THE RETURNS TO EDUCATION AT - Ricerca sull'intelligenza artificiale specializzata al MIT

THE RETURNS TO EDUCATION AT

COMMUNITY COLLEGES: NEW EVIDENCE

FROM THE EDUCATION LONGITUDINAL

SURVEY

Dave E. Marcotte

School of Public Aﬀairs

American University

Washington, DC 20016

marcotte@american.edu

Astratto
I use nationally representative data from the Education Longitudi-
nal Survey (ELS) to update the literature on returns to community
college education. I compare the experiences of the ELS cohort
that graduated high school in 2004 with those of the National Ed-
ucation Longitudinal Survey (NELS) cohort that graduated high
school more than a decade earlier, In 1992. I estimate that com-
munity college students from the ELS cohort were more likely to
be employed, and that those who were earned about 21 per cento
more than comparable peers with only a high school education.
This estimate is at least as large as that observed for the NELS co-
hort, though I ﬁnd some evidence that the value of an associate’s
degree is smaller for the more recent cohort. I compare these re-
sults with those from the burgeoning body of research using state
administrative data to answer similar questions.

D
o
w
N
o
UN
D
e
D

F
R
o
M
H

T
T

:
/
/

D
io
R
e
C
T
.

io
T
.

e
D
tu
e
D
P
UN
R
T
io
C
e
–
P
D

F
/

1
4
4
5
2
3
1
6
9
3
3
0
6
e
D
P
_
UN
_
0
0
2
6
7
P
D

B
sì
G
tu
e
S
T

o
N
0
8
S
e
P
e
M
B
e
R
2
0
2
3

523

https://doi.org/10.1162/edfp_a_00267

Returns to Education at Community Colleges

I N T RO D U C T I O N

1 .
Over the past several decades, the number and proportion of young Americans going
to college have steadily increased. The proportion going to two-year (community) col-
leges has grown especially fast. In 1990, 20.1 percent of recent high school graduates
enrolled in two-year colleges, E 40 percent enrolled in four-year colleges.1 By 2015,
the proportion enrolling in two-year colleges increased to 25.2 per cento, and rates of
enrollment at four-year colleges increased to 44 percent.2 As enrollment has grown,
so too has our understanding of the employment and earnings impacts of community
college education. Beginning in the 1990s, nationally representative survey data per-
mitted economists to measure the earnings eﬀects of community college, improving
on a literature that relied heavily on nonrepresentative, institution-based analyses. Over
the past decade, research on the topic has shifted toward the use of rich state adminis-
trative data combining postsecondary enrollment and Unemployment Insurance (UI)
wage records. Administrative data have many advantages, but for reasons I detail be-
low, research ﬁndings using these data have important limitations and their ﬁndings
can be diﬃcult to generalize broadly.

in questo documento, I revisit the use of nationally representative survey data to update
the literature on returns to community college education. This update is useful for
comparing the experiences of recent cohorts of high school graduates with those of
a previous generation, and for comparing results from survey data with those from
administrative records. I aim to do so to give the reader a sense of the range of estimates,
but also to illustrate the empirical challenges and limitations inherent in estimating
returns to postsecondary data, a problem that by its nature cannot rely on experimental
designs.

To update estimates from survey data, I study the experiences of students from the
Education Longitudinal Survey (ELS) cohort, collected by National Center for Educa-
tion Statistics. Students in the ELS were high school seniors in 2004, and began their
postsecondary and labor market careers at the doorstep of the Great Recession. The ex-
periences of this cohort are important in their own right, because they provide insight
into the experiences of American workers during and after one of the largest economic
downturns in modern history.

I also compare the employment and earnings eﬀects of community college en-
rollment for the ELS cohort with the cohort of students surveyed in the National
Education Longitudinal Survey (NELS). The NELS cohort began their postsecondary
education and their working careers in the early to mid-1990s—a very diﬀerent labor
market than the one young people in the ELS cohort entered. These diﬀerent settings
provide the opportunity to assess the merits of current policy proposals encouraging
sub-baccalaureate education.

Central to the task of estimating the eﬀects of postsecondary education is the omit-
ted variables problem. Infatti, estimating earnings eﬀects of education is often literally
the textbook example of this common empirical problem. Studies that use survey or ad-
ministrative data necessarily rely on quasi-experimental variation. Without random or

1. Estimates from Digest of Education Statistics, 2016, Tavolo 302.10: https://nces.ed.gov/programs/digest/d16

/tables/dt16_302.10.asp?current=yes.

2. During the end of the Great Recession, the relative increase in enrollment at two-year colleges was most pro-

nounced, peaking at 28.2 percent in 2012, whereas rates of enrollment in four-year colleges fell.

