|
|
|
IX. HISPANIC EMPLOYMENT AND EARNINGS
Out of the Barrio: Do Young Hispanics Benefit from Residential
Job Training Programs?
Alfonso Flores-Lagunes, Arturo Gonzalez,
and Todd Neumann
University of Arizona
Abstract
A four-year longitudinal study, the National Job Corps Study
(NJCS) was a randomized experiment in which over fifteen thousand Job
Corps eligible applicants were randomized into treatment and control groups.
Job Corps was found to have positive impacts in the weekly earnings of
whites and others forty-eight months after randomization, but not for
Hispanics. We argue that one reason for this finding is that the NJCS
did not create comparable treatment and control groups for Hispanics.
Given the failure of randomization for Hispanics, we employ non-experimental
estimators to examine the programmatic outcome of Hispanics. Our findings
suggest that the lack of programmatic gain is due to the large and unusual
earnings by Hispanic controls.
Introduction
During the late 1990s, the Department of Labor sponsored the
National Job Corps Study (NJCS) to assess the effectiveness and social
value of Job Corps. Eligible Job Corps applicants were assigned to a treatment
or control group, where the former could enroll in Job Corps, and the
latter were denied enrollment for thee years. Overall, the NJCS report
found that program participants earned 12 percent more than control-group
members during the forty-eight-month follow-up survey (Burghardt et al.
2001). At the same time, however, the NJCS revealed that Hispanics that
undertook Job Corps training did not have higher earnings than the Hispanic
control group, a finding that could not be explained by differences in
characteristics, education and training, length of program enrollment,
or quality of Job Corps centers.
The objective of this paper is to provide possible explanations to these
findings. In particular, we argue that the NJCS randomization, which was
applied to the whole sample, did not create comparable treatment and control
groups for Hispanics, and thus the Hispanic control group is not a valid
counterfactual to compare the outcomes of Hispanics that completed Job
Corps training. In such a case, non-experimental methods are appropriate
for evaluating the programmatic outcome for Hispanics. When we apply different
non-experimental estimators to the NJCS data we find smaller estimated
negative effects than the original experimentally-estimated impacts for
Hispanics, although our estimated effects are still negative and statistically
insignificant. We report evidence that the Hispanic control group is not
a valid counterfactual, but we are not able to distinguish the reason
why no programmatic effects are found for Hispanics with the current available
data.
Job Corps Program and the National Job Corps Study
The purpose of Job Corps is to provide low-skilled and less-educated
young people with marketable skills to enhance their labor market outcomes
by offering academic, vocational, and social skills training in over 115
residential and training centers. Nearly seventy thousand new students
participate every year at a cost of about $1 billion. Students are selected
based on several criteria, including age (between sixteen and twenty-four
years old), poverty status, residence in a disruptive environment such
as neighborhoods with low socioeconomic characteristics, not on parole,
being a high school dropout or in need of additional training or education,
and citizen or permanent resident (U.S. Department of Labor 1999; Schochet
et al. 2001).
From a national pool of over 80,000 Job Corps-eligible young
persons, 15,386 were selected for the National Job Corps Study (NJCS)
in the mid-1990s. The experimental study assigned 61 percent of the selected
Job Corps eligible young persons to the treatment group and 39 percent
were assigned to the control group. The control group was not permitted
to enroll in Job Corps for three years after randomization, yet they were
not prevented from enrolling in other training programs. In order to assess
the effects of Job Corps, both the control and treatment groups were tracked
with a series of interviews immediately after randomization, with more
interviews twelve, thirty, and forty-eight months after randomization.
Some Features of the NJCS Data
One of the main reasons why social experiments are employed
is the notion that, because of randomization, the treatment and control
groups are statistically identical, and this allows direct comparisons
between both groups. Of the initial intake sample of 15,386, 11,313 members
completed the final forty-eight-month follow-up survey. We include those
persons with a complete baseline survey and who provided income information
during the forty-eight-month follow-up survey in our working sample. A
table with the main variables of interest in this study by ethnic origin
and randomization outcome, and the z-statistic of the test of equality
of means within ethnic group is available from us upon request. Here,
we summarize some of the salient features.
