
direct wage return and zero life cycle wage growth for GED certification,
in contrast to the evidence presented by Jaeger and Clark (2006), who
used a biased sample.
Our evidence suggests that direct returns to GED certification are low.
Selection into the GED program on the basis of cognitive ability can
account for all wage differentials between those dropouts who do not
certify and those who choose to do so. The gap in cognitive skills appears
to be greater for older birth cohorts, and it is this greater ability bias that
produces the apparent growth in the return to the GED with age that is
found in the CPS data. No empirically significant life cycle wage growth
can be attributed to the GED title itself. Cognitive ability differences also
account for the positive effects found for GED certification among immigrants
in the CPS. This evidence highlights the importance of using
data with a rich set of family background and cognitive variables in order
to evaluate the true impact of social programs. When we control for ability
and other person-specific invariant components using longitudinal models
12
Our estimate of the effect of the GED on immigrants may be understated
because the test score we use to control for ability may be raised by preparation
time qualifying for the GED. Our NALS measure is post-GED.
Bias-Corrected Estimate of GED Returns 699
in the CPS, we find no causal effect of the GED. While CPS data provide
a foundation from which to begin an analysis of the GED program, it
cannot be considered a definitive data source. For this reason, we are
currently engaged in a more refined analysis of NLSY data and other data
sources to determine the treatment effects of GED certification among
different groups and to expand on the analysis of differences in GED
certification across cohorts reported here.
The available evidence suggests that GED certification for those who
do not obtain postsecondary schooling has little or no direct causal effect
on wages among men, women, older and more recent cohorts, and the
foreign born. All measured differences between GED recipients and dropouts
who do not certify can be accounted for by cognitive skill differences,
and these are highly correlated with schooling.
While the direct benefits of GED certification appear low, there may
still be an economic value to GED certification in opening postsecondary
schooling and training opportunities. We discuss this issue elsewhere (see
Heckman and LaFontaine 2007). As previously noted, from the CPS, we
do not know the GED status of those who go on to attend institutions
of higher learning. Thus, we cannot use these data to compute option
values from attaining the GED. From the NLSY data, we know that
about 40% of the GEDs go on to college. However, only a small percentage
finish 2-year or 4-year schools. The GED opens doors to opportunities
that are not realized. Overall, 3% of GEDs complete 4-year
college and 5% complete an associate’s degree at a 2-year college. Those
who obtain vocational skills certificates do so at the same rates as high
school dropouts.
What is true today was true 60 years ago when the GED program was
first started: there are no cheap substitutes for classroom instruction and
training.
References
Angrist, Joshua D., and Alan B. Krueger. 1999. Empirical strategies in
labor economics. In Handbook of labor economics, vol. 3A, ed. Orley
Ashenfelter and David Card, 1277–1366. New York: North-Holland.
Boesel, David, Nabeel Alsalam, and Thomas M. Smith. 1998. Educational
and labor market performance of GED recipients. Washington, DC:
Office of Educational Research and Improvement, National Library of
Education, U.S. Department of Education.
Bollinger, Christopher R., and Barry T. Hirsch. 2006. Match bias due to
earning imputation: The case of imperfect matching. Journal of Labor
Economics 24, no. 3:483–519.
Boudett, Katherine, Richard J. Murnane, and John B. Willett. 2000. “Sec-
700 Heckman/LaFontaine
ond chance” strategies for female school dropouts. Monthly Labor Review
123, no. 12:19–32.
Cameron, Stephen V. 1994. Assessing high school certification for women
who drop out. Unpublished manuscript, Department of Economics,
University of Chicago.
Cameron, Stephen V., and James J. Heckman. 1993. The nonequivalence
of high school equivalents. Journal of Labor Economics 11, no. 1, pt. 1:
1–47.
Carneiro, Pedro, Karsten Hansen, and James J. Heckman. 2003. Estimating
distributions of treatment effects with an application to the
returns to schooling and measurement of the effects of uncertainty on
college choice (2001 Lawrence R. Klein Lecture). International Economic
Review 44, no. 2:361–422.
Heckman, James J. 1979. Sample selection bias as a specification error.
Econometrica 47, no. 1:153–62.
Heckman, James J., and Paul A. LaFontaine. 2007. America’s dropout
problem: The GED and the importance of social and emotional skills.
Chicago: University of Chicago Press (forthcoming).
Heckman, James J. and Richard Robb. 1985. Using longitudinal data to
estimate age, period, and cohort effects in earnings equations. In Cohort
analysis in social research: Beyond the identification problem, ed. William
M. Mason and Stephen E. Fienberg. New York: Springer-Verlag.
Heckman, James J., Jora Stixrud, and Sergio Urzua. 2006. The effects of
cognitive and noncognitive abilities on labor market outcomes and social
behavior. Journal of Labor Economics 24, no. 3:411–82.
Hirsch, Barry T., and Edward J. Schumacher. 2004. Match bias in wage
gap estimates due to earnings. Journal of Labor Economics 22, no. 3:
689–722.
Jaeger, David A., and Melissa A. Clark. 2006. Natives, the foreign-born,
and high school equivalents: New evidence on the returns to the GED.
Journal of Population Economics (forthcoming).
Murnane, Richard J., John B. Willett, and Kathryn Parker Boudett. 1999.
Do male dropouts benefit from obtaining a GED, postsecondary education,
and training? Evaluation Review 22, no. 5:475–502.
Shao, Jun, and Randy R. Sitter. 1996. Bootstrap for imputed survey data.
Journal of the American Statistical Association 91, no. 435:1278–88.