524

D
o
w
N
o
UN
D
e
D

F
R
o
M
H

T
T

:
/
/

D
io
R
e
C
T
.

io
T
.

e
D
tu
e
D
P
UN
R
T
io
C
e
–
P
D

F
/

1
4
4
5
2
3
1
6
9
3
3
0
6
e
D
P
_
UN
_
0
0
2
6
7
P
D

B
sì
G
tu
e
S
T

o
N
0
8
S
e
P
e
M
B
e
R
2
0
2
3

Dave E. Marcotte

as-good-as random variation in access to community college, researchers using survey
and administrative data adopt diﬀerent strategies to limit unmeasured heterogeneity
between those with and without postsecondary education. Naturally, survey and admin-
istrative data have diﬀerent strengths and weaknesses. Though these are well known,
they are worth restating here: Although administrative records provide data on large
samples or even the universe of relevant units, they seldom provide much informa-
tion on relevant controls or include an obvious control group. E, though survey data
typically provide richer sets of potential controls, sample size and power can be limited.
In this context, these characterizations are germane. Researchers using adminis-
trative data limit the potential impact of omitted variables by diﬀerencing or control-
ling for pre-enrollment outcomes. The survey data used in the literature are of young
persons, where no meaningful pre-enrollment outcomes are observed. This advantage
for administrative data comes at a cost: Results only generalize to the population of
students with a work history. This leaves out traditional college students. E, using
pre-enrollment earnings requires researchers to rely on a common-trends assumption
that is often violated in this context. Further, because of the reliance on records from
community colleges, there is no natural control group. Così, treatment eﬀects are es-
timated only at the intensive margin. Researchers utilizing survey data attempt to limit
the omitted variables problem by saturating regression models. Tuttavia, most survey
data oﬀer limited opportunity to employ other strategies to limit the potential impact
of heterogeneity between treatment and control groups.

in questo documento, I revisit the use of nationally representative survey data and, in do-
ing so, address and attempt to assess the importance of these limitations. This is a
useful update to the literature for at least two reasons. Primo, the ELS data provide esti-
mates of employment and earnings outcomes of community college for young people
studying and starting their careers in the 2000s and 2010s, and will serve as an im-
portant comparison to studies of previous cohorts. Understanding whether or how the
economic value of community college study has changed for young Americans is vital
for evaluating policy proposals that encourage sub-baccalaureate study as a foundation
for improving college access and reducing costs. Secondo, comparing results from sur-
vey data to contemporaneous ﬁndings from studies using administrative data can help
calibrate the ﬁndings from studies at the state level.

I estimate that community college students from the ELS cohort were more likely
to be employed, and that those who were earned about 21 percent more than their high
school–educated peers. Further, students accumulating one to two full-time equivalent
(FTE) years’ worth of credits earn more than 30 percent more than their high school–
educated peers. This is slightly larger than the earnings diﬀerence for students from
the NELS cohort, and equivalent to results from research using cohorts graduating high
school in the 1970s and 1980s (Kane and Rouse 1995). I view this as evidence that the
returns to community college education are likely increasing over time, and at the least
are holding steady.

As with research using other survey data in this literature, the capacity to develop
convincing causal estimates is limited with the ELS. As a consequence, the main results
here should not be viewed as causal estimates. Nonetheless, I show that the estimates
obtained from regression adjusted and inverse probability weighting are quite similar.
Further, by comparing unconditional earnings with those conditional on observables, IO

D
o
w
N
o
UN
D
e
D

F
R
o
M
H

T
T

:
/
/

D
io
R
e
C
T
.

io
T
.

e
D
tu
e
D
P
UN
R
T
io
C
e
–
P
D

F
/

1
4
4
5
2
3
1
6
9
3
3
0
6
e
D
P
_
UN
_
0
0
2
6
7
P
D

B
sì
G
tu
e
S
T

o
N
0
8
S
e
P
e
M
B
e
R
2
0
2
3

525

Returns to Education at Community Colleges

ﬁnd large diﬀerences between community college– and high school–educated workers
under a wide variety of assumptions about the relative degree of selection on unobserv-
ables versus observables.

2 . B AC K G RO U N D
Community colleges have played a key role in access to postsecondary education among
both recent high school graduates, and older workers attempting to upgrade their
skills.3 More than 43 percent of all students enrolled in public postsecondary educa-
tion in 2014 were at two-year institutions—up from approximately 27 percent in 1970.4
Community colleges have a mission that includes providing open-access education to
adults, as well as lifelong learning and training for nontraditional learners. But a central
mission is to provide a low-cost, open-admission opportunity for students to take col-
lege coursework, earn sub-baccalaureate degrees, and potentially transfer to four-year
colleges.

Estimating the success of community colleges is made complicated by the variety of
educational objectives of their students. Some students are full-time, degree-seeking,
and right out of high school. Others are mid-career workers seeking speciﬁc skills with
no intent of earning an associate’s degree or pursuing continuing education. A further
source of heterogeneity is that some degree-seeking students include those intending
to earn an associate’s degree as a terminal degree whereas others intend to transfer
into colleges providing bachelor’s degrees. A diﬀerent group of degree-seeking students
enroll in career technical education with the aim of earning vocational certiﬁcates and
degrees that lead to employment in ﬁelds such as information technology and health
or protective services.

The task of estimating earnings diﬀerences by education level has been central to
the study of human capital. Because it often relied on cross-sectional survey data, like
the Current Population Surveys, early work on the topic measured education using
years of completed schooling reported by respondents, and estimated the earnings ef-
fects of college by deﬁning college graduates as those reporting at least four years of
education beyond twelfth grade (Levy and Murnane 1992; Murphy and Welch 1992).