Given the opportunity to enroll in Job Corps, 73 percent of those in the
treatment group did so, while a small percentage of the control group,
4.4 percent, also undertook Job Corps services. Tests of difference in
means show that for the whole sample, randomization was successful in
producing a comparable control group. However, the Hispanic treatment
and control groups show significant differences at baseline in the mean
number of children, mean age of oldest child, and the proportion of them
living in a PMSA (this is the only demographic group with a statistically
significant difference in this variable). Also, the Hispanic controls
have a unique education and training outcome that distinguishes them from
other groups. First, Hispanic controls with a high school diploma account
for 26.6 percent of all persons who attained a high school diploma or
GED, the highest percentage among all groups. Second, while 72.5 percent
of Hispanic control group members took some form of training or education
program, a greater percentage of Hispanic control group members (18 percent)
completed a vocational program than of whites (13.4 percent) and others
(15.2 percent).
To consider further differences between the control and treatment group
members, Figures 1 and 2 show the growth in earnings over the sixteen-quarter
period for each of the demographic groups split by individuals receiving
and not receiving any training (including Job Corps), and by whether they
were assigned to the control or treatment group. Within individuals receiving
training (Figures 1A and 2A), whites show higher earnings in all quarters,
followed by Hispanics and others. For some reason, earnings for treated
Hispanics in the treatment group stagnate, while the earnings of those
treated in the control group keep growing over those final quarters. Undoubtedly,
this difference contributes to the negative estimate of the effect of
the program.
In Figure 2B, Hispanics in the control
group not receiving training show surprising growth in earnings in the
first twelve quarters that allows them to overtake the level of earnings
of whites for a few quarters (Hispanics do not earn more than whites in
any other subgroup in any other quarter). While the earnings growth for
this Hispanic subgroup also stagnates, the previous high growth allows
them to finish with higher earnings than whites in this subgroup, and
comparable to the earnings of whites who receive training. It is interesting
to note that Hispanics show the highest earnings growth within those receiving
training in the treatment group (13.82 percent) and within those receiving
training in the control group (9.24 percent). Additionally, Hispanics
not receiving training in the control group have high earnings growth
(8.14 percent), which is significantly higher than the growth of any other
group not receiving training, either in the control or treatment group,
by over 1.5 percentage points. There is, therefore, some evidence that
the Hispanic control group that does not receive training is somewhat
different from whites and others in the same category.
Insights from Non-experimental Estimators
Since individuals in the control group have access to alternative
programs, and a large number of individuals indeed enrolled in them, we
can define the following two parameters that can be estimated under some
assumptions. The first one is the (average) treatment effect of Job Corps
relative to other training programs available to eligible applicants,
known as the effect of the program (Heckman et al. 2000). The second parameter
is the effect of training relative to no training at all, called the effect
of training (Heckman et al. 2000).
The assumption that the control and treatment groups are comparable is
not valid under certain situations. As the data shows, just over 27 percent
of those randomized into training never actually enroll in Job Corps,
while slightly less than 72 percent of those in the control group enroll
in substitute training. It is possible that those that enroll are somehow
different than non-enrollees. If we are interested in evaluating the effect
of training and relax the assumption that the training inside and outside
Job Corps provides the same mean benefits, then randomization no longer
yields comparable treatment and control groups, and non-experimental methods
are necessary to obtain estimates of the benefits of any training. Additionally,
as mentioned above, given that randomization was used for the whole sample
and not applied specifically to the different demographic subgroups, then
randomization does not guarantee that the treatment and control groups
by race are comparable. This might be particularly true for Hispanics,
as they represent the smallest group (compared to whites and others),
and since they are more likely to be geographically concentrated: 44 percent
of Hispanics, in contrast to 35.8 percent of others and 15.6 percent of
whites, live in a PMSA.
As a reference, using an experimental estimator similar to the one employed
in the original NJCS study (which adjusts for enrollment in Job Corps
by treatment and control group members) shows that the impact on wages
in quarter sixteen on those that completed Job Corps are as follows: treated
Hispanics earned a statistically insignificant $15 less per week relative
to their control group, while whites and others earned a statistically
significant $46 and $22 more, respectively (all income is in 1995 dollars).