Evidence from Panel Data in the 1980s and 1990s
Improvements in this measurement strategy were made possible with the advent of sev-
eral panel datasets surveying young people in the years following high school. These in-
cluded the National Longitudinal Survey of Youth, High School and Beyond, and NELS.
Early work using panel data and detailed enrollment information found that an addi-
tional year of enrollment in postsecondary education increased earnings by about 5 A 8
percent for traditional-age college students (Kane and Rouse 1995) and for older adults
returning to school (Gill and Leigh 1997). Kane and Rouse also estimate that women
who receive an associate’s degree earned about 29 percent more than their comparable
peers with only a high school diploma. For men, they estimate the earnings premium
for a community college degree at about 8 per cento.

3. Community colleges are public two-year postsecondary institutions that award associate degrees as their high-

est degrees. This includes junior colleges, but not proprietary schools.

4. See https://nces.ed.gov/programs/digest/d15/tables/dt15_303.25.asp.

526

D
o
w
N
o
UN
D
e
D

F
R
o
M
H

T
T

:
/
/

D
io
R
e
C
T
.

io
T
.

e
D
tu
e
D
P
UN
R
T
io
C
e
–
P
D

F
/

1
4
4
5
2
3
1
6
9
3
3
0
6
e
D
P
_
UN
_
0
0
2
6
7
P
D

B
sì
G
tu
e
S
T

o
N
0
8
S
e
P
e
M
B
e
R
2
0
2
3

Dave E. Marcotte

Belﬁeld and Bailey (2011) review the literature of the eﬀects of community college
on earnings, highlighting a number of studies using the NLS and National Longitu-
dinal Survey of Youth data for this purpose. Despite diﬀerences across studies (due in
part to deﬁnitions, timing of outcome measures, speciﬁcation diﬀerences, or sample
inclusion or exclusion restrictions), these studies report earnings premia for associate’s
degrees over high school diplomas of 10 A 15 percent for men, E 20 A 25 percent for
women

Marcotte et al. (2005) and Marcotte (2010) updated this early work on the earnings
eﬀects of community colleges. These studies used data from NELS. This cohort matric-
ulated into college and started working in the 1990s, whereas previous work focused
mainly on students graduating high school in the 1970s. Despite the fact that the rel-
ative earnings of college-educated workers rose over the period, the authors’ estimates
of earnings premia for young workers with community college educations in the 1990s
were similar to those in earlier decades: They estimated that full-time enrollment in a
community college increases earnings between 5 E 8 percent for each year enrolled,
even if no degree was received—and that earning an associate’s degree increases earn-
ings by about 15 A 30 per cento.

Evidence from State Administrative Data
The recent literature on the earnings eﬀects of community college education has fo-
cused heavily on the use of state administrative data that provide the opportunity to
link students attending community college to UI wage records. Among the earliest pa-
pers of this type is the work of Jacobson, LaLonde, and Sullivan (2005), who examine
the impact of community college coursework for workers dislocated from jobs in the
early 1990s in Washington state. They estimate that an academic year’s worth of com-
munity college education increased displaced workers’ earnings by about 9 percent for
men, E 13 percent for women. They also found that returns varied substantially, con
those taking technical, vocational coursework earning substantially more than those
taking nonvocational coursework.

A more recent and burgeoning literature using administrative records has studied
the eﬀects of community college course taking on students who are not (necessarily)
displaced workers. This includes studies using data from Kentucky (Jepsen, Troske, E
Coomes 2014), North Carolina (Liu, Belﬁeld, and Trimble 2015; Xu and Trimble 2016),
Virginia (Xu and Trimble 2016), Arkansas (Belﬁeld 2015), and Washington (Dadgar and
Trimble 2015). The relative earnings diﬀerences estimated in these studies vary a good
bit. Per esempio, using data from Kentucky, Jepsen, Troske, and Coomes (2014) report
substantial earnings returns for students completing associate’s degrees from 2002 A
2004, with an increase in earnings for degree holding women of more than 50 per-
cent, and smaller increases (less than 10 per cento) for men. Xu and Trimble (2016) use
data from North Carolina and Virginia and ﬁnd signiﬁcant but smaller earnings ef-
fects for students receiving certiﬁcates at community colleges. They report adjusted
earnings diﬀerences of associate’s degree recipients to be about 30 percent higher for
women and 18 percent higher for men. In other states, researchers using administrative
data report substantively smaller relative earnings gains for associate’s degree holders
(Belﬁeld 2015; Dadgar and Trimble 2015).

D
o
w
N
o
UN
D
e
D

F
R
o
M
H

T
T

:
/
/

D
io
R
e
C
T
.

io
T
.

e
D
tu
e
D
P
UN
R
T
io
C
e
–
P
D

F
/

1
4
4
5
2
3
1
6
9
3
3
0
6
e
D
P
_
UN
_
0
0
2
6
7
P
D

B
sì
G
tu
e
S
T

o
N
0
8
S
e
P
e
M
B
e
R
2
0
2
3

527

Returns to Education at Community Colleges

3 . E M P I R I C A L M E T H O D S
Although research using state administrative data has many advantages for estimating
employment and earnings eﬀects of many aspects of education at community colleges,5
the task of estimating causal eﬀects of community college education on employment
and earnings remains empirically challenging. The principal challenge is the evalua-
tion problem inherent in establishing the counterfactual in a setting where access to
treatment cannot be feasibly assigned at random. A second (and related) problem is
identifying the treatment of interest when some students enroll with no intent to earn
a diploma. Piuttosto, they might be taking a class or two to learn a skill they perceive impor-
tant in the labor market. One approach to dealing with the latter problem is to estimate
the eﬀects of any enrollment, separate from credits earned or diplomas received.

Tuttavia, this does not resolve the central problem: The choices of whether to enroll
in community college and what to study once enrolled are surely aﬀected by factors that
cannot be controlled by the researcher but nonetheless shape anticipated outcomes. Re-
searchers using survey data have attempted to deal with this problem primarily by using
the relatively rich sets of control variables those data aﬀord. These include measures
of student and family socioeconomic and demographic attributes, as well as measures
of academic preparation and ability measured prior to postsecondary enrollment. Such
approaches provide causal estimates insofar as selection is on observables.

Researchers using administrative data have access to much more limited sets of ob-
servable attributes for students and their families. Ma, because they have access to quar-
terly wage records from state UI systems, these researchers typically can use pre-post
earnings diﬀerences to approximate causal eﬀects. Tuttavia, relying on within-student
earnings diﬀerences requires knowledge or assumptions about pre-enrollment earn-
ings trends for treatment and control groups. The well-known (Ashenfelter’s) earnings
dip prior to enrollment for adults in job training or vocational programs is relevant
here, because all administrative data studies focus on adults with a work history prior
to community college enrollment. Jacobson, LaLonde, and Sullivan (2005) illustrate
several empirical speciﬁcation issues important for estimating earnings eﬀects of com-
munity college courses for these students, including the need to measure earnings after
a suﬃcient job-search period following enrollment.

Although the use of within-student variation in studies using administrative records
potentially strengthens internal validity for estimates of employment and earnings ef-
fects, this comes at the cost of limiting external validity. These studies rely on data on
community college students who have previous work experience in UI-covered jobs.
Hence, these results cannot generalize to students with no work history or who are em-
ployed in part-time or contract work—as is the case for most teenagers or those right
out of high school. Further, studies using administrative data typically do not have a
control group of comparable adults with no postsecondary education. Piuttosto, they fo-
cus on variation in credits or degrees among persons enrolling in community colleges.
So, the treatment-control comparison is between those with degrees or many credits

5. Per esempio, a subset of the literature using administrative data examines the labor market eﬀects of career
technical education at community colleges (per esempio., Stevens, Kurlaender, and Grosz 2015). This is related substan-
tially to the questions taken up by Jacobson, LaLonde, and Sullivan (2005).

528

D
o
w
N
o
UN
D
e
D

F
R
o
M
H

T
T

:
/
/

D
io
R
e
C
T
.

io
T
.

e
D
tu
e
D
P
UN
R
T
io
C
e
–
P
D

F
/

1
4
4
5
2
3
1
6
9
3
3
0
6
e
D
P
_
UN
_
0
0
2
6
7
P
D

B
sì
G
tu
e
S
T

o
N
0
8
S
e
P
e
M
B
e
R
2
0
2
3

Dave E. Marcotte

and those dropping out. For most policy debates, the counterfactual of greatest interest
is not enrolling in postsecondary education at all.

Finalmente, studies that rely on state administrative data can only infer employment
eﬀects: If a wage record is found in a state for a person observed attending community
college, then the person is assumed to be employed, and if no wage record is found,
they are assumed to be unemployed. This is often a safe assumption, but certainly not
always. Prior research suggests interstate mobility increases with education (Wozniak
2010; Malamud and Wozniak 2012). If those with the most postsecondary education
are more likely to be lost to interstate migration, this may result in an underestimate
of the impact of community college on employment and earnings.6

Although these external validity problems are less relevant for nationally representa-
tive panel survey data, the lack of information on earnings over time is often a limitation
in those data. In the section below, I describe the survey data used in this paper, E
describe the estimation problems and models used here.

Survey Data
The ELS is a high school clustered random sample of students in the tenth grade in
2002. Student respondents were interviewed (along with school administrators, teach-
ers, and parents) in the initial year, and in 2004, 2006, E 2012. The overwhelming
majority of high school graduates in the ELS cohort received their diplomas in 2004.7 IO
restrict my analysis to the students who graduated high school on time (In 2004) and ei-
ther enrolled in community or four-year college by the time of the 2006 interview (cioè.,
within about two years of graduation), or did not enroll in postsecondary education at
all.8

I exclude those who delay enrollment in postsecondary education because there is
less information about postsecondary education for those enrolling after 2006. Dopo
2006, il prossimo (and last) follow-up is 2012. So, unlike those who start postsecondary
study by 2006, for those who start after 2006 information data on postsecondary study
would come from the same follow-up survey year (2012) as the outcome data, and there
would be insuﬃcient opportunity to observe postsecondary outcomes for this group.9
Further, this would require a much longer recall period than for those enrolling between
IL 2004 follow-up during the high school graduation year and the 2006 follow-up
two years later. Therefore, the comparison of interest is between those who enroll in
community college within two years of completing high school, and those with only a
high school degree.10 For those who enrolled in college, I distinguish between those