These estimates are similar to the ones reported in the NJCS study.
We report in Table 1 the estimated effects of Job Corps and any training
on weekly earnings in quarter sixteen using the following non-experimental
estimators: differences-in-differences (DID), sample selection
(SS), and matching estimator (ME). Given the DID
estimates, we infer that the effect of time-invariant differences in covariates
and unobserved traits fail to explain the observed lack of an effect of
Job Corps on Hispanics, while the DID estimates of the effect of
training Hispanics is positive but statistically insignificant.
We employ the widely used sample selection
(SS) model by Heckman (1979), using randomization into the experimental
treatment group as an exclusion restriction. A salient feature of the SS
results is the evidence of selection into training based on unobservable
characteristics: the selection variable ( ) is statistically
significant in most cases. There is evidence of negative selection for whites
and others, suggesting that the unobserved factors that influence the probability
of receiving training are negatively correlated with the unobserved factors
influencing earnings or the probability of working. Interestingly, the results
for Hispanics suggest a positive relationship between the unobservable characteristics
in the selection into training equation and those in the outcome equation.
These findings suggest that what motivated Hispanics to enroll in Job Corps
and any training was different from what motivated other ethnic groups.
The ME we employ is the bias-corrected simple matching estimator
of Abadie and Imbens (2002). This matching estimator is easy to implement,
and has desirable large-sample properties and good finite sample properties
compared to other matching estimators available.1 The ME interestingly
shows for Hispanics that the estimated effect of Job Corps is positive
while the effect of any training is negative, although both are statistically
insignificant.
Further Analysis of the Hispanic Subsample
Hispanics Controls versus non-Hispanic Trained Groups
To address our suspicion that the Hispanic control group might
be different from the other demographic control groups, we undertake the
experiment of matching individuals in the Hispanic treatment group with
individuals in the control group of a different race. Conversely, we match
non-Hispanic individuals who received training with Hispanic individuals
who do not. The results of this experiment are in the first four columns
of Table 2.
Taken together, these results seem to indicate that Hispanics receiving
training (Job Corps or any type) compare favorably with non-Hispanics
not receiving it, whereas when non-Hispanics receiving training are compared
to Hispanics not receiving it, the previously estimated positive effects
shrink dramatically and sometimes become statistically insignificant.
We regard these results as further supporting evidence for the hypothesis
that the Hispanic subgroup, especially those not receiving any training,
is particularly unique in various unobserved ways that make it a very
suspect control group.
Evidence from a Control Group Drawn from the 2000
Census
We further experiment with an alternative control group of
Hispanics constructed from Census 2000 data, again using the ME.
Even though we take particular care in drawing a comparison group with
similar characteristics as the eligible applicants to Job Corps, the use
of such a control group is disputable. However, it is possible that, if
the Hispanic control group members exhibit implausibly high outcomes,
the comparison with this more representative alternative control group
should circumvent this problem. The results in the last two columns of
Table 2 show that the estimated effect using the census data as a control
group is large and statistically significant. While this suggests that
Hispanic individuals in the treatment group are clearly better off than
those in the census control group, we hesitate to attribute this exclusively
to the effect of training, and regard these results as suggestive only.
Some Potential Explanations for the Lack of Programmatic
Effects on Hispanics
We consider two possible explanations to the lack of effects
in earnings of Job Corps and training on Hispanics: sample attrition and
geographic mismatch. Unreported figures of the change in samples for each
ethnic group due to attrition reveals that the change in sample sizes
is not uniform across ethnic groups: whites and others have a drop-off
in sample size of between 16.9 to 18.5 percent, while for Hispanics the
decline is 21.5 percent. The other overall pattern is that the reduction
in sample size is greatest with program-group members, 22.5 versus 18.7
percent for control-group members, respectively.