6. This is an issue likely to be pertinent in many of the states relevant to this literature. Per esempio, in Kentucky,
the Louisville metro area straddles the Indiana border, and the Cincinnati (Ohio) metro area is among the
largest in Kentucky. Virginia’s most populous region is Northern Virginia, a part of a metro area with Washing-
ton DC, Maryland, and West Virginia. Charlotte, the largest metro area in North Carolina, straddles the border
with South Carolina.

I exclude those who enroll in for-proﬁt colleges.

7. A total of 96.2 percent of the ELS respondents with a high school diploma graduated high school by 2004.
8.
9. Approximately 14 percent of the ELS sample who enrolled in community college as their initial postsecondary
education did so after 2006. The study here is limited to the 86 percent who enrolled within two years of high
school graduation.

10. This feature of the ELS hampers comparison with ﬁndings from research using administrative data, Quale
focuses on students enrolling in school after some period in the labor force. I discuss this external validity
constraint in the conclusions.

D
o
w
N
o
UN
D
e
D

F
R
o
M
H

T
T

:
/
/

D
io
R
e
C
T
.

io
T
.

e
D
tu
e
D
P
UN
R
T
io
C
e
–
P
D

F
/

1
4
4
5
2
3
1
6
9
3
3
0
6
e
D
P
_
UN
_
0
0
2
6
7
P
D

B
sì
G
tu
e
S
T

o
N
0
8
S
e
P
e
M
B
e
R
2
0
2
3

529

Returns to Education at Community Colleges

enrolling only in community college, or who enrolled ﬁrst in a community and later a
four-year college.

The ELS collects detailed information about students and schools, and provides
information on family and community life. This includes information about students’
prior achievement, college plans, and college enrollment decisions. As I describe below,
I attempt to limit diﬀerences between students who enroll in community college and
those who do not by controlling for student attributes, and the educational level and
income of their parents. I also control for standardized math and reading scores on
tests administered to all students while still in high school. Naturally, students with
higher academic ability are more likely to enroll in postsecondary study, so controlling
for precollege achievement levels helps isolate the impact of community college on
employment and earnings.

I measure employment and earnings outcomes from the 2012 follow-up survey.
These are self-reported measures of any employment for pay, and total annual labor
earnings in the 2011 calendar year (NCES 2014). The students in this cohort were en-
tering the labor market and/or ﬁnishing college at the start of the Great Recession.
Infatti, the labor market prospects of young workers during this period were among
the worst in a generation. At the time of the ELS follow-up, the unemployment rate for
teens exceeded 25 per cento, and was above 15 percent for those in their 20s.11

To limit academic and skills diﬀerences between community college students and
their high school–educated peers that may aﬀect subsequent labor market outcomes, IO
control for standardized scores from math and reading tests administered to the ELS
sample while in the tenth grade. The math tests included questions on algebra, geom-
etry, probability and summary statistics, and select advanced topics (NCES 2014). IL
reading assessment tested comprehension and other fundamental reading and English
language skills. The ELS tests weighted problem solving and applications more heavily
than did its predecessor, the NELS (NCES 2014). The scores are based on Item Re-
sponse Theory, which uses the pattern of responses (correct vs. incorrect) to estimate
the probability of correct answers for unanswered questions (weighted by question dif-
ﬁculty). The test scores used here are norm-referenced to the population of eligible
tenth graders.

In table 1, I present descriptive statistics for the ELS sample of 3,025 who earned a
high school diploma and ended their education there, or who enrolled in a community
college within two years of high school graduation.12 Just under two thirds of the sample
(65.9 per cento) enrolled in community college. The sample is unremarkable on basic
demographic characteristics, such as gender and race/ethnicity composition. The mean
values of the math and reading assessments are about 48, with standard deviations of
Di 9. About 56 percent of sample members’ parent attended at least some college,
con 17 percent earning a bachelor’s degree.

In addition to estimating the earnings premium associated with sub-baccalaureate
education for Millennials, another goal for this paper is to understand whether this
premium has changed over time. To assess this, I construct an identical sample from

11. NOI. Bureau of Labor Statistics (http://data.bls.gov).
12. The exclusion restrictions dropped 1,056 records of those who did not earn a high school diploma, E 5,890

who enrolled in four-year colleges after high school.

530

D
o
w
N
o
UN
D
e
D

F
R
o
M
H

T
T

:
/
/

D
io
R
e
C
T
.

io
T
.

e
D
tu
e
D
P
UN
R
T
io
C
e
–
P
D

F
/

1
4
4
5
2
3
1
6
9
3
3
0
6
e
D
P
_
UN
_
0
0
2
6
7
P
D

B
sì
G
tu
e
S
T

o
N
0
8
S
e
P
e
M
B
e
R
2
0
2
3

Dave E. Marcotte

Tavolo 1. Descriptive Statistics of the Education Longitudinal Survey Sample

Variable

Enrolled in community college

Female

White, non-Hispanic

Black, non-Hispanic

Hispanic

Asian

Math score (10th grade)

Reading score (10th grade)

Native English speaker (0/1)

Parent highest education level

Some college

College graduate

N = 3,025

Mean

0.659

0.505

0.631

0.121

0.165

0.031

47.89

47.97

0.84

0.393

0.170

Standard Dev.

0.474

0.500

0.483

0.326

0.371

0.173

8.553

8.928

0.367

0.489

0.376

NELS graduates from the high school class of 1992. I deﬁne control and independent
variables to be directly comparable to the ELS. So, for the NELS cohort I include stu-
dents matriculating at a community college between 1992 E 1994 as their ﬁrst post-
secondary enrollment. Employment outcomes were measured during the 2000 NELS
follow-up survey. In entrambi i casi, employment outcomes were measured eight years after
high school graduation, and at least six years after the onset of any postsecondary edu-
catione. Because the main and new analyses here are based on the ELS cohort, I direct
the interested reader to Marcotte (2010) for a detailed discussion of the NELS.

Empirical Methods
Regardless of whether data come from administrative records or surveys, the absence of
as-good-as random assignment to college attendance means the potential for an omit-
ted variables problem is an inherent complication for researchers in this area. To ad-
dress this problem, using the ELS survey data, I start with models of the following type:

yi = α + βCCi + γ Xi + δSi + (cid:6)io,

(1)

where yi is either a 0/1 indicator of whether individual i is employed or a measure of the
log of individual i’s annual earnings. The treatment variable(S) of interest in the initial
models is CCi, which measures whether individual i ever enrolled for credit in a com-
munity college. To limit heterogeneity between those who enroll in community college
and those who receive no education beyond high school, in all models I include mea-
sures of family and demographic attributes (Xi) and previous schooling aptitude (Si), COME
described above. (cid:6)i is a random disturbance term assumed to be normally distributed.
In this model, the coeﬃcient β measures the average eﬀect of enrollment in com-
munity college, conditional on observed family and student attributes. Speciﬁcally, Esso
is a weighted average of the relative diﬀerences in the likelihood of employment for
those with more/fewer earned credits and with/without degrees. In separate models,
I distinguish between students who complete diﬀerent levels of credits or who earn
degrees.

D
o
w
N
o
UN
D
e
D

F
R
o
M
H

T
T

:
/
/

D
io
R
e
C
T
.

io
T
.

e
D
tu
e
D
P
UN
R
T
io
C
e
–
P
D

F
/

1
4
4
5
2
3
1
6
9
3
3
0
6
e
D
P
_
UN
_
0
0
2
6
7
P
D

B
sì
G
tu
e
S
T

o
N
0
8
S
e
P
e
M
B
e
R
2
0
2
3

531

Returns to Education at Community Colleges

To attempt to further limit the potential impact of unobserved diﬀerences between
community college students and high school graduates, I ﬁrst make use of the cluster
design of the ELS and models of the following type:

yis = α + βCCi + γ Xi + δSi + μs + (cid:6)È,

(2)

where μs is a ﬁxed eﬀect for high school s. The employment and earnings outcomes
and the residual (cid:6)is now each have idiosyncratic and high school speciﬁc components.
Including a high school ﬁxed eﬀect, outcome diﬀerences are estimated by comparing
students who enrolled in community college with peers from their same high school who
did not, conditional on observed student and family attributes. Note that model 2 can
be estimated for individuals who graduated from ELS high schools where at least one
sample member enrolled in community college and at least one sample member ob-
tained no education beyond the high school diploma. In total, the ELS survey collected
data on students in 750 high schools. Six hundred twenty-eight high schools contribute
to the estimation of the ﬁxed eﬀect models of employment and 615 contribute to the
earnings models.13

The high school ﬁxed eﬀects models limit threats to validity due to the possibility
that high schools vary in their academic culture and quality, or are in diﬀerent labor
markets, both of which can aﬀect the likelihood of postsecondary study as well as em-
ployment prospects. Ovviamente, models that include high school ﬁxed eﬀects assume
that within-school factors that aﬀect postsecondary enrollment decisions are captured
by observable variables. This is a weaker assumption than the previous models that also
assume observables adequately control for local economic and social factors that shape
postsecondary enrollment decisions and employment prospects.

In addition to utilizing high school ﬁxed eﬀects to limit threats to validity, I esti-
mate outcome diﬀerences between community college–educated (treatment) and high
school–educated (controllo) members of the ELS sample using nonparametric matching
estimates. Matching estimators can be an improvement over regression analysis be-
cause they reduce model dependence (King et al. 2011). I estimate employment/
earnings diﬀerences by estimating treatment propensity as a function of observable
individual and family attributes, as well as high school attended. I then use the propen-
sity scores as inverse probability weights (IPWs), where the weight for individual i is:

wi = Ti
ˆpi

+ 1 − Ti
1 − ˆpi

(3)

where ˆpi is the propensity that individual i received treatment.14 This strategy weights
both the treatment and control groups up to the full sample, just as probability sam-
pling weights are used to generate population estimates for disproportionately sampled
subgroups in surveys (Stuart 2010). As with all matching estimators, IPW requires trim-
ming samples to enforce common support. I estimate propensity scores with replace-
ment, and drop all observations not in the area of common support. Though IPW rests
on the same assumptions about unconfoundedness as the ﬁxed eﬀects regression mod-
els above, it requires no assumptions about functional form on treatment eﬀects. IL

13. See https://nces.ed.gov/surveys/els2002/surveydesign.asp.
14. For discussion of propensity score matching, see Imbens (2004) and Stuart (2010).