A consistent pattern with the Hispanic sample is the extent to which they
reside in large cities. Among the non-treated, 47.7 percent of Hispanics,
35.8 of others, and 14.3 percent of whites reside in a PMSA. This characteristic
may play a significant role in the findings above. Heckman, Ichimura,
Smith, and Todd (1998) stress the importance of comparing individuals
in the same geographic locations to control for potential differences
in the local labor markets. We believe this is a potentially important
factor to control for given the documented differences in the Hispanic
subpopulations. In an unreported table available upon request, we demonstrate
that treated Hispanics living in PMSAs average $36.3 less per week than
non-treated Hispanics. This effect exists only among Hispanics, as treated
and non-treated whites and others have statistically equal earnings in
PMSAs. We are in the process of obtaining the restricted-use data to address
its potential effect in the estimated impacts.
Conclusions
Our findings shed some light on the NCJS study's results for
the Hispanics sub-sample. While in many cases the non-experimental estimators
we employ find zero or negative effects of Job Corps for Hispanics, our
examination of the data shows that Hispanics exhibit some unique characteristics,
especially among the non-trained control group. A more detailed analysis
of the possible explanations for this finding will be addressed in the
future.
We suspect that the Hispanic control group, especially the untrained,
pose problems when used as a comparison group, and it is this group that
is partly responsible for these outcomes. We find that the programmatic
effects of treated Hispanics improve when compared to non-Hispanics not
treated, while the effects of treated non-Hispanics decrease substantially
and sometimes become insignificant when they are compared to non-treated
Hispanics. These results hold even when including only whites in the non-Hispanic
group, which traditionally have higher earnings than Hispanics.
We also compare treated Hispanics to an ad-hoc control group constructed
following Job Corps eligibility guidelines from the 2000 Census, finding
that treated Hispanics fare extremely well compared to the Census control
group, which we regard as additional evidence that the Hispanic control
group is somehow odd. We believe that the reason why randomization did
not yield comparable Hispanic treatment and control groups is that it
was applied to the whole sample and not to the different subpopulations
of interest, such as Hispanics. In this respect, this study raises some
concerns that policy makers and social experiments should keep in mind.
Across the estimates we obtain in this paper with different methodologies,
we still find that Job Corps appears to have insignificant effects on
Hispanics. Our ultimate goal, which we have not accomplished here, is
to find explanations for this result. It remains possible that Job Corps
training for Hispanic youth is no more effective than substitute training
programs. However, in our view, the most important issue that needs to
be considered is the impact that local labor markets in large metropolitan
areas have on Hispanic youth.
Notes
1. Abadie and Imbens (2002) provide some Monte Carlo evidence
about the finite-sample properties of the bias-corrected estimator.
References
Abadie, Alberto, and Guido W. Imbens. 2002.
"Simple and Bias-Corrected Matching Estimators for Average Treatment Effects."
http://emlab.berkeley.edu/users/imbens/. Department of Economics, Harvard
University and University of California-Berkeley.
Burghardt, John, Peter Z. Schochet, Sheena McConnell, Terry Johnson, R.
Mark Gritz et al. 2001. "Does Job Corps Work? Summary of the National Job
Corps Study." Princeton: Mathematica Policy Research, pp. 8140-8530.
Heckman, James J. 1979. "Sample Selection Bias as a Specification Error."
Econometrica, Vol. 47, no. 1, pp. 153-61.
Heckman, James J., Neil Hohmann, Jeffrey Smith, and Michael Khoo. 2000.
"Substitution and Dropout Bias in Social Experiments: A Study of an Influential
Social Experiment." Quarterly Journal of Economics, Vol. 115, pp.
651-94.
Heckman, James J., Hidehiko Ichimura, Jeffrey A. Smith, and Petra Todd.
1998. "Characterizing Selection Bias Using Experimental Data." Econometrica,
Vol. 66, no. 5, pp. 1017-98.
Schochet, Peter Z., John Burghardt, and Steven Glazerman. 2001. "National
Job Corps Study: The Impacts of Job Corps on Participants' Employment and
Related Outcomes." Princeton: Mathematica Policy Research, pp. 8140-8530.
U.S. Department of Labor, Employment, and Training Administration. 1999.
Job Corps Fact Sheet.
.
|