532

D
o
w
N
o
UN
D
e
D

F
R
o
M
H

T
T

:
/
/

D
io
R
e
C
T
.

io
T
.

e
D
tu
e
D
P
UN
R
T
io
C
e
–
P
D

F
/

1
4
4
5
2
3
1
6
9
3
3
0
6
e
D
P
_
UN
_
0
0
2
6
7
P
D

B
sì
G
tu
e
S
T

o
N
0
8
S
e
P
e
M
B
e
R
2
0
2
3

Dave E. Marcotte

Figura 1. Distribution of Completed Postsecondary Credits of Community College Matriculants

IPW estimates will provide a point of comparison to the regression-based estimates of
average treatment eﬀects of community college education.

I also estimate models of the relationship between degree and credit completion at
community colleges and employment and earnings. To assess the relationship between
degrees and employment outcomes, I include mutually exclusive dummy variables of
the highest academic degree earned by students matriculating at a community college.
These are either an associate’s (AA) degree, or a bachelor’s (BA) degree earned after
transferring to a four-year college (even if an AA degree was earned). Of the ELS sample
who started college at a community college, 13.9 percent earned an AA as their highest
degree and 23.1 percent earned a BA. These are generally comparable to estimates from
the National Student Clearinghouse data that about 39 percent of ﬁrst-time college stu-
dents enrolling in community college in Fall 2010 earned either an AA or BA after six
years (Shapiro et al. 2016).

To study the impact of credit hours, I diﬀerentiate between community college stu-
dents earning various multiples of 15 credit hours (a full load for one semester). Ap-
proximately 20 percent of those starting at a community college had earned less than
15 credit hours after eight years (ﬁgure 1). The distribution of completed credits is bi-
modal, as the proportion earning less than two, three, and four full semesters’ worth
of credits is smaller than the proportion earning less than 15 credit hours—while the
largest group earns at least 60 credits (or two full years). Typically, 60 credits are re-
quired for an AA. Students transferring to a four-year college seeking a BA are often
required to complete 120 credits.

It is important to point out that the completion of various milestones or credits may
be associated with underlying diﬀerences in student ability or intent. Or, they may pick

D
o
w
N
o
UN
D
e
D

F
R
o
M
H

T
T

:
/
/

D
io
R
e
C
T
.

io
T
.

e
D
tu
e
D
P
UN
R
T
io
C
e
–
P
D

F
/

1
4
4
5
2
3
1
6
9
3
3
0
6
e
D
P
_
UN
_
0
0
2
6
7
P
D

B
sì
G
tu
e
S
T

o
N
0
8
S
e
P
e
M
B
e
R
2
0
2
3

533

Returns to Education at Community Colleges

up essential heterogeneity in the impact of community college on student employment
outcomes, since students learn about the value (and costs) of continued enrollment
during their studies, and this may shape both decisions about persistence as well as
subsequent outcomes.

To estimate the relationship between community college education, and employ-
ment and earnings, I estimate a series of models in which the dependent variables are
either indicators of being employed, or the log of annual labor earnings (conditional
on employment) at the time of the 2012 follow-up, when the modal age of respondents
era 26 years. I ﬁrst control only for student demographic characteristics and parents’
income and education. I then add in scores on math and reading achievement tests ad-
ministered when respondents were in tenth grade, to control for diﬀerences in ability
that might be correlated both with the likelihood of postsecondary study and labor mar-
ket outcomes. I control for potential labor market experience, measured as a quadratic
in months since last enrollment in school.

To assess whether the earnings and employment outcomes for community college–
educated workers have changed, I develop a comparison sample from the NELS, E
deﬁne outcome, treatment, and control measures identically, and estimate the same
models, described above.15 In the case of the NELS, outcomes were measured in 2000.

4 . R E S U LT S
As a ﬁrst step in understanding earnings and employment diﬀerences between com-
munity college–educated workers and their high school–educated peers, in ﬁgure 2
I present characteristics of ELS sample members who attended community college
compared with those whose education ended with a high school diploma. Each panel
presents diﬀerences between community college– and high school–educated ELS sam-
ple members. Panel (UN) shows diﬀerences between the two groups while still in high
school (In 2002). Those who would go on to community college were twice as likely to
be from high-income families (with annual incomes above $75,000) than those who would get no education beyond high school. They were also more likely to have parents who graduated from college and scored higher on standardized tests.16 All diﬀerences are statistically signiﬁcant at the 5 percent level. In panel (B), I present employment diﬀerences between the two groups in 2012, when they were typically 26 years old. Respondents with postsecondary education had better employment and earnings outcomes than their high school–educated peers. Among those with postsecondary education, 82 percent were employed at age 26 years, compared with 76 percent of high school graduates. Further, the average earnings of those with at least some college was $24,200, compared with $20,700 for their high school–educated peers. To further assess the employment outcomes of community college students com- pared with their high school counterparts, table 2 presents results of the estimation of models 1 E 2 and the IPW matching estimator. In each case, the coeﬃcient of inter- est provides an estimate of diﬀerences in the expected value of employment outcomes 15. The NELS provides an ideal comparison to the ELS for several reasons. These include readily comparable measures of family background, student ability, and employment outcomes. 16. Figura 2 presents math score diﬀerences. Reading score diﬀerences are nearly identical. In entrambi i casi, the diﬀerences are about 0.3 standard deviation. 534 l D o w n o a d e d f r o m h t t p : / / d i r e c t . m i t . F / / e d u e d p a r t i c e – p d l f / / / / 1 4 4 5 2 3 1 6 9 3 3 0 6 e d p _ a _ 0 0 2 6 7 p d f / . f b y g u e s t t o n 0 8 S e p e m b e r 2 0 2 3 Dave E. Marcotte l D o w n o a d e d f r o m h t t p : / / d i r e c t . m i t . / F / e d u e d p a r t i c e – p d l f / / / / 1 4 4 5 2 3 1 6 9 3 3 0 6 e d p _ a _ 0 0 2 6 7 p d . / f Note: All differences by education level are significant at the 5% level. Figura 2. Comparison of High School— and Community College—Educated Education Longitudinal Survey Sample: In (UN) 2002 (Tenth Grade) E (B) 2012 f b y g u e s t t o n 0 8 S e p e m b e r 2 0 2 3 535 Returns to Education at Community Colleges Table 2. Community College Enrollment and Subsequent Employment and Earnings: Education Longitudinal Survey, High School Class of 2004 Employment ln(Earnings) Independent variable (1) (2) (3) (4) (5) (6) Enrolled in community college 0.052 (0.032) 0.026 (0.036) 0.094 (0.054) 0.284** (0.091) 0.199* (0.097) 0.197** (0.067) High school fixed effects Estimator N Number of high schools R2 No OLS 3,025 – 0.065 Yes OLS 3,025 628 0.301 Yes IPW 2,272 225 No OLS 2,669 – 0.069 Yes OLS 2,669 615 0.278 Yes IPW 1,903 243 Notes: All models control for respondent race, genere, potential labor market experience (quadratic), parental education, family income when in 10th grade, and performance on standardized reading and math assess- ments in high school. See text for details. Standard errors in parentheses. OLS = ordinary least squares; IPW = inverse probability weight. *P < 0.05; **p < 0.01. between those with any enrollment in community college and those with no educa- tion beyond high school, conditional on observed individual and family characteristics. The results in model 2 and the IPW estimator also condition/match on high school attended. The left side of the table presents models where the dependent variable is employ- ment at the time of the last follow-up survey. The ﬁrst two columns are parametric estimates from linear probability models.17 The third column shows the IPW matching estimate. Regardless of the model or estimate, I ﬁnd no signiﬁcant diﬀerence in em- ployment likelihood between community college– and high school–educated workers. The right side of table 2 presents results from the models where the dependent vari- able is earnings conditional on employment. In column 4, I present the results from model 1. The results suggest that on average, by their late 20s, Millennial workers who enrolled in community college earned approximately 32.8 percent more annually than their high school–educated peers (p < 0.01), conditioning on observed demographic, family, and academic background.18 In column 5, I present results from model 2, which includes high school ﬁxed eﬀects. The conditional earnings diﬀerence between com- munity college– and high school–educated young workers falls to 22 percent. Notably, the earnings diﬀerence between observationally identical high school– and community college–educated workers falls by about a third when we compare students who attend the same high school. This suggests that some of the diﬀerences observed in column 4 are due to diﬀerences in earnings that would have been expected anyway, because college students on average attended better high schools or lived in areas with better labor markets. Because of the importance of high school ﬁxed eﬀects, the IPW estimator matches on observable characteristics for between workers with and without postsecondary ed- ucation who attend the same high school. Of course, this is the same variation that 17. Marginal eﬀects at the means from logistic regression estimates are substantively similar to the linear proba- bility model coeﬃcients. I exponentiate coeﬃcients from log-linear models for more precise estimated percent changes. 18. 536 l D o w n o a d e d f r o m h t t p : / / d i r e c t . m i t . / / f e d u e d p a r t i c e - p d l f / / / / 1 4 4 5 2 3 1 6 9 3 3 0 6 e d p _ a _ 0 0 2 6 7 p d f / . f b y g u e s t t o n 0 8 S e p e m b e r 2 0 2 3 Dave E. Marcotte Table 3. Community College Enrollment and Subsequent Employment Outcomes Independent Variable Intensity of enrollment Any enrollment <15 Credits 15 to 30 45 60 More than Highest earned degree Associate’s Bachelor’s (0>
THE RETURNS TO EDUCATION AT image

Scarica il pdf